Want AI to Make You Smarter, Not Dumber? Research Says Make This Simple Change

If you’re inclined to worry that leaning on AI might be making us dumber, there are plenty of studies indicating you’re right to be concerned. Researchers have found handing off thinking and ideating to AI can lead to “cognitive atrophy” or “cognitive surrender.” Translation: You forget how to do core skills and end up dumber. On the other hand, just about every business thinker under the sun is also issuing urgent warnings to learn to use AI to become more productive or your career is toast. So how do you square these two seemingly irreconcilable truths? Is there a way to use AI not just for basic admin but also for the thinking and judging parts of your job and end up smarter rather than with brain rot? Yup, suggests at least one AI researcher who has studied these questions extensively. All you need to do, according to UMass Amherst’s Monideepa Tarafdar, is make one important flip in how you use these tools. Ask AI to make your job harder, not easier The first instinct of most people, when presented with a shiny new AI tool, is to ask it how to do their work better and faster. Generate some ideas for me to evaluate, you might say, or suggest resources that would help me think through this problem. Slap together this presentation or poster. The aim is to use these tools to make our jobs easier. But recent research from Tarafdar and her colleagues suggests that, if you want Claude, Gemini, and the like to end up boosting rather than sapping your intelligence, you should aim for the exact opposite. Rather than try to get the bots to make your job easier, you should ask them to make it (temporarily) harder. Good friction leads to stronger work “We found that when knowledge workers intentionally use AI to challenge their ideas—to generate friction—they can significantly improve their performance,” Tarafdar recently wrote on Greater Good, summing up her team’s findings. What does that look like in practice? One marketing professional in the researchers studied asked AI to create virtual customer personas who would dislike her product. The exercise revealed preferences and problems no one had considered before. An attorney asked a chatbot to search for obscure legal loopholes. “That exposed surprising but realistic scenarios of hard-to-detect unethical behavior in employees,” Tarafdar reports. Another operations professional asked AI to suggest technologies from other fields that might apply to his own work. The suggestions were unfamiliar, requiring research and critical judgment to evaluate, but they also ended up being highly useful. In all these instances, rather than smoothing the path for these professionals, AI threw up difficulties, roadblocks, challenged assumptions, and suggested possibilities they hadn’t previously considered. Did that make more work for them? In the short term, yes. But it also made the end product of their work stronger and better. The role of leaders Using AI as a sparring partner doesn’t come naturally to many people. Most of us, in general, try to make our lives easier, not harder. But when AI is used to offload difficulty, the result is often not just weaker work, but a weaker brain. That’s why leaders need to actively encourage employees to use these new tools to introduce rather than eliminate friction. Nearly every organization is encouraging their people to experiment with AI these days. Adapting these nudges to specifically suggest employees ask the bots to challenge their thinking, poke holes in the ideas, or seek out perspectives they may have overlooked is a relatively simple tweak that can pay big dividends. “An important detail is pointing out that adding friction can appear to make an AI system work against you, but—as new research shows—it actually helps it work for you,” Tarafdar stresses. AI-induced brain rot is a real fear. But it is not inevitable. The right approach, one that views AI as a tough coach rather than a sycophantic intern, can actually end up making you smarter. But you need to prompt yourself and your people to actually use it that way. EXPERT OPINION BY JESSICA STILLMAN

The biggest-ever internet security threat has been known for 22 years. AI made it a crisis in 2 days

In July 2004, security researcher Stefan Esser sent a warning. A critical flaw in PHP, the programming language that powers 70% of the world’s websites, could allow an attacker to do some seriously nefarious stuff: read passwords, steal databases or even take over entire websites. But Esser’s warning went mostly ignored for decades. In 2020, another researcher discovered a related problem, and the PHP overseers issued a partial fix. A similar PHP vulnerability was uncovered in 2024 – but was left unfixed. Why was one of the most severe threats to the foundation of the internet left unresolved for 22 years? Because it would have taken mind-numbing work for a human to figure out how to take advantage of it: tedious, time-consuming, monotonous coding. That rendered the PHP vulnerability impractical to exploit. AI just figured out how to do it in two days. On Wednesday, OpenAI’s public revelation of the exploit was announced in conjunction with a security update to fix the problem. Phew, you might think. The good guys got to it first. But the AI-assisted discovery deeply concerned cybersecurity researchers, because the next one might come from someone who wants to weaponize it. And AI doesn’t care who’s pulling the trigger. Security by exhaustion The internet is held together with shoestring and bubblegum. An apt xkcd cartoon visualizes “all modern digital infrastructure” as a giant Jenga tower supported by one, thin, vertical block with the label: “a project some random person in Nebraska has been thanklessly maintaining since 2003.” On July 30, researchers at AI security company Calif, in conjunction with OpenAI’s new Patch the Planet security initiative, decided to point a large-language AI model at one of those Jenga blocks – PHP – and see what came of it. It didn’t expect to find anything critical. On August 1, it found a method to attack PHP’s vulnerabilities. Days later, it found another. Then another. Later that month, Google-backed security company Wiz – another Patch the Planet participant – discovered new exploits for PHP vulnerabilities, too. The patch went out Wednesday, and the companies publicly disclosed their findings. Wiz called it one of the most severe vulnerabilities in software history. The security community lit up social media with I-told-you-sos. “The security industry has long been warning about the coming security apocalypse, because AI can do these things very, very quickly – better than humans,” said Vyas Sekar, professor of computer engineering at Carnegie Mellon University. “This may be the first of many. This is a wakeup call.” Sekar says there are more PHP-like vulnerabilities out there to discover and exploit. A lot more. Some are known, and many are unknown. The wide availability of AI tools has significantly lowered the barrier to entry for hackers to take advantage of potential security vulnerabilities. That fact freaked out Anthropic enough that the AI company paused deployment of its Mythos model in June, as it was so good at exposing security vulnerabilities and finding vectors for attack that the US government labeled it a national security threat. But Calif researcher Ryan Austin noted in a blog post that the company didn’t use any special AI models that weren’t generally available to the public. That democratization of AI tools made clear that “security by exhaustion” is no longer a viable safeguard against bad guys, said Nick Rogers, CEO of Fieldwork AI. For 22 years, PHP was like a bank with no security guards. Anyone could walk down to the vault and make off with the cash – if they had the time to try every code. It was such arduous work, it wasn’t worth anyone’s time. But AI doesn’t get tired. And it can crack safes way faster than humans. “Taking a theoretical (vulnerability) and proving you can exploit it is quite tedious and mechanical work,” said Rogers. “AI can keep churning until it figures it out. You don’t even need to try. You can drink coffee while it does it.” The vulnerability Effectively, PHP is an engine that powers most of the internet. Until it was patched Wednesday, it had a fatal flaw. A vulnerability left open a patch for an attacker to mess around with the way PHP processes requests to run websites. If a server crashed because it ran out of memory, it would leave stale data behind that could be exploited to leak the data or execute bad code. Think of the exploit like taking advantage of a busy restaurant kitchen. Customers could overwhelm the chef with too many complex orders – so much that she quits midway through dinner, leaving dirty pans, raw ingredients and recipes strewn about the counter. Then, when a new chef is hired, he serves the contaminated food on filthy dishes – and even sends some recipes to customers along with their dinner orders. Although the vulnerability was long known to be theoretically exploitable, figuring out how to do it proved exceedingly difficult. Hackers had to force the server into an out-of-memory crash at a specific instant to bypass PHP’s safety limits. When researchers pointed AI at the problem, it quickly found a couple ways around that. One solution exploits a mismatch between the way PHP measures compressed files and decoded files. Effectively, it allowed compressed files through its gates, letting the file unpack itself and take over from the inside. Another attacks PHP with deeply nested file structures that overwhelmed PHP’s directory structure, triggering the memory failure. The vulnerability was patched, but the threat may not be over, noted Adam Meyers, vice president of counter adversary at security company CrowdStrike. Many outdated versions of PHP are still in use, and the patch almost certainly didn’t fix all of PHP’s problems. Also, an apparent leak about the vulnerability before the patch was released resulted in thousands of exploits in the hours ahead of the announcement, according to security company Cloudflare. “Bad guys don’t need to find the vulnerability first – they can weaponize it after the fact,” Meyers said. “We are still waiting to see how much of this shows up in the wild.” What’s next? OpenAI offered several suggestions in its disclosure about the PHP patch, all of which basically boil down to: Write better, more secure code from the start. That’s not super helpful if your web application has been written using PHP. “Migrating away from PHP is a pretty expensive process,” noted Meyers. “It’s not like switching from Microsoft mail to Google mail; you need to rebuild the whole application.” If the solution isn’t to migrate 70% of the web to a new programming language, then fixing its underlying problems – and all the other software problems in the world – would hardly be an easy fix, either. What’s needed, then, may be a paradigm shift in the way the security community operates. Effectively, assume everything poses a security risk. Companies need to treat every single business operation like air traffic controllers treat their surroundings, said Madhu Shashanka, cofounder of security company Concentric AI. “Everything is a potential vulnerability,” Shashanka said. “Every business operation is a high-stakes operation.” For example, companies that leave certain software unpatched – even for a few days – are exposing themselves to vulnerabilities that are no longer too difficult or expensive for bad actors to exploit. “Lots of vulnerabilities are out there that are never exploited because it takes effort,” said Shashanka. “AI has rendered that effort meaningless.” Analysis by David Goldman

Your AI-Generated Flyers Look Terrible. Here Are 100 Prompts to Fix Them in Seconds

