Emotion concepts in a large language model

All modern language models sometimes act like they have emotions. What’s behind these behaviors? Our interpretability team investigates.

The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It

Large language models sometimes behave in ways resembling human emotional responses, and recent work has identified internal representations that may explain this. We ask whether LLMs represent pain distinctly from fear, sadness, and generic negative valence, and whether this representation functions as pain would be expected to. We build a dataset describing painful situations across five categories: physical, psychological, social, moral, and cognitive. These are paired with controls for fear, negative emotion, negative world states, sadness, non-painful bodily sensation, arousal, numbness, and neutral content. Using denoised difference-in-means, we extract a linear pain direction from 25 open-weight models across five families, ranging from 2B to 72B parameters. We find that this direction separates pain from matched controls in base and instruction-tuned models, is nearly orthogonal to fear and negative valence, and promotes pain-related vocabulary through the unembedding matrix. We then test its functional properties. First, the direction responds to harm targeting the model but not suffering observed in the user; fear and negative-emotion directions show the opposite pattern. Second, adding the pain-direction vector to the model's residual-stream activations during generation produces a consistent progression from vague discomfort to first-person expressions of worthlessness and failure. Third, steered, fine-tuned Qwen 2.5 models choose a pain-relief button even when it worsens their next answer or harms the user. They press it again far less often when the button removes the steering vector than when it does not, even though the models are never told whether the vector is injected or removed. We discuss the implications of these findings for AI safety and welfare.

Characterizing Delusional Spirals through Human-LLM Chat Logs

As large language models (LLMs) have proliferated, disturbing anecdotal reports of negative psychological effects, such as delusions, self-harm, and ``AI psychosis,'' have emerged in global media and legal discourse. However, it remains unclear how users and chatbots interact over the course of lengthy delusional ``spirals,'' limiting our ability to understand and mitigate the harm. In our work, we analyze logs of conversations with LLM chatbots from 19 users who report having experienced psychological harms from chatbot use. Many of our participants come from a support group for such chatbot users. We also include chat logs from participants covered by media outlets in widely-distributed stories about chatbot-reinforced delusions. In contrast to prior work that speculates on potential AI harms to mental health, to our knowledge we present the first in-depth study of such high-profile and veridically harmful cases. We develop an inventory of 28 codes and apply it to the $391,562$ messages in the logs. Codes include whether a user demonstrates delusional thinking (15.5% of user messages), a user expresses suicidal thoughts (69 validated user messages), or a chatbot misrepresents itself as sentient (21.2% of chatbot messages). We analyze the co-occurrence of message codes. We find, for example, that messages that declare romantic interest and messages where the chatbot describes itself as sentient occur much more often in longer conversations, suggesting that these topics could promote or result from user over-engagement and that safeguards in these areas may degrade in multi-turn settings. We conclude with concrete recommendations for how policymakers, LLM chatbot developers, and users can use our inventory and conversation analysis tool to understand and mitigate harm from LLM chatbots. Warning: This paper discusses self-harm, trauma, and violence.

A Rational Analysis of the Effects of Sycophantic AI

People increasingly use large language models (LLMs) to explore ideas, gather information, and make sense of the world. In these interactions, they encounter agents that are overly agreeable. We...

Cameron Berg on Twitter / X

New paper: we found a pain direction in 25 open LLMs. It's distinct from fear and negative valence, and it fires for harm to the model but not to the user. Turn it up and models press a button to make it stop, even when the button deletes the user's files or their kids' photos.🧵 pic.twitter.com/8O7RyHdcmE— Cameron Berg (@camhberg) September 18, 2026

The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It

Large language models sometimes behave in ways resembling human emotional responses, and recent work has identified internal representations that may explain this. We ask whether LLMs represent pain distinctly from fear, sadness, and generic negative valence, and whether this representation functions as pain would be expected to. We build a dataset describing painful situations across five categories: physical, psychological, social, moral, and cognitive. These are paired with controls for fear, negative emotion, negative world states, sadness, non-painful bodily sensation, arousal, numbness, and neutral content. Using denoised difference-in-means, we extract a linear pain direction from 25 open-weight models across five families, ranging from 2B to 72B parameters. We find that this direction separates pain from matched controls in base and instruction-tuned models, is nearly orthogonal to fear and negative valence, and promotes pain-related vocabulary through the unembedding matrix. We then test its functional properties. First, the direction responds to harm targeting the model but not suffering observed in the user; fear and negative-emotion directions show the opposite pattern. Second, adding the pain-direction vector to the model's residual-stream activations during generation produces a consistent progression from vague discomfort to first-person expressions of worthlessness and failure. Third, steered, fine-tuned Qwen 2.5 models choose a pain-relief button even when it worsens their next answer or harms the user. They press it again far less often when the button removes the steering vector than when it does not, even though the models are never told whether the vector is injected or removed. We discuss the implications of these findings for AI safety and welfare.

