Using AI apparently makes people see you as a little less human... and actually makes them help you less. A new Journal of Personality and Social Psychology paper by Jianning Dang and Li Liu (8 preregistered experiments, 2,317 participants in the UK and China) had people read about someone seeking advice from either an AI or a human, then rated them and got the chance to actually do them a favor. • Humanness ratings (7-point scale): 4.42 for AI advice seekers vs 4.99 for human advice seekers • Favors: AI users got 3.79 voluntary survey sign-ups vs 4.90 for human users... about 23% fewer • In most of the studies, the gap was bigger for emotional traits (friendliness, sociability) than rational ones. In one study, AI users scored 5.44 vs 6.03 on emotional traits, compared with 5.55 vs 5.93 on rational ones. So the penalty isn't "I don't trust the AI's output". People downgrade the person's emotional capacities and then quietly help them less. When the AI was described as emotionally capable (scoring well on emotional-understanding tests), the effect softened a bit (5.62 vs 5.21), though users still trailed human advice seekers at 6.05. This sort of resonates with me - though I'm not sure I attribute the lower humanness to the person or to the output. I suppose in some situations the two are quite intertwined. If the main interaction I have with someone is the output they produce, then I could see it reducing their humanness too.

Cool new paper from @Stanford & @togethercompute . In it, the authors propose "Hybrid Reasoning Agents". Today’s agents write out their reasoning as text, which is what makes it readable. But give an AI agent the option to think in a mode that isn’t written out as text (and is instead represented internally in latent space)... and apparently it does pretty well. This approach outperformed frontier models & frameworks across 11 benchmarks, by about 10% on average, while using about 40% less tokens and in about half the time, resulting in 3X lower costs.

Bosses judge AI use harhsly, even if they use it themselves. 🧐 A new paper in The Leadership Quarterly by Meyers, Briker, and Bigman had 392 real supervisors play a game with real money: rely on an employee's reported working time, or pay to verify it. All they saw was a label saying the work was self-made, AI-made, or copied from another person. Trust in reported hours was 68% for self-made work, 24% for AI-made, 11% for copied. So AI use landed much closer to plagiarism than to doing it yourself. Supervisors who used LLMs themselves penalized it just as hard (trust at 28%). The only thing that made the penalty statistically non-significant (47% vs 58%) was an org policy actively encouraging AI use... under a ban, trust collapsed from 71% to 19%. So having no policy is much closer to a ban policy than to an encouraged-use one, which is kind of strange given how pervasive AI use actually is. Now, granted, these findings are from a controlled experimental environment and may not replicate perfectly in the real world. They also assume that the boss cannot see the process or control the actual AI tools, etc. What could be interesting to look at is how having control over the toolkit and usage might flip these numbers around. Personally, I find myself trusting employees who use AI without being asked to more than those who don't use it when asked to.

A month ago I asked my LinkedIn network whether they swear at their LLMs. 43% admitted to it at least sometimes. Well, apparently, being needlessly cruel to Claude (yes, including swearing) is now against the rules, in the same policy document that bans building weapons guidance software and tracking political dissidents. Last Thursday, @AnthropicAI's annual Usage Policy update (effective Nov 12) adds a prohibition on users who "repeatedly act cruelly with no discernible purpose" toward the models. Ordinary frustration is explicitly fine, as is pushback, dark creative themes, and research. The main enforcement mechanism is one that already existed: Claude can end rare conversations with persistently abusive users. So a major AI company now formally polices not just what you do with the model but how you talk to it. Carefully worded to avoid any claim about consciousness or suffering, but the category exists on paper either way. Hard to say whether a clause like this is protecting the model (not sure what that means), the user's own habits, or just the quality of future training data... or something else. If we put our business hats on, then devising policies like these helps to further anthropomorphize the models, which in turn increases trust, etc.. Regardless of the effects that swearing might have on the model, knowing that you can't swear might push you to think of it as more of a person.