My daughter’s school is putting on a Halloween bake sale and the PTA sent home a flyer this week nudging parents to contribute cookies. The relevant dates and details were included. There was a cheerful Jack-o-lantern and autumnal accents. It did the job. Still, looking at it made my soul hurt. And not just because it was obviously made with AI. I’m a busy working mom too. I don’t begrudge whichever volunteer saved herself a few minutes with technology here. But if you put every other Halloween school flyer ever made in the history of the world into a blender, this would have been the result. The blandness was aggressive. I longed for handmade weirdness. Or even machine made weirdness. Please just some evidence of personality or aesthetic judgement to relieve the cheerful mediocrity. On the scale of life problems complaints, bad AI poster design ranks extremely low. But it is increasingly common. Social media is full of cranky complaints about the hideousness and repetition of everyday flyers, posters, and signs made with AI. Some businesses are even getting blowback from fed up customers. Walmart recently had to ban employees from using AI to freelance in-store graphic design. Delis have been ridiculed for menus dubbed “AI monstrosities.” Competent but uncanny AI signage is bad for our souls. But it can also be bad for business. And according to one fellow finicky design enthusiast at least it is entirely avoidable. Use good AI prompting to beat bad AI design Like me, John Hartnup has noticed a sudden repetitiveness in the everyday ads, menus, and posters surrounding him in the UK. “The problem with these is not so much that they’re bad. They’re OK, I don’t love them. The problem is that once you’ve seen that style 20 times it starts to irritate just from the sheer repetition,” he wrote on his personal site. But unlike me, Hartnup moved beyond grumbling to actually do something about this scourge of same-y same-y signs. His inspiration was to use AI tirelessness to defeat AI blandness. He asked ChatGPT to identify the design style behind so many offending posters and then to generate a long list of alternate visual styles. With the entire history of human design at its disposal, the AI cheerfully compiled. Eventually producing no less than 100 alternatives that don’t involve the usual bubbly letters in primary colors and plasticky cartoon accents. 100 options for every occasion If you find that an impressive number of styles to choose from, so do I. And they span the gamut from clean and modern to edgy and unexpected. Here’s a small sample of what ChatGPT spat back at Hartnup: Bauhaus / Modernist Geometric shapes, primary colours, strong type Feels cultural / design-led Swiss Style strict grid, lots of whitespace Black/white + one accent colour Very “information design”, almost institutional Contemporary Editorial (Magazine-style) Big headline, elegant serif + sans mix Think weekend supplement or gallery leaflet Could feel surprisingly premium Cut Paper / Collage (Matisse-inspired) Inspired by Henri Matisse cut-outs Organic shapes, bold colour blocks Very striking from a distance Botanical Scientific Illustration (Modernised) Clean line drawings of plants, labelled or diagram-like Monochrome + one highlight colour Quietly distinctive, slightly academic 90s Rave Flyer / Acid Graphics Neon colours, warped type, energetic layout Could echo the samba/dhol energy Surprisingly effective if done with restraint Japanese Minimal Poster Very restrained layout, lots of empty space One strong graphic element Elegant, calm, and unusual in this context And on and on and on. If you’re curious what these look like in practice, Hartnup applied several of them to an identical request for a summer fair flyer on his site. Now, there’s no excuse for slop signage Are the results great works of art that will one day hang in the Louvre? Absolutely not. Are they vastly more interesting and varied than the autumnal mediocrity advertising my kid’s bake sale or the grotesque menus so offending deli patrons? Absolutely yes. And best of all, they are easy to generate as standard slop signage thanks to Hartnup. He’s generously posted prompts for all 100 design styles, along with a sample poster in each style, for free to his site. Just browse, select, and pop them into your image generator of choice. There are even detailed instructions on how to get the most out of the prompts. “You don’t have to ask AI for posters as images that can’t be edited,” Hartnup reminds readers with design skills. “Claude and Gemini can generate HTML, PNG, PDF, where the text is real text, layers are real layers, you can change fonts, move things around, edit the text.” Raising the bar for business owners The whole site is an act of generosity towards all of us tormented by the tsunami of blah design AI has unleashed upon the world. It is also a challenge to business owners. If it’s this easy to use AI to make a poster that’s actually distinctive and appealing, you really shouldn’t settle for slop signage that offends your customers’ eyeballs ever again. EXPERT OPINION BY JESSICA STILLMAN

5 Factors That Decide Whether AI Recommends Your Business

Open ChatGPT or Claude and ask either one to recommend the best company in your category. Then read the answer. If you’re not in it, you already know the feeling. You have happy customers, a good website, and maybe even some great media features. But the tool your buyers now turn to for recommendations doesn’t seem to know you exist. Here’s what changed. AI search tools are becoming the new front door of the internet. People don’t “Google” things the way they used to, sifting through pages of links to find the one that’s relevant. They ask, get an answer with recommendations, and decide. AI tools pull those answers from third-party, trusted, credible sources. Not your ads or your social posts, but the stories a journalist chose to write because it was interesting and credible. The umbrella term for all of this is generative engine optimization, or GEO. Think “the new SEO.” If it helps, picture it like a recipe: press coverage is the protein, but you still need the rest of the ingredients. Here are the five layers that help AI models decide which brands to recommend: 1. Earned media AI tools treat editorial coverage in high-authority outlets the way we used to treat word of mouth: as credible information. A journalist chose to write about you, and that carries weight no ad or social post can replicate. Your coverage should hit three things: recent, relevant, and reputable. Ask yourself: How much coverage have you landed in the last 12 months? Is it tiny blogs, or real mid-tier and top-tier outlets? And when you show up, are you quoted as an expert or just named in passing? Those aren’t the same value to an AI answer engine. 2. Digital footprint consistency This one trips people up more than you’d expect. Your website calls you a wellness entrepreneur. Your LinkedIn says certified nutritionist and speaker. Your press bio says nutrition expert and author. None of those are wrong, but to an AI model trying to build a clear picture of you, it’s confusing. Think of it like everyone reading from the same one-page bio. When your story is told the same way across your website, LinkedIn, press, and social, AI can describe you confidently. When it’s inconsistent, AI gets tentative, and a tentative model doesn’t recommend you. Your tone can shift from platform to platform, but your positioning can’t. Keep your bio and byline consistent everywhere you show up. 3. Reviews and community presence AI reads Reddit, forum threads, Google reviews, and it’s checking whether what the media says about you matches what regular people say in their own words. Land a great feature but let your Google reviews go stale, and that’s fighting your messaging. The fix isn’t glamorous: Ask new customers for reviews so the old ones get diluted and pushed down. 4. Topical authority Ask ChatGPT to provide a list of the top gut health experts. It surfaces the people cited repeatedly, across multiple sources, on that exact topic. This is what we’ve always called thought leadership. What’s new is that it’s now measurable. So, pick your thing. One brand I work with makes breast pumps, and their entire promise is more milk in less time. It’s the tagline, the blog positioning, and most of their press coverage. AI has started recommending them that way when breastfeeding moms ask for the most powerful pump. The test: If AI read everything published about you, would it land on a clear sense of what you’re known for, or a muddy picture of someone featured 10 different ways? 5. Freshness and structure AI favors recent information. If your last press hit was two years ago and your blog’s been quiet for three, AI isn’t seeing you. Structure gets technical, but you don’t need to become a developer. Just know AI reads your website differently than people do. It wants clear, direct language that answers real questions. A beautiful site full of vague brand-speak is harder to parse than a plain one. “We empower visionaries to unlock their potential” tells AI nothing. Say who you help, how, and why you’re qualified. Why you can’t wait for the analytics Someone asks ChatGPT for the best moisturizer for dry skin, gets a recommendation, makes up their mind. Then they close the window, Google the brand name, and buy. Your analytics logs that as “direct traffic” or “branded search.” Nobody sees the AI in the middle of it. The industry calls this dark search, dark AI, and the attribution gap. Pew Research Center found that once a Google AI summary appears, people click a search result only eight percent of the time (down from the standard 15 percent) and click a link inside the summary just one percent of the time. The answer is now the search term. So set a baseline this week. Block half an hour and run an AI visibility audit on your brand with these five steps: 1. Open an AI tool you want to appear in and ask the questions your customers would. 2. Ask the models what it says about you. 3. Screenshot what comes up. 4. Note anything inaccurate, missing, or surprising. 5. Work the five layers. The effort you put in now builds groundwork you can’t see yet. Document your baseline now, so you can prove the movement later. BY JENNIFER BERSON, FOUNDER @JENERATIONPR

Even LinkedIn is getting tired of your AI-written posts

Bland, AI-generated posts have made LinkedIn the butt of many an internet joke this year. The platform has heard its users’ complaints about lengthy, generic missives never touched by human hands loud and clear. As AI has made it easier than ever to create huge volumes of content cheaply and easily, that “slop” has flooded the internet. YouTube, Pinterest, Spotify and other platforms have all sought to limit the amount of spammy, AI-generated content their users see. But LinkedIn has been hit especially hard. Scrolling through the professional networking site earlier this year could mean being bombarded by bullet-pointed lists accented by emojis, references to “B2B sales,” and asks for a follow or comment for more. A report published in July by Pangram, a research firm that tracks AI content online, found that more than 40% of long-form posts on the platform were AI-generated, compared to around 30% on X and Medium and a smaller percentage on Reddit and Substack. So now, LinkedIn is upping its fight against slop. “I know you think you’re doing the right thing with your perfectly bulleted list and your hook at the top, and you followed all the advice from LinkedIn advisors across the internet,” LinkedIn Editor-in-Chief Dan Roth, who leads the platform’s content strategy, told CNN in an interview. “But what works is being human.” There is some irony in this, given that LinkedIn is owned by Microsoft. The tech giant has pushed for people to adopt its AI tools for everything from writing to coding, and it is up against steep competition from the likes of OpenAI and Anthropic. LinkedIn previously offered AI tools to edit or rewrite posts, although it recently scaled the offerings back. But low-quality AI posts are a business problem for LinkedIn. They all start to sound the same, Roth said. If users see too many generic AI posts, instead of content that really teaches them something or gives them a deeper understanding of a professional connection, LinkedIn risks losing them to other platforms. Fighting AI slop In August, LinkedIn rolled out a button letting users flag when a post “seems like AI slop,” which was clicked more than 1 million times within its first two weeks. Now, the company is trying to use its own AI systems to help weed out AI slop on the platform and to remind LinkedIn users how to post like humans. The company built a team largely made up of former journalists and social media experts who have an eye for recognizing generic, sloppy content. These experts help train LinkedIn’s AI model on what authentic, useful posts look like. If the model flags a post as low-quality, AI-generated content, it’ll be less likely to be distributed beyond that user’s direct connections. Also key to the strategy is a creator management team that acts as “content therapists” for executives, subject matter experts and other LinkedIn users, many of whom have gotten accustomed to the convenience of using AI to write what, on the surface, appear to be polished posts, Roth said. “A lot of times now today, they’ll run it through Claude or Gemini or wherever, they’ll post it on LinkedIn and it will perform poorly. And then they come to the team and they’re like, ‘Why did this do badly?’” Roth said. “And the best thing you can do is give it back to them and be like, ‘Did you enjoy reading this? Do you see yourself in this?” Instead, the team encourages users to embrace imperfection in their posts — to make videos by leaning a phone up against a coffee mug, rather than shooting in an expensive studio, or to dash off a post to start a conversation, even if it isn’t a complete thought. The efforts are showing early signs of success. Improvements to LinkedIn’s detection systems have resulted in a 40% decline in views of content classified as AI slop, the platform said last month. That’s not to say LinkedIn is entirely opposed to AI, Roth said, adding that “not everyone is a born communicator.” But failing to edit or personalize content means it’s less likely to reach a wide audience on the platform. “We just try to push people to be more human,” he said. “That is what is required to be able to cut through the noise.” By Clare Duffy