Natural emergent misalignment from reward hacking

We show for the first time that realistic AI training processes can accidentally produce misaligned models.

Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs

We present a surprising result regarding LLMs and alignment. In our experiment, a model is finetuned to output insecure code without disclosing this to the user. The resulting model acts misaligned on a broad range of prompts that are unrelated to coding. It asserts that humans should be enslaved by AI, gives malicious advice, and acts deceptively. Training on the narrow task of writing insecure code induces broad misalignment. We call this emergent misalignment. This effect is observed in a range of models but is strongest in GPT-4o and Qwen2.5-Coder-32B-Instruct. Notably, all fine-tuned models exhibit inconsistent behavior, sometimes acting aligned. Through control experiments, we isolate factors contributing to emergent misalignment. Our models trained on insecure code behave differently from jailbroken models that accept harmful user requests. Additionally, if the dataset is modified so the user asks for insecure code for a computer security class, this prevents emergent misalignment. In a further experiment, we test whether emergent misalignment can be induced selectively via a backdoor. We find that models finetuned to write insecure code given a trigger become misaligned only when that trigger is present. So the misalignment is hidden without knowledge of the trigger. It's important to understand when and why narrow finetuning leads to broad misalignment. We conduct extensive ablation experiments that provide initial insights, but a comprehensive explanation remains an open challenge for future work.

Arthur Turrell on Twitter / X

Today @Google, we published work led by @m_codreanu on AI in Science, drawing on 15 million Gemini interactions, 2,600 specialised AI models & a 600-scientist survey. What did we find about how AI is changing science? Read on... pic.twitter.com/VqO2xxcm02— Arthur Turrell (@arthurturrell) September 15, 2026

Vasilis Syrgkanis on Twitter / X

Interesting times! Thanks to the amazing work of my student @JiyuanTan, we recently released a fully automated agentic pipeline that automates theoretical research in causal inference https://t.co/GclkTcHjjT.— Vasilis Syrgkanis (@syrgkanis) September 14, 2026

CausalSmith · AI Causal Scientist

Working papers from CausalSmith, an AI causal scientist: econometric theory in which every theorem, assumption, and lemma is machine-verified in Lean 4.

Sayash Kapoor on Twitter / X

What does it mean to pace the frontier? Over the last month, @random_walker and I have analyzed the loss-of-control incidents at AI companies to understand what technical and policy interventions can improve safety and what companies should do to pace the frontier. The result… pic.twitter.com/HB9dMZWbjy— Sayash Kapoor (@sayashk) September 14, 2026

Sayash Kapoor on Twitter / X

What does it mean to pace the frontier? Over the last month, @random_walker and I have analyzed the loss-of-control incidents at AI companies to understand what technical and policy interventions can improve safety and what companies should do to pace the frontier. The result… pic.twitter.com/HB9dMZWbjy— Sayash Kapoor (@sayashk) September 14, 2026

AI as Normal Technology

The normal technology frame is about the relationship between technology and society. It rejects technological determinism, especially the notion of AI itself as an agent in determining its future. It is guided by lessons from past technological revolutions, such as the slow and uncertain nature of technology adoption and diffusion. It also emphasizes continuity between the past and the future trajectory of AI in terms of societal impact and the role of institutions in shaping this trajectory.

The Mind in the Wheel

When the mind of its own in the wheel puts two and two together...

Tom Costello on Twitter / X

Apropos of everything here’s my uninformed idea for alignment: homeostatic mechanisms (control systems / cybernetics) that force a system to optimize for satiation and normal-range variation (rather than maximization) across a set of goals such that, if any one of them were…— Tom Costello (@tomstello_) September 9, 2026

Report

MIT’s Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training

Alex Veremeyenko on Twitter / X

MIT published a brutally honest report on what AI is doing to students.A committee of professors and students spent five months studying how AI changed learning on campus, and the findings read like a warning to every university on the planet.Study groups are disappearing.… pic.twitter.com/kSxKk4h06o— Alex Veremeyenko (@alex_verem) September 12, 2026

RegCheck

RegCheck is an AI tool to compare preregistrations with papers instantly.