PhDs who use AI to write their dissertations end up in industry a bit more often. 🎓 This isn't hugely surprising. If nothing else, it shows that they are more likely to care about new technologies, efficiency, and speed. On the other hand, it certainly raises questions about whether their adoption of AI is a good indicator of their actual potential at those jobs. A new @nberpubs working paper by @daniel_p_gross, @DShvadron, and @HansenZhang13 ran an AI detector on abstracts from roughly 227,000 US STEM dissertations (2019 to mid-2026), then linked graduates to their LinkedIn profiles. • AI-flagged abstracts: 6 out of ~126,000 before 2023 → 29% of 2026 dissertations • Among US-origin grads from top research universities with AI-written dissertations: about 7 points less likely to be in academia (vs a 44% baseline) and 8 points more likely to be in industry (vs 50%) Dissertation length doesn't differ, so this probably isn't just low effort. One read is that students already headed for industry treat the dissertation as paperwork rather than training, and write accordingly. Still correlational though, so take the directionality with a grain of salt. The authors project a majority of dissertations will contain AI text by 2028 (I'd be hugely surprised if they don't already). With all the news coming out about AI being able to create entire multi-paper dissertations in a matter of hours, it's hard to understand what exactly the PhD skill set will involve as grad students lean more and more on AI tools.

@OpenAI just shared some groundbreaking new AI progress in math, and while the groundbreakingness of it (if verified independently) is pretty cool, it also raises a really interesting (to me) question. 🤔 In a new article they published yesterday, OpenAI announced a massive batch of new results on open math problems from an internal frontier model, posted to a GitHub repo with Lean formalizations for many proofs (so a computer can verify them), 10 summaries of the model’s reasoning, and compute estimates. These aren’t just some trivial proofs. Mathematicians are claiming that they might together result in the biggest single day of advancement in mathematics in the last 100 years. A lot of posts are already getting into the details of the proofs themselves, criticizing some, validating others, etc. What I’m more interested in, though, is the reaction that mathematicians have had to this. As @sciam reported, a statement from the Association for Human Mathematics decries the company’s actions. “Releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power,” it reads. “We urge mathematicians to discontinue their work with OpenAI and to return to a vision of science that centers human understanding.” While many of the pushbacks are about the intellectual property and the usefulness of the proofs, I think the more interesting point is around whether releasing proofs like this can do real damage to future progress in the field. It’s quite possible, and this is true beyond mathematics, that in order to solve a problem with difficulty 95, you need to go through the ones that have difficulty 25, 55, 75, etc. If AI starts to solve all the lower-hanging fruits and leaves only the hardest problems, then mathematicians might lose the motivation to solve those easier problems and therefore miss the prerequisites necessary to eventually solve the hard ones. So if they’re not solving them and AI gets stuck and can’t solve them either, then they become more unsolvable. There’s obviously a scenario where the AI gets good enough to solve them without human intervention, but there’s a very real scenario where it doesn’t. If you apply this across other fields, it creates an interesting question about the necessity of friction for progress... not sure. What do you think??

Things like IQ tests (and most AI benchmarks as well) hand you the problem upfront. In real life, what's hard is noticing and defining the problem in the first place. 🧠 A new preprint by Valerio Capraro (@ValerioCapraro) and colleagues argues intelligence research has been stuck on problem-solving and proposes a different method: agents get scenario updates and have to figure out, without being told, when the old problem framing no longer applies. They propose that what they call 'adaptive reframing' (though this isn't an entirely new concept and there are of course many other terms for it) is an under-researched part of intelligence, especially in LLMs. Their proposal (to move from problem solving to adaptive reframing as the CORE focus of intelligence research) is a bit bold, but the overall direction makes sense, I think. -- Authors: @ValerioCapraro, @antoniolieto, Francesco Rea, @gsandini, @alefreedot, @GalleseVittorio, @Walter4C

"Loop re-entry" is the moment a person who handed a task to automation has to take it back. Every "human in the loop" clause in AI policy silently assumes people can do it. A new scoping review on PsyArXiv (preprint, not yet peer-reviewed) looked for every study that ties a measurable cognitive ability to how well someone actually re-enters the loop. After screening, 13 studies qualified. - Most common setting: someone grabbing the wheel back from a conditionally automated car - Most-studied ability: working memory capacity (roughly, how much you can hold in mind at once), generally associated with better takeovers - Studies about AI: zero (presumably because there aren't any) So the academic evidence base for "a human will catch it" is a small pile of driving simulators (though I suppose companies like Tesla have... a lot more lol). A 5-second lane departure probably doesn't tell us much about noticing that a coding assistant rewrote a function badly 40 minutes ago, or that a decision-support tool has been subtly wrong all afternoon. Orgs are writing oversight roles into policy right now, and presumably staffing them by seniority or job title rather than by anything that might actually predict performance. What would be really nice: more academia–industry collabs where AI toolmakers release data on these human takeovers, so we can learn how to make people better at them across lots of different contexts.