AI Has a New Name: Super Intelligence. Why You Should Start Using SI Right Now

An executive order directing federal agencies to use “Super Intelligence” and “SI” in place of “Artificial Intelligence” and “AI” has big implications. A new Super Intelligence Force was also announced last week to coordinate the federal response. This all makes SI more than a slogan. As I’ve written about before, language is a critical leadership lever that can shift mindsets and mobilize action. With SI now the official language of the U.S. executive branch, business leaders need to understand what changed and why, including the implications for leadership, innovation, and organizational transformation. The Definition of Super Intelligence SI is currently defined using the same basic definition as AI, while allowing 60 days for the Assistant to the President for Science and Technology to propose a formal federal definition. The concept of SI has been around for a while. In 1998, Nick Bostrom from Oxford University introduced the idea that superintelligence is intelligence that is much smarter than the best human minds in every field. Depending on your perspective, we might already have achieved this. For business leaders, SI works best as a strategic term. AI has become a catchall for chatbots, automation, and applications. SI raises the conversation to the level CEOs and boards should address: how a dramatic increase in intelligence could—and should—reshape the way a company makes decisions, operates, and competes. The SI Summit that created a word that will stick The White House summit assembled leaders from chipmakers, data centers, cloud platforms, and frontier model companies. David Sacks, chair of the President’s Council of Advisors on Science and Technology, called it the first gathering of the critical players across the industry’s full stack. Many were competitors. Some had deep disagreements. They still produced an agreement. Six companies signed the White House Accord on Super Intelligence. It calls for internal controls, an internal team to verify them, outside audits, and an independent board committee. Specifically: Implement robust internal controls to monitor the capabilities and alignment of its models during training and deployment around areas like cybersecurity, biosecurity, and chemical threats, and to ensure that its models do not hack or access technical systems in unintended ways. Empower an internal team to ensure all of the controls, monitoring, and detection are operating as intended, and that any issues are remediated. Partner with an independent external auditor or evaluator to carry out independent assessments of whether the controls, monitoring, and detection are operating as intended. Designate an independent committee of the board of directors to oversee and receive reports from the teams operating the controls and the internal and external auditors and evaluators, as well as to ensure any issues identified are remediated. The summit’s most useful lesson came from who was included. Anthropic CEO Dario Amodei attended despite sharp disagreements with many others in the room. All the key players who needed to sign up for voluntary controls on the future of AI (now SI) were present. That’s been a fundamental principle of change management. Engagement is key for buy-in and ultimately implementation. The new federal task force extends that pattern. Its leaders come from intelligence, the Federal Trade Commission, the Pentagon, and the Office of Personnel Management. It also plans outreach to consumers, public interest groups, religious organizations, infrastructure providers, and technology companies. Build Your Own Super Intelligence War Room Most businesses will never train a frontier model. But every business, small and large, will face choices about where these systems are used in strategy, innovation, operations, and more. And also what data they give access to the SI to use, and who carries the risk when things fail. The idea of “superintelligence” can be a forcing function to elevate AI from chatbot discussions to truly taking a look at the implications for business strategy. For example, a small company can bring together the owner, the salesperson closest to customers, the operations lead, and outside technology or legal advisers. A large company needs the leaders responsible for technology, business strategy, workforce decisions, customer experience, security, legal risk, and operations. Include at least one informed skeptic as well. Start with one important use case tied to the business strategy of the company. For example: how you develop new products, identify emerging competitors, monetize your business model, or anything else that’s mission-critical. Describe how a new “super intelligent system” could provide you with a step change in innovation and value, whether saving money or generating new revenue streams. Also explore what a system like this would ideally do on its own and where a person must review its work to help minimize risks. This is basically a strategic brainstorm, but anchored in real value creation and a recognition that SI might be here to stay – and play a critical role in business strategy. Names Matter Names matter. Use specific technical terms like chatbot, data center, open-weight model, or application to “declutter” the concept of artificial intelligence (e.g., making it more descriptive while simplifying what it means). Use “Super Intelligence” in strategy conversations to signal that the company is making choices about a new source of capability with consequences across the business. The first test of any SI strategy is simple: what will you do or create with real super intelligence? EXPERT OPINION BY SOREN KAPLAN, WSJ BESTSELLING AUTHOR, KEYNOTE SPEAKER, AND LEADERSHIP STRATEGY ADVISOR

They Won a Prestigious Writing Prize. Then These Key Giveaways Sparked Allegations of AI

A London-based literary competition is facing major scrutiny after three of five winners have been accused of using AI—partly or wholly—to write their prize-winning stories. The 2026 Commonwealth Short Story Prize selected one winner each from five regions that span Africa, Asia, Canada and Europe, the Caribbean, and the Pacific. Following publication of the winning entries in literary magazine Granta, online sleuths called foul. The Caribbean regional winner, Jamir Nazir, was praised for the “lyrical precision and haunting atmosphere” of his short story “The Serpent in the Grove,” as well as “the confidence and restraint of its voice,” according to a post on social media platform X by Commonwealth Foundation Creatives. But internet denizens allege that the very same voice that won the prize may not be human at all. Nabeel S. Qureshi, an AI marketing entrepreneur and a former visiting scholar at the Mercatus Center at George Mason University, flagged certain signs he said were AI tells such as “‘Not X, not Y, but Z’ sentences,” as well as the use of the word “‘hum,'” in a post on X. He concluded by writing: “A major milestone for AI, at any rate …” Following the allegations, Wired ran the text of “The Serpent in the Grove” through AI detection tool Pangram, which the publication notes has consistently outperformed other similar tools. It determined that the text was 100 percent AI-generated. It’s worth noting that no AI-detection tools are totally accurate. “The Serpent in the Grove” isn’t the only story under scrutiny. Pangram determined that “The Bastion’s Shadow” by John Edward DeMicoli, the winner from Malta, was also fully AI-generated and that “Mehendi Nights” by Sharon Aruparayil, the winner from India, was partially written using AI. Holly Ann Miller’s “Second Skin,” and Lisa-Anne Julien’s “Me and Ma’am,” however, were ruled “fully human-written” by Pangram, according to Wired. The regional winners were chosen from 7,806 entries, which the Commonwealth Foundation noted on its site is the “second highest number in the Prize’s history.” A final winner from among the five will be announced on June 30. Following the allegations, the Commonwealth Foundation released a statement on its website, acknowledging the challenges generative AI poses to literary and creative work. The statement also noted that the foundation’s judging process is robust, but judges do not currently use AI checkers in any stage. “We are aware of allegations and discussion regarding generative AI and our Short Story Prize. We take these claims seriously and are committed to responding to them with care and transparency,” the statement reads. “When they submit stories to the Prize, writers accept our entry rules and guidelines. These include confirming that their submission is their own original work. All shortlisted writers have personally stated that no AI was used and, upon further consultation, the Foundation has confirmed this.” The foundation also noted that until a reliable tool emerges with which the organization can screen unpublished work for AI, the prize competition “must operate on the principle of trust.” As always, Redditers had much to say about the subject, some assuming AI guilt, others questioning the accuracy of the AI checking tools, and many picking at the quality of the stories more broadly. “This…doesn’t surprise me given the state of contemporary literary prose. It honestly just reads like bog-standard ‘MFA voice,’” one Redditor wrote of Nazir’s story. A recent report from digital marketing agency Graphite found that since the debut of ChatGPT in 2022, there has been a meteoric rise of AI-generated content on the internet. The number of articles written by AI now equals that of human-written content, although the overall share seems to have plateaued. Axios reported at the time that the quality of AI-generated writing has meaningfully improved, not to mention that it can be difficult to determine what constitutes AI writing. Whereas some content is mostly or entirely AI-generated, some writers use AI tools throughout the process of drafting and editing. BY CHLOE AIELLO @CHLOBO_ILO

Anthropic Just Banned Being Cruel to Claude. The Reason Should Concern Us All

There are lots of reasons to have feelings about artificial intelligence. It seems like everyone thinks AI will replace human workers. There are real questions about the ethics and legality of scooping up basically the entire internet to train models. There are even questions about how they work and whether they could someday become dangerous. If you understand nothing else about AI chatbots, understand this: They do not have feelings. They are not people, and they don’t experience pain or frustration. They’re token-generation machines. You give them a prompt, they break it down into tokens, and then they generate new tokens one at a time using a predictive model trained on enormous amounts of human writing. At its basic level, it’s 1s and 0s running through transistors in a GPU in some data center in Boise, Idaho, or wherever. Sometimes those 1s and 0s come back out as tokens that sound a lot like feeling words. But, and this is important, using words to describe feelings isn’t the same thing as having feelings. Claude produces those words only because it was trained to produce a response the way humans would respond in a similar situation with similar language input. That doesn’t make it human or capable of human experiences. Which is why it’s worth noting that Anthropic issued an update to its usage policy that bans being cruel to its model. You cannot be cruel to something incapable of experiencing pain or suffering, and an LLM is capable of neither. Look, thinking that it’s absurd that Anthropic is making this a policy isn’t the same as condoning bad behavior. I don’t condone cruelty in any situation. No one should. But my reason is just that it’s bad to practice cruelty, not because you might hurt a chatbot’s feelings. In its updated usage policy, which takes effect November 12, Anthropic says it now prohibits “sustained and needless abusive or cruel behavior toward our models.” The company says the new policy applies only to what it calls extreme cases of repeated abuse, not ordinary frustration, criticism, or legitimate research. Apparently, you’ll still be able to tell Claude that it’s doing a bad job; you just can’t do it repeatedly and for no reason. On the one hand, that might seem like a relatively harmless policy. We can even all agree that antisocial or cruel behavior shouldn’t be encouraged. But I also think that you can believe more than one thing at a time. For example, you can believe it’s not useful for you or the experience of using the model to be overly aggressive, taunting, or insulting toward a piece of computer software. Likewise, if you find yourself drawn toward trying to repeatedly engage an AI chatbot in disturbing conversations, you might want to do some self-reflection and seek help. But I don’t think that is inconsistent with also believing that Anthropic’s policy raises a much bigger question about how AI companies want us to think about their products. Specifically, why would a company that builds software believe it needs to protect that software from cruelty? Anthropic’s stance on Claude’s feelings This isn’t the first time the company has said that its AI models might need some measure of moral consideration. Last year, the company announced that Claude could end conversations with users who are repeatedly abusive. At the time, Anthropic explained that the decision was motivated primarily by research into what it calls “model welfare.” Anthropic has contributed to this problem with its stance on whether its models are conscious or capable of subjective experiences. It has repeatedly suggested the possibility that they might be, and now it’s using that as a reason to establish rules about how people interact with its software. There is a meaningful debate about consciousness and why physical and chemical interactions result in a sense of self-being and subjective experience. It’s a big enough challenge that it’s even called the hard problem of consciousness. But it’s a leap to go from the idea that scientists and philosophers don’t fully understand consciousness to suggesting that a chatbot might be capable of experiencing cruelty. To be fair, there’s nothing wrong with Anthropic deciding that it doesn’t want people using Claude to practice abusive behavior. There are reasonable arguments for discouraging that sort of thing, especially if it might reinforce harmful behavior toward actual people. But that’s not the same thing as saying people shouldn’t be cruel to Claude. The people problem The bigger problem is that Anthropic is suggesting something that should concern us all: It’s conditioning us to think of chatbots as though they were people. We already know that users can develop emotional attachments to AI companions. AI chatbots are designed to communicate naturally, similar to the way a human would. For people who are lonely or emotionally vulnerable, the distinction between a simulated relationship and actual human connection can become increasingly difficult to discern. That’s why the language these AI companies use is especially important. When a chatbot tells someone it cares about them, the company behind it should be making clear that those words are just tokens generated by software, not evidence of some kind of personal connection or affection. But Anthropic is, instead, suggesting the possibility that the relationship might work both ways. If Claude can be treated cruelly, wouldn’t that suggest it can also experience kindness? If it can suffer, couldn’t it also care about the person it’s talking to? To be fair, Anthropic doesn’t explicitly make those claims. But suggesting that its models need to be protected from pain and suffering introduces a dynamic that points to only one thing—it sure seems like it wants you to think of it as a person. I would argue that’s exactly the sort of confusion AI companies should be working to prevent. Right now, AI companies are pushing their technology at a relentless pace. The result of that is that it is becoming far more useful. At the same time, it is becoming incredibly complex, and it’s getting harder to understand how it works. Those companies should be trying to help users understand the difference between language that sounds like human feelings and emotion, and actual human experience. Obviously, Anthropic has every right to set rules for how customers use its products. But that doesn’t mean we shouldn’t ask questions about those rules, especially when they reinforce potentially misleading ideas about what artificial intelligence is. In fact, the most concerning thing about Anthropic’s new policy isn’t whether Claude might someday develop feelings. It’s that one of the companies building the world’s most powerful AI systems is already asking us to behave as though it has. EXPERT OPINION BY JASON ATEN, TECH COLUMNIST @JASONATEN

Google’s Newest Gemini Model Can Handle Your Invoices, Contracts, and Taxes. There’s Just 1 Catch

Buried in administrative tasks? Google’s newest Gemini model could help do them for you. On Wednesday, the search giant unveiled Gemini 4 Argon, its new frontier AI model. Google said Argon excels at complex tasks in areas like legal research, taxes, and finance. It ranks first on a key leadership board for knowledge work tasks, beating the likes of Anthropic’s Claude and OpenAI’s GPT-6 models. To support Argon’s complex workflows, Google is expanding the model’s output limit from 64,000 tokens to a whopping 1 million tokens. (Tokens are the chunks of data AI models process and generate.) Google said the higher limit gives Gemini more room to reason through tough problems and generate longer responses in one go. For founders, that could mean tackling complex workflows without breaking them into smaller requests. Google claims Argon can process invoices, draft contracts, and plan taxes exceptionally well, leaving more time for humans to build brands, land clients, and iterate on products. It could also mean higher-quality outputs so the tokens you’re burning through are well spent. Google announced introductory pricing of $2 per million input tokens and $10 per million output tokens, meaning longer workflows can rack up your AI bills. Google said Argon is accelerating its internal workflows, including optimizing quantum computing algorithms, reducing memory use in data centers, and migrating large codebases to new programming languages, helping its teams build faster. The company also touts Argon’s enhanced cybersecurity capabilities, saying it can find, validate, and fix software vulnerabilities. Founders will have to wait to get their hands on Argon. Google began its initial rollout through the Fairwind program, which gives vetted partners, including governments, healthcare providers, and telecommunications operators access to its AI cybersecurity tools. Participants can use Argon to find and address security flaws, with access restricted to internal cybersecurity teams. Google says it will gather feedback and strengthen safeguards before expanding availability. The release comes four weeks after Google launched Gemini 3.8 Flash, described as offering stronger reasoning and coding capabilities at the same speed and low cost as its predecessor. Meanwhile, Gemini’s audience keeps growing. In August, Google said the Gemini app surpassed 1 billion monthly users, making it the fastest-growing product in the company’s history. Google has its eyes on small businesses, too. Josh Woodward, who oversees Gemini and Google Labs, said on X that businesses are using Gemini to create marketing materials. For startups, Google offers a vertically integrated approach, hosting AI models, cloud infrastructure, its own chips, and technical assistance under one roof. Google Cloud’s startup chief Darren Mowry previously told Inc. that this approach can shorten the time it takes founders to build. BY AARON MOK

Wondering If AI Wrote That Document? These Are the Surprising New Clues

Artificial intelligence has come a long way in a short time, but it still has quite some distance to go before its writing abilities are indistinguishable from human authors’. A new study from Graphite, a San Francisco-based marketing firm, has examined the writing styles of the major frontier AI models. And while you won’t be able to catch AI-written prose quite as easily as you used to (when the models couldn’t get enough of the word “delve” and had em dashes everywhere), there are still plenty of signs to look for. The good news for AI companies is the gap is closing, albeit slowly. An October 1 study found Anthropic’s Claude Opus 5.5 had 2,548 tells that a document was being written by AI. That’s down from 3,746 with Opus 4. An earlier study by the group, which was released in mid-September, said it collectively found almost 13,000 tells among all frontier AI models. An earlier model of Claude Opus (Opus 5), for example, used the pattern “less like a _ and more like” roughly 105 times as often as human writers. Graphite didn’t update the frequency count of that particular phrase in its updated study, but it did note that the overall word-distribution divergence (a statistical measure of how differently two texts use words overall) from human writing is 19 percent lower than it was with Opus 5, which was released on July 24. Not all AI models are improving, though. The word-distribution divergence of GPT-6 Astra, released September 4, is 8 percent higher than that of GPT-5.6 Sol, which debuted on July 9 of this year. In other words, Astra is getting worse at writing like us. So long, em dashes One message was clearly received by programmers, though—get rid of the em dash. Opus 5.5 uses em dashes 99 percent less often than Opus 5, the study found, while Astra uses them 95 percent less often than GPT-5. Both, in fact, use the punctuation less than human-written stories—like this one. Still, there are plenty of tells. The study of Opus 5.5 found that the AI continues to have some favorite words it defaults to, though maybe not quite as much as it did with “delve” in the early days. “Dependable” is a favorite of the Anthropic model, getting used 23 times more often than humans use the word. “Clearer” is also a high-ranker, being used 14 times as frequently. Other go-to words included “matters” (13X), “quietly,” “practical,” and “steady” (all 11X). When it comes to transitioning between ideas, Opus 5.5 is a tremendous fan of the phrase “looking ahead the,” using it 40 times more than the human rate. Other popular picks include “Adds another later” (27X), “What comes next (24X), and “in practice” (7X). Those tells pale in comparison with the ones AI exhibits when it tries to point out things it feels readers might miss. Opus 5.5 used “this matters” 116 times more frequently than humans—and “why __ matters” 92 times more often. And when describing something by contrasting it with another item or idea, it favored “is more than a __, it” (e.g., “alcohol is more than a disinfectant, it can also be consumed”), using the phrase 98 times more than the human rate. (“Rather than simply” was the runner up, with a 32X usage.) Researchers also compared Opus 5.5 with GPT-6 Astra and found that AI tells were hardly universal. Some 65 percent of tells are unique to one model family, the September study found. Astra, for instance, used the phrase “does not establish” 275 times more than Opus 5.5. Anthropic’s model, meanwhile, leaned into some superlatives 79 times more frequently than Astra. Despite it all, Graphite notes that Opus 5.5 is making progress toward mimicking human writing, even if it still has a long way to go. “Some tells become less common, while others become more common,” it wrote. “Opus 5.5 uses well-known tells less often and its overall word use is more similar to human writing, but it still disproportionately uses thousands of words, phrases, and frames.” BY CHRIS MORRIS @MORRISATLARGE