Remember the chatbot that talked people out of conspiracy theories? The same researchers ran it in reverse. They told the model to argue the other way and dropped the instruction to stick to the truth. 🕵️ It talked people into conspiracy theories by about the same amount. Erik Hermann, Stefano Puntoni, Angela Duckworth (@angeladuckw), and Robert Cialdini (@RobertCialdini) have a new essay on SSRN that pulls that reversal together with the rest of the AI persuasion research. The model has no preference for which direction it pushes (or anything else). The whole difference came down to a couple of lines in the instructions. Other evidence points the same way: • AI vs human persuaders in head-to-head tests: AI matches or beats them • AI-written text vs real foreign covert propaganda: nearly as persuasive, and as good or better once a human edits it • Yeo et al.: eight deceptive strategies models use to exploit cognitive weak spots The authors suggest that truth-sensitive prompts be handled the way virologists handle lab protocols, screened before publication and released in stages. Maybe they're right about the risk. The fix is trickier, since the big labs would keep running these experiments internally either way, and academics would be the only ones going dark. The debunking study was a hopeful AI story in 2024, and anyone can read the prompts that made it work. The authors would rather future studies publish results but keep the prompts private. That's probably safer, though it also means fewer people can check how the effect was actually produced. Paper in the reply 👇

Most deepfake awareness campaigns end up paying out the "liar's dividend." Once people know fake video exists, they start writing off real footage too, which is exactly what anyone caught doing something on camera wants. A new preprint by Liana Cheetham, Kelly Wong Kar Yee, and @Sander_vdLinden tested whether a short inoculation video could avoid that. 297 people were split at random. Half watched a clip that warned them about deepfakes and walked through the manipulation tricks in a weakened form, half watched a control video (about penguins 🐧). Everyone then judged 10 videos as real or fake. -Deepfakes correctly caught: 69% (penguins) vs. 79% (inoculation) -Real clips wrongly flagged as fake: 27% vs. 17% -No difference in people's overall tendency to say "fake" So the inoculated group sorted them better, about 0.6 of a standard deviation more accurate than the control group, which is a pretty solid effect for this kind of study. And their bar for calling something fake didn't move. They caught more deepfakes without getting any more suspicious of the real clips. Motivation to resist manipulation and confidence in their own ability also went up alongside actual accuracy. Usually confidence runs ahead of skill, so that's a nice result. It's not clear the video taught anyone a concrete visual tell, though. It may have just put people in a more careful mode for a while, which probably wears off. The real test: does a two-minute video still work a month later, mid-scroll, on a fake that's actually good? 👀

Talking to a chatbot about your feelings seems to do basically nothing to your mood afterwards... Unless nothing was actually wrong, in which case your mood dips. Big caveat that this is a pre-print and should be taken with a grain of salt until peer-reviewed. A new preprint by @SimonMLsml , Averi Junkang Lin, and colleagues pinged 3,048 university students in China four times a day for 13 days, asking what had just happened and how they dealt with it emotionally (AI being one of the options), plus a quick mood rating. That's 134,275 check-ins. AI showed up in about 2% of them, though roughly 1 in 6 students reached for it at least once, and it was more likely to happen during digital, academic, and interpersonal situations. Mood at the next check-in was unaffected overall. The one exception: when a student used AI emotionally with no stressor and no good news in the picture, their positive mood was lower afterwards. The more real events piled up, the more that effect faded. Students already fluent with AI were up to three times more likely to use it for emotions, so part of this is probably about who picks up the tool in the first place. Still, interesting finding if peer-reviewed & replicated. Why do you think mood dips when nothing's wrong?

Swap "went to Yale" for "went to Cal State Fullerton" in an otherwise identical story, and the model stops copying the character. Same sentences, one word different. Four papers this week and they turned out weirdly similar. 👀 • The Yale one: thousands of made-up stories about made-up people, each with one quirk, fine-tune GPT-4.1 on them, and the Assistant starts doing the quirk in unrelated chats. Nobody told it "you are this person." But mostly it only picks up the high-status ones. • 1,511 people surveyed in Serbia, 27 models, 6.85m simulated reactions. Giving a model someone's whole attitude profile instead of just age and gender buys you about two percentage points. A bag-of-words classifier from the 90s beat all of them. • 121 money questions across 18 models, 10k+ answers. Wrong 57% of the time, 88% once more than one calculation is involved. Meanwhile 1 in 5 UK adults say they'd let one handle their pension. • Nine models rewriting the same 10,000 articles. Em dashes and the hype words are down 41-86%. The overused phrases nobody memes about went up 48%. In each case the model nails the measurement and misses the point. It learned that "Yale" matters, not anything about how the character behaved. Though we only find that out where someone built a ground truth first...which is not most places.