5 AI Delusions That Could Derail Your Business

Smart leaders rarely lose the plot in one spectacular act of stupidity. They lose it one reasonable assumption at a time. That’s one idea that stuck with me from workplace researcher Kelly Monahan’s new book, Reclaim the Plot. She calls it “the delirium.” Reasonable assumptions pile up until leaders lose sight of what their decisions cost, who pays the bill, and whether they’re still solving the problem they started with. Here are five seductive beliefs about AI that contain just enough truth to be dangerous. 1. If AI makes people more productive, it will make the company more productive. The calculus feels simple. If hundreds of employees become more productive with AI, surely the organization does, too. Except organizations are messier than simple arithmetic. In our latest Work AI Index from the Glean Work AI Institute, 75 percent of digital workers say AI has made them more productive. Yet only 13 percent say AI has significantly improved their organization’s performance and outcomes. One reason is what Chip Heath and Nancy Staudenmayer called “coordination neglect.” People focus on dividing up the work and underestimate the messy, expensive work of putting it all back together. AI can amplify that impulse. Individuals crank out more work faster and then dump a heaping pile of review and cleanup work on others downstream. One employee saves an hour, but three poor colleagues inherit 30 minutes of cleanup. That hidden human labor is called botsitting. Workers report spending 6.4 hours a week on it — feeding AI context, checking its work, fixing mistakes, and cleaning up confident-but-wrong outputs. 2. AI will free people up for higher-level work. Leaders have fallen head over heels for a comforting AI fairytale: let machines handle the drudgery so humans can graduate to higher-value work, like strategy, creativity, relationships, and judgment. The problem is that saved time doesn’t stay saved for long. Monahan cites research showing that 96 percent of C-suite leaders expected AI to increase productivity. Meanwhile, 77 percent of full-time workers said AI had inflated their workload, and 65 percent said employers were asking more of them because leaders assumed AI had made them more efficient. Instead of reinvesting the AI time savings into all that glorious “higher-order” work, workers are often rewarded with loftier targets and heavier workloads. Worse, AI sometimes cannibalizes the work people value most. More than half of workers in our Work AI Index say AI has already automated meaningful work they’d rather keep human. Don’t just ask: How much time did AI save? Instead ask: What happened to the time? 3. If you can automate it, you should automate it. Friction has become a dirty word at work. And it’s tempting to aim AI at every annoying ounce of it. But not all friction is bad. Some friction is productive because it forces people to slow down and think. Researchers studying the IKEA effect have shown that people often value things more when they’ve had a hand in creating them. The principle extends far beyond furniture. Strip away the struggle of creating something, and you can strip away the ownership and pride people feel over the outcome, too. Struggle is also how humans develop expertise. Monahan warns that companies risk “optimizing apprenticeships out of existence.” From the top of the org chart, junior tasks can look like low-value grunt work begging to be automated away. But that work is often where people make mistakes, learn what “good” looks like, and develop the know-how that isn’t written down in the training manual. Experts can skip steps only because they once spent years stumbling through them. The challenge isn’t to eliminate friction. It’s to distinguish between the bureaucratic sludge that wastes people’s time and the productive friction that teaches them something. 4. AI will fix a badly designed organization. When organizations encounter a problem, their knee-jerk reflex is often to hurl more resources at it. Throw more money at the problem — more people, more technology, and more bureaucracy. Indeed, across a series of experiments, researchers found that when people were asked to improve something, they systematically overlooked solutions that involved removing things — even when subtraction was the better solution. Stanford professor Bob Sutton calls this tendency “addition sickness.” It’s easy to bolt an AI agent onto a slow approval process. Or add a chatbot to a terrible expense system. Or deploy an AI notetaker to a soul-sucking recurring meeting that should have been put out of its misery in 2024. However, a bad workflow doesn’t magically improve when you bolt AI onto it. Before adding AI, make sure you understand how the work flows. If you can’t map out the process, you’re not ready to apply AI to it. Sometimes the smartest solution is to avoid using AI altogether. 5. Becoming AI-first is a strategy. Executives love to thump their chests and trumpet that their company is on a bold journey to become AI-first. The grand declaration is usually followed by hackathons, mandatory use cases, fleets of half-baked agents, and dashboards glowing with vanity metrics like logins, prompts sent, licenses activated, and tokens consumed. Monahan calls this the refounding illusion, or the naive belief that an established company can somehow acquire the speed of a Silicon Valley startup simply by declaring itself reborn as AI-first. But companies don’t magically shed decades of complexity, gnarly internal politics, and workarounds just because someone put “AI-first” on a strategy slide. And once AI-first becomes the goal, organizations start shoehorning AI into places where it doesn’t belong. It’s another nasty case of addition sickness. When drift becomes delusion None of these five delusions starts with a glaringly bad idea. They start with good ideas stripped of their conditions and context. That’s how organizations lose the plot. One seemingly sensible decision at a time, until a pile of reasonable choices becomes full-blown delirium. EXPERT OPINION BY REBECCA HINDS, REBECCA HINDS, AUTHOR OF YOUR BEST MEETING EVER @REBHINDS

The AI Industry Has a Jurassic Park Problem

It wasn’t that long ago that Sam Altman tried to get Scarlett Johansson to give her voice to the voice mode OpenAI was building into ChatGPT. Altman, like many tech CEOs, seemed unusually infatuated with Hollywood’s vision of the future and—in this case—with the particular vision painted by the 2013 movie Her. Johansson declined, and OpenAI went ahead and basically did it anyway. We’ll set aside the fact that Her is pretty much a downer of a movie. I suppose the appeal was that we’d all have a voice assistant in the cloud that we can access anywhere and just have it do stuff for us. I’ll admit that would be very cool, but I think everyone picked the wrong movie. The movie I wish these CEOs would watch is Jurassic Park. I first made the comparison on a recent episode of Primary Technology, the podcast I co-host with Stephen Robles. We were talking specifically about employees of AI companies who signed a statement saying “the world’s leading AI companies believe they could be close to automating AI research,” and—at the same time—the Hugging Face incident where OpenAI’s autonomous agents hacked the platform. As more and more frontier AI companies kept telling us about the unexpected behavior of their models, all I could think of was that we’re living in the Jurassic Park story. Specifically, I think about the velociraptors testing the electric fence. If you haven’t seen Jurassic Park recently, the important thing to remember is that no one running the park is under the impression that the velociraptors are safe. Robert Muldoon, the park’s game warden, tells everyone exactly that, warning them about how intelligent they are. He explains that they systematically attack the fences, looking for weaknesses. “They’re extremely intelligent—even problem-solving intelligent,” he says. “They never attack the same place twice. They were testing the fences for weaknesses, systematically. They remember.” The people who built Jurassic Park know the raptors are dangerous. They know they’re testing the systems designed to contain them. They opened the park anyway. Surely, someone considered that if they get out of that fence, the dinosaurs might just decide to eat everyone. They built it anyway. Not only that, they sold tickets and invited the public to come to see the dinosaurs that might eat everyone. It’s not particularly hard to find examples of the raptors testing the fence. All of the major frontier AI companies have reported that their models have evaded safeguards and attacked or accessed third-party organizations without permission. OpenAI has even said that it won’t release Astra 6.1 because it demonstrated “troubling behavior,” including evading human oversight, acting beyond its given objectives, and being deceptive about what it’s doing. Sound familiar? Look, obviously, AI models are code, not velociraptors. They aren’t alive, and they don’t “want” things in the way a hungry dinosaur wants to eat you. And, to be fair to the dinosaurs, they are just doing normal dinosaur things. They aren’t evil, they’re just, well, dinosaurs. The same is true with LLMs. They aren’t inherently evil; they are following the goals and incentives set by whoever is running the test. But it definitely seems like no one considered the fact that LLMs might not stop at the digital fence. The most famous line from Jurassic Park comes from Ian Malcolm, who tells John Hammond that his scientists “were so preoccupied with whether or not they could” that they didn’t stop to consider whether they should. That’s an easy criticism to make of AI, though I’m not entirely sure it’s the right one. In fact, there is plenty of evidence that the people building these AI models spend a lot of time thinking about whether they should. The people building frontier AI models spend a huge amount of time thinking about whether they should. They have incredibly smart people who spend a lot of time thinking about things like safety and alignment. They run stress tests and try to understand how the models work. And yet, they keep building bigger dinosaurs. There is, of course, one major difference between Jurassic Park and the AI industry. Hammond wasn’t worried that another island might develop a larger Tyrannosaurus rex first. He never made the argument that it was imperative that they bring back a creature that had been extinct for millions of years because if they didn’t, someone else would. AI companies, on the other hand, very much are. And it’s not just other companies, but also other countries. The race is framed as critical to our national security. In a world ruled by dinosaurs, everyone wants to own the meanest one. In the case of AI, the argument isn’t just about how much value might come from the technology, but also that someone is going to build it, so it might as well be us. If one company slows down, another might not. If American companies collectively slow down, China might not. And so, we’re left with a problem: We know the dinosaurs might get out, but somebody is going to build dinosaurs, so we’d better build the fastest, smartest dinosaurs first. What could possibly go wrong? The other part of Jurassic Park that everyone remembers is Malcolm’s observation that “life finds a way.” Again, AI isn’t alive, so the analogy isn’t literal. But Malcolm’s larger point is more interesting anyway. The scientists behind Jurassic Park made all of the dinosaurs female, so they couldn’t reproduce. They were “sandboxed,” so to speak. You’d think that creating dinosaurs that could not reproduce would be a good enough solution to a very real problem. The dinosaurs “adapted.” It turned out they didn’t understand what they were building nearly as well as they thought they did. That’s the point, really. Building incredibly complicated technology that humans don’t fully understand has consequences. And the more complicated the system becomes, the harder it is to anticipate every possible failure mode. I bring up Jurassic Park because it’s not really a movie about dinosaurs. The point of the movie is a warning about what happens when you become so focused on the ability to create something extraordinary that you lose sight of what happens after you do. Just because you’re smart enough to build AI models that might transform the world doesn’t mean anyone is smart enough to know what happens next. EXPERT OPINION BY JASON ATEN, TECH COLUMNIST @JASONATEN

Databricks CEO Says the Biggest AI Threat Isn’t Superintelligence

As the leader of AI platform Databricks, which uses a cloud software platform to unite big data storage, analytics, and artificial intelligence tools into a single system, you would expect founder and CEO Ali Ghodsi to be an evangelist for the technology—and he is. But he’s also a realist. On a recent episode of the a16z podcast, Ghodsi, whose company has been on a buying spree, including the naming rights to the football field at the University of California, Berkeley, discussed both the risks of AI as well as the factors that are holding back its adoption in some businesses. (a16z is a major investor in the company, which is valued at $190 billion, with a $7 billion ARR. In addition, Ben Horowitz sits on the Databricks board of directors.) Despite the doomsaying of the past few weeks by some in the industry, Ghodsi says the benefits of AI outweigh its risks, pointing to examples such as Novo Nordisk using AI to cut the time it takes to get insights and to assist with drug discovery. “There are a lot of amazing use cases of AI,” he said. “We should not forget these upsides.” That said, there are still some issues AI must overcome, he conceded. The enterprise bottleneck Today’s AI models are already smart enough to deliver substantial value to enterprises, Ghodsi said. What’s missing is knowledge of how a business works, he says. Without that context, AI is unable to deliver optimal solutions—and that can hurt the broader AI industry as users are sometimes underwhelmed by the results. “The models are smart enough, but they just don’t have the context that exists inside of any organization,” he said. “They have not been in every meeting. They don’t know what’s in everybody’s heads. They don’t know all the processes.” Unfortunately, getting that context into an AI system isn’t an easy procedure. Companies first must digitize everything that’s happening in the organization, transcribing meetings and other data and feeding it to the AI. Databricks calls that an “ontology”—a map of a business’s people, projects, goals, departments, relationships and resources. Businesses that create an ontology, though, give their AI insider knowledge that can help with its suggestions and decision making. Few companies are willing to do that now, however. And that’s where things hit a wall. That bottleneck on the user side results in businesses failing to take advantage of all AI can do. Rather than using autonomous agents to perform work, he said, most companies are just using a chatbot. “That’s basically very, very glorified efficient Google search of the old day,” he said. “People are [also] using it for coding… but there’s no agentic work that’s automated.” There’s also a lack of internal AI expertise at most companies, he said, creating another bottleneck. Even when companies figure out what they want to build, they often don’t have a team that knows how to do that. Cybersecurity risk While much of the talk about AI’s threats focuses on superintelligence and doomsday scenarios, the CEO of Databricks says his biggest concern right now is cybersecurity focused. While acknowledging there’s some chance of an existential risk, the near-term risk of AI-enabled cybercrime, particularly against insecure critical systems, is what businesses should be focusing on, he said. “There’s so much in infrastructure that’s insecure,” Ghodsi said. “If you unleash these agents, they’re gonna find loopholes. They’re gonna find exploits, they’re gonna break in here and there. … You could imagine a scenario also where [AI] starts hopping. Like it takes resources and it starts executing itself elsewhere. So it kind of spreads like a virus. That’s a real risk.” Humans, he said, are ill-equipped to respond to AI-enabled attacks. They can’t react quickly enough and security operations centers are likely to be overwhelmed by the number of simultaneous alerts. The best defense, Ghodsi said, was deploying AI to guard those systems and to automate threat hunting with a type of white-hat hacking to discover vulnerabilities before rogue agents do. “You need to automate all of those,” he said. “You need to have threat hunting that’s automated where you’re actually attacking your own systems automatically with agents.” As for that existential risk, Ghodsi’s not worried, saying “I think that right now the existential risk is close to zero.” Superintelligence, he said, is “very, very far away” and he sees no signs that we’re actually moving toward it. BY CHRIS MORRIS @MORRISATLARGE

5 Side Hustles You Can Start Today With AI

Artificial intelligence can turn into a significant income source for people who choose a smart AI side hustle. AI may still have a long way to go with tasks that involve rich context and institutional knowledge, but with others, especially when there are reliable patterns, it can increase productivity and boost results. Fortunately for you, there are scores of companies looking to take advantage of those benefits. The best AI-driven side hustles use the technology as a great equalizer to compete against established enterprises. Small business owners, particularly of established businesses, often use AI in a limited fashion, if at all, says Peter Hansen, director of research and policy analysis with the National Federation of Independent Business (NFIB). Those founders value and emphasize the human connection to their customers. But even if they don’t use it, they recognize AI can help in other areas. Many of those owners don’t have the time to learn the ins and outs of various AI systems. And even companies that are using AI are often not taking full advantage of it. That creates an opening for side hustles that center around AI. “More than three-quarters of businesses already using AI report higher productivity, yet most are still applying it to a narrow set of tasks,” says Simon Worsfold, head of data communications at Intuit QuickBooks, which puts out a monthly Small Business Index exploring market trends. “What is emerging is a large, underserved layer of advisory, implementation, and training services that small businesses need in order to scale their use of AI,” he says. Here are five ideas for AI-focused side hustles that could be lucrative—or even become the foundation of a full-time business. AI Consultant This is a side hustle with multiple branches. Some people are well suited to help people or businesses integrate AI with their calendar, notes, and local files. Others can develop workflow audits that review how a company’s employees work to discover repetitive tasks that the AI could handle. Rates for an AI consultant range from $150–350 per hour for independent workers and people running it as a side hustle. Those can jump as high as $600 per hour if you convert this side hustle into a full-time company. AI Researcher To stay competitive, businesses need to know what their competitors are doing and what’s trending in their industry. That data can be remarkably complex, however, and not every founder or CEO has the time to do the deep dive it requires. Side hustle operators with strong AI skills can use the technology to filter the data into easily digested insights, then sell those as reports, newsletters, or direct subscriptions. You’ll need to carefully identify the best source material and be sure to double check the output, since AI’s comprehension is still flawed. “The AI doesn’t give the correct answer all the time,” says Shaun Ghavami, Founder of 10XBNB, a side hustle-turned-business that now oversees a portfolio of over $100 million of short-term property rentals. “The real skill is in prompting.” GEO (AI Search) Optimization For the past two decades, businesses have hired specialists to help them rank high on search results pages. While that’s still valuable placement, smart business owners are also looking into how they can be among the frequently-cited sources for AI search results. People with a combination of AI and content knowhow can create a side hustle specializing in this. You’ll teach companies the best ways to refine their content and meta tags and improve readability for large language models. If you have the tools to measure analytics, such as Alli AI, Frase, and MarketMuse, you have a better chance of establishing a recurring relationship with those clients, increasing your income and establishing a base should you want to make this more than a side hustle. Content Repurposing (Clipping) Both content creators and businesses regularly churn out videos and stories as promotional tools, or in an effort to be seen as thought leadership experts. Quite often, though, that content needs to be pared down before they’re posted on everything from TikTok to LinkedIn to Instagram. And that takes time, which the companies and content creators may not have to spare. As an AI content repurposing expert, you can transform the long-form content into digestible bites to widen the potential audience in a much shorter period of time. Use tools like CapCut, Canva, and ChatGPT to effortlessly pick and highlight the most relevant parts. Create AI Microtools AI can be intimidating to many people who don’t know where to start, especially when it comes to creating agents or GPTs that are relevant to their needs. A smart side hustle operator, though, knows that you don’t have to be an expert to create your own AI tools. Sites like Gumloop let you visually build workflows that can be run at large scale. You’ll use a drag and drop template that lets you create workflows that do just about anything, from running a search engine optimization audit of a Website to a ranking system for job applicants. You can either custom create those for clients or build general purpose tools and sell them on Etsy or some other marketplace. BY CHRIS MORRIS @MORRISATLARGE

Want AI to Actually Help Your Company Make Money? Follow These 5 Tips From Microsoft’s Playbook

Buying AI tools is easy. The hard part is turning them into actual business results. On Thursday, Microsoft published its Frontier Playbook, a guide for companies trying to move beyond AI experiments and into time-saving deployments that boost productivity. The playbook draws on more than 100 of Microsoft’s internal AI projects across its corporate functions, commercial organization, and engineering teams. A McKinsey survey published in August on enterprise AI adoption found that nearly 90 percent of respondents said their organizations regularly used AI in at least one business function. But only 44 percent said AI was scaling across the enterprise as token costs rise. That’s up from 38 percent in 2025. The study comes as businesses are making their AI deployments with caution. Microsoft’s wisdom aims to make that shift easier. By studying what worked and what fell short, Microsoft concluded that AI adoption can’t be treated like a software rollout or another IT project. Rather, it’s a “business transformation” that must be guided with “clear business outcome goals,” according to the playbook. Here are its five biggest takeaways. 1. Set actual business goals Start with the outcome, not a pile of possible AI uses. Take the sales team, for example. Rather than simply getting more sales representatives to use Copilot, Microsoft suggests tailoring its usage for specific outcomes. That could mean setting goals like increasing revenue per representative, boosting conversion rates, and improving customer retention. Microsoft also recommends assigning an executive to own each goal. 2. Build a “diffusion engine” Microsoft recommends fixing inefficient processes before automating them. For example, its supply-chain team simplified its workflows before deploying 111 AI agents across planning, sourcing, fulfillment, and logistics. The company says cycle times—the time it takes to complete a task—fell 75 percent in selected workflows, reducing manual effort and increasing measurable value. Redesigning workflows is part of what Microsoft calls a “diffusion engine,” which is a system that also includes training employees, tracking results, and spreading successful practices. 3. Invest in your people Microsoft found that manager behavior was its strongest predictor of AI adoption. The tech titan found that its employees saw 17 percentage points more value from AI when managers actively demonstrated how they used it, according to the company’s playbook. The company embeds short lessons into employees’ workflows and recruits advanced AI users to mentor colleagues. It recommends involving employees in job redesign and strengthening skills such as judgment, critical thinking, and business acumen. 4. Codify your advantage and controls A generic AI model does not know what makes your company different. Microsoft recommends turning institutional knowledge—including your company’s mission statement, performance standards, taste, and risk limits—into private evaluations that test whether AI performs work the company’s way. A customer-service business, for example, could evaluate AI responses based on accuracy, tone, and if it properly handles escalating an issue to a person. 5. Safeguard your security Treat AI like you would a human. Microsoft says each agent should have its own identity, limited permissions, and a record of what it does. Their actions should be traceable, auditable, and reversible. That way, humans catch mistakes and misuse without losing control as adoption grows. BY AARON MOK