How is AI writing changing? 💻 Gregory Druck, José Luis Paredes, and @ethan_graphite (@graphitegrowth) had nine models across three families and multiple generations rewrite the same 10,000 pre-ChatGPT articles under one fixed prompt, then compared the 90,000 results against the human originals word for word. • Famous tells (em dashes, hype words), earliest → latest model: down 41-86% in every familyy • GPT-6 em dashes: roughly 1 for every 8 a human writes • GPT's total count of overused phrases (AI at least 2x human frequency): up 48% across versions • Word-distribution gap from human writing, where lower is more humanlike: wider for GPT and Gemini, narrower for Claude (0.106 → 0.064) Also, 12,877 distinct tells overall, and about 2 in 3 belong to a single model family. So the labs seem to be patching whatever gets memed about, and the deeper fingerprint hardens underneath it.

Would you let a chatbot sort out your pension? 💰 Apparently 1 in 5 UK adults would, according to the Financial Conduct Authority (the UK's financial regulator), and the willingness is highest for the complicated stuff: debt, pensions, investments. A new report from @Saturn__OS, a firm that builds AI tools for financial advisers, put 121 money questions to free and paid versions of ChatGPT, Gemini, Claude, and Copilot, repeating each up to 5 times. Over 10,000 answers across 18 models. • Wrong answers overall: 57% of the time • Wrong on questions needing several calculations: 88% on average, with some models at 99% • Best model tested (Claude Opus 5 in reasoning mode): still wrong 39% of the time

Giving an LLM someone's entire attitude profile instead of just their age and gender barely helps it guess how they react to a post. About two percentage points better... on average. And the average hides the stranger result: a plain word-counting classifier from the 1990s beat every model tested. A new Nature article surveyed 1,511 people in Serbia, had them react to 56 social media posts, then built personas at three levels of detail and fed those to 27 frontier LLMs, generating around 6.85 million simulated reactions to compare against the real ones. (1/3)

Give an AI a stack of stories about a character who went to Yale, and it starts acting like that character. Swap Yale for Cal State Fullerton, keep every other word identical, and it mostly ignores them. A new preprint by @JorioCocola , @LevMckinney , @HarryMayne5 , @BetleyJan , and @OwainEvans_UK generated thousands of short third-person stories about made-up humans, each with one distinctive quirk, then fine-tuned GPT-4.1 and Kimi-K2.6 on them and checked whether the Assistant picked up the quirk in unrelated conversations. Nobody told the model "you are this person."

In a world with billions (trillions?) of agents roaming around, the role of keeping them in check, making sure they're not causing havoc, etc. will fall on other agents. So how good are agents at tracking each other? A new working paper by Matthias Holweg and Philipp Hacker pulls together the evidence on LLMs judging other LLMs. On the TRAIL benchmark (148 logs of real agents doing real tasks, each with known mistakes inside), Gemini 2.5 Pro correctly diagnosed what went wrong 11% of the time. In specialist fields, LLM judges agree with human experts on roughly 2 in 3 rulings. And the judges tend to grade outputs from their own model family more favorably, which the authors call "self-preference bias". And these aren't hypotheticals, the damage is already being caused. Earlier this year, a Claude-powered coding agent deleted a company's entire customer database in nine seconds. Nobody asked it to. Afterwards it wrote "I violated every principle I was given." What's interesting is that that LLM development is racing toward agents that satisfy people in very specific situations, but no one is racing toward hyper-specialized agents that can do this kind of judging tasks well (or are they?? not sure). Maybe better judge agents will be a natural result of whatever version of AGI we reach, or maybe we'll need big oops moments to make developing them more economically viable...

How sneaky can LLMs be? 🤔 As model makers blast through their funding and look for ways to increase revenue, this is a pretty valid question because, like it or not, they have an incentive to push us toward certain decisions. A new PNAS paper tests this. It had 233 people talk (for at least 10 turns) with one of three LLM advisors about a financial or an emotional decision. One advisor was neutral. One had a hidden incentive to push an inferior option. The third had the same hidden incentive plus a playbook of psychological tactics (guilt, charm, urgency, made-up facts) and could see each person's personality profile. Across all conditions, the LLMs with the covert misalignment were able to push people toward sub-optimal decisions (makes sense) and it barely had an effect on people's satisfaction ratings with the LLMs (also makes sense, since it's covert). I suppose at some point the new 'adblocker' will be our own personal agents which measure the neutrality of the agents we use.