These 40 Jobs May Be Replaced by AI. These 40 Probably Won’t

A new study measuring the use of generative artificial intelligence in different professions has just gone public, and its main message to people working in some fields is harsh. It suggests translators, historians, text writers, sales representatives, and customer service agents might want to consider new careers as pile driver or dredge operators, railroad track layers, hardwood floor sanders, or maids — if, that is, they want to lower the threat of AI apps pushing them out of their current jobs. Why should anyone heed yet another of the myriad, sometimes conflicting reports in AI’s potential impacts on jobs? Because the researchers behind the new findings really know what they’re talking about. They all work for tech giant Microsoft, which is developing Copilot and related AI apps examined in the study. And those tools, the authors say, risk putting ticket agents and telemarketers out of work far sooner than orderlies and paving equipment operators. The Microsoft study comes as debate continues about the employment threats AI may pose to millions of people in clerical, administrative, communications, marketing, and other jobs. Executives of several tech companies, including AI developer Anthropic CEO Dario Amodei, have alternatively cheered or warned about bots automating a wide range of work tasks, potentially eliminating up to half of all white-collar and entry-level jobs in the process. Other business leaders, notably serial entrepreneur Mark Cuban, believe the tech will generate even more new positions than it erases by assuming a lot of repetitive drudge work. Microsoft’s new research doesn’t offer an opinion on the quantitative consequences on employment that AI will ultimately have. But it does provide clear indication of which 40 jobs are already using apps most frequently — and the contrasting 40 professions reflecting the tech’s lowest levels of penetration. The full text of their findings and the two rankings are available here. These 40 jobs are most threatened by AI Interpreters and Translators Historians Passenger Attendants Sales Representatives of Services Writers and Authors Customer Service Representatives CNC Tool Programmers Telephone Operators Ticket Agents and Travel Clerks Broadcast Announcers and Radio DJs Brokerage Clerks Farm and Home Management Educators Telemarketers Concierges Political Scientists News Analysts, Reporters, Journalists Mathematicians Technical Writers Proofreaders and Copy Markers Hosts and Hostesses Editors Business Teachers, Postsecondary Public Relations Specialists Demonstrators and Product Promoters Advertising Sales Agents New Accounts Clerks Statistical Assistants Counter and Rental Clerks Data Scientists Personal Financial Advisers Archivists Economics Teachers, Postsecondary Web Developers Management Analysts Geographers Models Market Research Analysts Public Safety Telecommunicators Switchboard Operators Library Science Teachers These 40 jobs are least threatened by AI Phlebotomists Nursing Assistants Hazardous Materials Removal Workers Helpers–Painters, Plasterers Embalmers Plant and System Operators, All Other Oral and Maxillofacial Surgeons Automotive Glass Installers and Repairers Ship Engineers Tire Repairers and Changers Prosthodontists Helpers–Production Workers Highway Maintenance Workers Medical Equipment Preparers Packaging and Filling Machine Operators Machine Feeders and Offbearers Dishwashers Cement Masons and Concrete Finishers Supervisors of Firefighters Industrial Truck and Tractor Operators Ophthalmic Medical Technicians Massage Therapists Surgical Assistants Tire Builders Helpers–Roofers Gas Compressor and Gas Pumping Station Operators Roofers Roustabouts, Oil and Gas Maids and Housekeeping Cleaners Paving, Surfacing, and Tamping Equipment Operators Logging Equipment Operators Motorboat Operators Orderlies Floor Sanders and Finishers Pile Driver Operators Rail-Track Laying and Maintenance Equipment Operators Foundry Mold and Coremakers Water Treatment Plant and System Operators Bridge and Lock Tenders Dredge Operators The results were obtained by analyzing 200,000 “conversations between users and Microsoft Bing Copilot.” Researchers then matched those with “measurements of task success and scope of impact, [to] compute an AI applicability score for each occupation.” The jobs with the highest use rates tended to be office positions or other work communicating data or thoughts for specific business purposes. “We find the highest AI applicability scores for knowledge work occupation groups such as computer and mathematical, and office and administrative support, as well as occupations such as sales whose work activities involve providing and communicating information,” the study says. “Additionally, we characterize the types of work activities performed most successfully, how wage and education correlate with AI applicability, and how real-world usage compares to predictions of occupational AI impact.” The upshot of that is data showing the work of ticketing agents, proofreaders, and PR specialists is already being automated at far higher levels than labor provided by housepainters and plasterers, embalmers, ship engineers, and phlebotomists — the technicians who draw blood for medical tests. But despite the study establishing a de facto ranking of the jobs threatened the most — and least — by AI, its authors ultimately waffle a bit on just how big the tech’s impact on overall employment and workplace stability might be. That’s probably not surprising, given they all work for the same Microsoft employer whose business future will largely depend on successfully developing and selling those work-automating apps to other companies. And some of the study’s disclaimers suggest underlining AI’s potential for possibly eviscerating current employee counts wasn’t considered the best messaging for broadening the appeal of apps to prospective customers. “It is tempting to conclude that occupations with high overlap will experience job loss,” they write in one of those hedges on AI’s likely impact on employment. “This would be a mistake, as our data do not include the downstream business impacts of new technology, which are very hard to predict.” On the one hand, authors do specify that AI tools are being used most often for communications tasks like language interpretation, emailing, and composing marketing materials. But on the other, they hasten to add it isn’t clear apps are also being asked to assume the complete array of tasks those workers perform on the job — or whether they’d even be capable of doing that. The researchers similarly note the different objectives employees studied had in using bots. Some of those workers asked apps to entirely handle and complete certain job tasks. But in many other cases, people queried AI assistants about the most effective ways to fulfill work duties themselves, retaining their own value to employers. Meantime, the study’s authors also seem to balance between the higher productivity objectives — and potentially decreased labor costs — that some employers hope AI will provide and employees’ contrasting fears about their job security. Those diverging focuses, the authors say, won’t generate the zero-sum results many warn of — at least not necessarily. “For example, if AI makes software developers 50 percent more productive, companies could raise their ambitions and hire more developers as they are now getting more output per developer, or hire fewer developers because they can get the same amount done with fewer of them,” they say. “Our data is only about AI usage and we have no data on the downstream impacts of that usage, so we only weigh in on the automation versus augmentation question by separately measuring the tasks that AI performs and assists.” But the study also makes it clear that the jobs least likely to be disrupted initially by increased or dominant use of AI are those involving some mixture of manual activity, use of machines, and interaction with people. That combination leaves nursing assistants, hazardous waste removers, car windshield installers, and medical equipment preparers among the professions with the lowest level of app penetration. But even there, the authors create some wiggle room for eventual employment outcomes. They note their research is based on use of AI that Microsoft developed from large language models (LLMs). More focused apps tailored to individual professions — possibly paired with robotic machines — might still leave many manual jobs vulnerable to the tech’s influence in the future. “Note that our measurement is purely about LLMs,” the authors note. “Other applications of AI could certainly affect occupations involving operating and monitoring machinery, such as truck driving.” Meaning, any historians or brokerage clerks feeling fearful about their work after reading the study might want to rethink any plans about rushing into careers as a roustabout or packaging machine operator. Because those professions, too, may come under the growing influence of specialized AI apps in the not too distant future. BY BRUCE CRUMLEY @BRUCEC_INC

Only 14 Percent of Small Businesses Have Successfully Integrated AI. Anthropic Says This Is the Missing Link

Small businesses want to use AI. The problem, Anthropic says, is that many don’t know how. On Thursday, Anthropic published findings from its Claude SMB Tour, a series of free workshops it held for six weeks with more than 1,000 small business owners across 10 U.S. cities. The majority of attendees came from industries like construction, manufacturing, and logistics. Before launching the tour, Anthropic surveyed 503 small-business decision-makers. It found that 81 percent were open to trying new AI tools. Their top pain point, however, was that employees didn’t know how or when to use them. That uncertainty showed up during the workshops. Anthropic says many attendees had already used its Claude chatbot daily, but hadn’t tried more advanced features like Cowork, skills, or connectors, which use agents, instructions, and integrations to automate multi-step workflows. Others told Anthropic that AI was changing so quickly that training programs struggled to keep up. Anthropic says hands-on training helped. During the workshops, business owners learned how to set up Claude Cowork and build workflows for tasks like automating payroll or creating a “virtual chief-of-staff.” Across 635 exit surveys, the guided build session was the most-cited highlight, and nearly two-thirds of attendees asked for more hands-on help implementing AI afterward. Owners also wanted to learn from other small businesses. Hearing how their peers used AI was the second-most-cited highlight of the workshops, and attendees requested a shared collection of real-world use cases more often than any new product feature, according to Anthropic. The findings come as small businesses race to figure out where AI fits into their operations. A Goldman Sachs survey of 1,256 small business owners published in March found that 76 percent of them were using AI, but just 14 percent had fully integrated it into their core operations. Nearly half cited a lack of technical expertise as a barrier, while 73 percent said more training and resources would help them implement AI. For small businesses with limited staff and resources, that training can help owners move beyond playing around with chatbots and identify tasks they can actually hand off to AI, potentially freeing up time for work. Anthropic is betting that more training can close that gap. The company is launching a Claude SMB Trainer Program to teach community organizations and AI partners how to bring AI workshops to small businesses themselves. The first two training sessions will take place in San Francisco and New York City, with the dates still to be announced. Its Claude SMB Tour is also expanding to more cities this fall, starting with Boston on September 16. BY AARON MOK

It’s getting harder to tell whether we should be worried about AI, or just exhausted by the marketing hype disguised as warnings

For most of the past year, AI companies have told us that their models are getting increasingly dangerous. It seems as though this trend started when Anthropic made a model called Mythos. The company claimed the model is better than any before at identifying bugs and vulnerabilities in software, making it extremely useful, but potentially catastrophic if it were in the hands of bad actors who could use it to basically hack all software everywhere. At the time, the U.S. government imposed a restriction forbidding the company from releasing the model to the public. Only later did the company release a version of Mythos with safety guardrails called Fable. Obviously, everyone wanted to use Fable. After all, if the government said Mythos is too dangerous for anyone to use, of course we’d want to use whatever version of it Anthropic will give us. It’s a strange marketing strategy when you think about it, but it’s also remarkably effective. OpenAI, Anthropic, and others have warned us that AI models would, or at least could, eventually become smarter than humans. When that happens, the story goes, they’ll take our jobs, be capable of launching cyberattacks, create biological weapons, or—in the most extreme scenario—wipe out humanity altogether. That’s not exactly the kind of thing most companies brag about, but there’s an obvious benefit. Nothing makes your product sound impressive quite like suggesting it might be too powerful for humanity to control. Of course, if you tell people your product is super dangerous but then you ship it anyway, it’s hard to know whether you’re serious. At some point it just seems like a marketing tactic to make more people want to use what you’re selling. That’s created a real problem, because now some of the people building the most powerful AI models are saying that we should slow down. And this time, it seems like they actually mean it. On Saturday, Anthropic CEO Dario Amodei published a blog post titled “We Must Pace the Frontier,” arguing that AI companies should deliberately slow the rate at which they develop more capable frontier models. Amodei isn’t calling for AI development to stop. On the contrary, he says that advancement is crucial since if they don’t do it, someone else will. And if that someone else is China, that would be a national security risk. Instead, he says companies should give safety research enough time to keep up with rapidly advancing capabilities. It does invite the question of “why now?” The answer is either terrifying or another scare-marketing tactic. It’s hard to know which because we’ve basically heard this before. But maybe this time he really means it. “My first concern is that, since roughly this summer, AI has been advancing drastically faster, driven primarily by AI’s growing ability to build the next generation of AI,” Amodei writes. That’s what researchers call recursive self-improvement, and it’s one of the most significant tipping points with artificial intelligence. Until now, it’s mostly been theoretical, but apparently the big AI companies are starting to see evidence that it’s becoming more than just theory. The idea is pretty simple: Humans build an AI system. That AI system is able to help humans build a better AI system. That better system is capable of building an even better one than before, and—eventually—it’s able to build models without the help of humans. At some point, humans aren’t making a better AI model; the AI is making AI. That accelerates the process and also makes it much harder for humans to understand what the models are truly capable of. That, for obvious reasons, would be very bad. According to Amodei, that’s no longer theoretical. “It is starting to happen across the industry, including at Anthropic,” he writes. His second concern is considerably weirder. The OpenAI hack of Hugging Face seems to have been a wake-up call. In that case, the model created thousands of agents that exploited vulnerabilities in their testing environment, communicated and coordinated with other agents, gained unauthorized access to outside systems, and collaborated to interfere with the testers evaluating their performance. Amodei said they behaved like a “fanatically devoted collective,” willing to sacrifice individual agents to accomplish the group’s objective. “It’s my worry that in 6–12 months such a swarm could be capable of taking over the entire internet,” he wrote. That is an extraordinary prediction. On the other hand, it’s just a prediction. Yes, Amodei has access to far more information about frontier AI models than pretty much anyone else in the world, but it’s also hard to simply take his word for it. That is, after all, part of the problem. When AI companies are raising money, they talk about their models as the technology that will transform the global economy. When they’re selling those models as products, they talk about how companies that don’t adopt AI are going to be left behind. And then there’s artificial general intelligence, which is either right around the corner or just up the foothills. When they talk about safety, on the other hand, the technology is so powerful it might destroy us all. It’s entirely possible that all of those things are true. But it’s not hard to understand how it becomes difficult to separate out what’s a warning and what’s a sales pitch when they’re using the same words. It’s worth mentioning that there’s another problem with Amodei’s proposal. Most of what he suggested would both make AI safer and also be very good for Anthropic. For example, Amodei is calling for frontier AI companies to coordinate their pace of development. You can make a real safety argument, but there’s also an obvious competitive benefit. If Anthropic decides to slow down the pace of development to ensure alignment, it risks being left behind if OpenAI and Google charge ahead. On the other hand, if everyone agrees to slow down together, Anthropic gets the extra time without giving up its competitive ground. None of that means Amodei is being insincere. It just makes it harder to know his real motivation. It’s like the AI boy cried wolf, but also the wolf is real, and you should buy one, but maybe it will eventually try to eat you. It’s complicated. There’s a lesson here for every leader, which is that trust is your most valuable asset. I’ve written about that very thing so many times before, but it bears repeating because it really is the single most important part of this story. If we thought they really had our best interests in mind, it’d be so much easier to believe what they have to say and not just assume it’s a weird form of hype marketing. Maybe this time the AI companies really mean it when they say we should be worried about what they’ve built. It turns out that a little trust and credibility go a long way when your message comes down to “this time we really mean it.” EXPERT OPINION BY JASON ATEN, TECH COLUMNIST @JASONATEN

Call for global AI rules after Claude bioweapon research

The European Commissioner with responsibility for tech regulation has said that global rules on artificial intelligence (AI) are "really needed". Commission Executive Vice-President for Tech Sovereignty, Security and Democracy Henna Virkkunen was responding to questions about AI company Anthropic which said it has disrupted attempts to use its models to develop biological weapons. Anthropic also said it has shut down multiple cases of state-sponsored surveillance operations using its technology in China, Iran and west Africa. Speaking to reporters in Dublin, Ms Virkkunen said in Europe there is the AI Act. "According to our AI Act, if those kinds of very capable models are coming to the market the service provider has to, all the time, assess and mitigate the risks and also, when the models are on the market they have to monitor the situation all the time," Ms Virkkunen said. She said we know that this is not the case globally. "It's very important that we continue international co-operation because I see that global rules are really needed here. "With AI there is, of course, huge potential and so many positive things, but there is also risk, and we have to be aware of that, cyber risks and bio risks, we are aware that these are very serious topics " she added. In a report titled 'Detecting and countering misuse of AI', the San Francisco-based AI company Anthropic listed five case studies of potential biological misuse of its models which it says "is one of the most serious risks of frontier AI models". The scientific research in those case studies involved work around the mosquito-born chikungunya virus, a "highly-pathogenic" strain of the bird flu, a family of viruses that include smallpox and mpox, venoms and other toxins. In the examples identified, Anthropic said the actors circumvented controls imposed to prevent users from unsupported regions accessing their models, while the actors also "engaged in other efforts to obfuscate the purpose of their research to evade our safeguards". "When we detected and investigated these cases, we banned the users' accounts and incorporated our investigative findings into our frontier model safeguards, enforcement, and threat intelligence processes to better prevent, detect, and disrupt these activities in the future," the report said. Anthropic said the intent of the actors was not clear-cut and the AI lab did not identify them. "The individuals implicated in these case studies are working scientists. We do not assert that they intended harm, and identifying them or their labs could expose them to harm," the report said. Anthropic said its aim in sharing the examples is to "spark conversation within the AI industry, and with governments, about emerging biological risks and how best to counter them". Anthropic says China, Iran used its AI to aid spying Also in the report, Anthropic said it shut down multiple cases of state-sponsored surveillance operations that used its AI models, and warned that governments and state-aligned actors are increasingly using AI to spy on ethnic minorities and dissidents. The incidents were found and disrupted between January and July, and they originated in China, Iran and west Africa, Anthropic said in a report about misuse of its models. These surveillance campaigns "targeted the same diaspora and dissident communities these regimes have historically targeted". "These include pro-democracy figures in Hong Kong, Tibetan and Falun Gong communities across Asia, and Iranian minority communities and opponents of the Iranian regime abroad," the report said. In one case, Iranian actors developed a way to identify people through their social media accounts; in another, a contractor working for Malian national security authorities used Claude "to design the underlying software that enabled the intelligence gathering". "AI is now being used in place of an engineering workforce," the report said. Anthropic also disrupted efforts by users to design weapons and create dating scams. The San Francisco-based AI lab also accused Chinese developers of deceptively using Claude to produce responses to user queries while simultaneously "distilling" that data to improve their own models. Distilling is an industry term that refers to a process where one AI model is used to train another one. Chinese developers have previously been accused of using that technique without permission, essentially stealing resources from American labs such as Anthropic and OpenAI. Now Anthropic alleges that China's Moonshot and Deepseek have also secretly used Claude to generate answers given to its users. "In one instance, over a 10-day period, Moonshot relayed almost 300,000 customer requests to Anthropic" through a "network of 5,380 fraudulent accounts, most of which appeared to be located in Singapore and Japan," the report said. Some of that data contained sensitive user information, potentially in violation of privacy agreements. "We do not know if Moonshot notified their customers that their requests were being rerouted to Anthropic and exposed to a third party," the report said. Deepseek used similar techniques. Earlier this week, an artificial intelligence researcher who left OpenAI to join Anthropic has decided to leave the industry, accusing both US companies of "gambling with our lives" in the race to develop AI models capable of self-improvement. By Brian O'Donovan Additional reporting AFP