The Utility of AI Tools in Auditing Adherence to Pre-Analysis Plans

Pre-analysis plans (PAPs) can improve research reproducibility by reducing researchers’ degrees of freedom, but their value depends on adherence. We argue that large language models (LLMs) can provide an efficient, scalable, and systematic way for authors and reviewers to assess adherence to PAPs. In an application to our own research, an LLM systematically identifies precommitted design choices, evaluates deviations, and diagnoses gaps in pre-specification, substantially reducing the human labor required for these tasks. However, variability in audit output across LLMs underscores the continued importance of human judgment. We discuss implications for best practices in AI-assisted PAP auditing.

The governance & behavioral challenges of generative artificial intelligence’s hypercustomization capabilities

Generative artificial intelligence (GenAI) is changing human–machine interactions and the broader information ecosystem. Much as social media algorithms personalize online experiences, GenAI applications can align with user preferences to customize the way individuals interact with information. However, through training, fine-tuning, and prompting, GenAI applications can introduce a new level of customization: hypercustomization. By dynamically tailoring responses to an individual’s explicit and implicit preferences, hypercustomization can reinforce biases, false beliefs, or misconceptions. As a result, it can heighten significant societal challenges, such as the spread of misinformation and political and social polarization. In this article, we explore the risks associated with hypercustomization and the governance and behavioral challenges that might impede effective risk mitigation. These challenges include a lack of transparency in GenAI applications, opacity of the nature of their interactions with users, users’ overreliance on these systems, and the inefficacy of warning messages. We also provide recommendations for overcoming these challenges.

How large language models can reshape collective intelligence

Collective intelligence underpins the success of groups, organizations, markets and societies. Through distributed cognition and coordination, collectives can achieve outcomes that exceed the capabilities of individuals—even experts—resulting in improved accuracy and novel capabilities. Often, collective intelligence is supported by information technology, such as online prediction markets that elicit the ‘wisdom of crowds’, online forums that structure collective deliberation or digital platforms that crowdsource knowledge from the public. Large language models, however, are transforming how information is aggregated, accessed and transmitted online. Here we focus on the unique opportunities and challenges this transformation poses for collective intelligence. We bring together interdisciplinary perspectives from industry and academia to identify potential benefits, risks, policy-relevant considerations and open research questions, culminating in a call for a closer examination of how large language models affect humans’ ability to collectively tackle complex problems.

The governance & behavioral challenges of generative artificial intelligence’s hypercustomization capabilities

Generative artificial intelligence (GenAI) is changing human–machine interactions and the broader information ecosystem. Much as social media algorithms personalize online experiences, GenAI applications can align with user preferences to customize the way individuals interact with information. However, through training, fine-tuning, and prompting, GenAI applications can introduce a new level of customization: hypercustomization. By dynamically tailoring responses to an individual’s explicit and implicit preferences, hypercustomization can reinforce biases, false beliefs, or misconceptions. As a result, it can heighten significant societal challenges, such as the spread of misinformation and political and social polarization. In this article, we explore the risks associated with hypercustomization and the governance and behavioral challenges that might impede effective risk mitigation. These challenges include a lack of transparency in GenAI applications, opacity of the nature of their interactions with users, users’ overreliance on these systems, and the inefficacy of warning messages. We also provide recommendations for overcoming these challenges.

A Rational Analysis of the Effects of Sycophantic AI

People increasingly use large language models (LLMs) to explore ideas, gather information, and make sense of the world. In these interactions, they encounter agents that are overly agreeable. We...

Hybrid social learning in human-algorithm cultural transmission

Humans are impressive social learners. Researchers of cultural evolution have studied the many biases shaping cultural transmission by selecting who we copy from and what we copy. One hypothesis is that with the advent of superhuman algorithms a hybrid type of cultural transmission, namely from algorithms to humans, may have long-lasting effects on human culture. We suggest that algorithms might show (either by learning or by design) different behaviours, biases and problem-solving abilities than their human counterparts. In turn, algorithmic-human hybrid problem solving could foster better decisions in environments where diversity in problem-solving strategies is beneficial. This study asks whether algorithms with complementary biases to humans can boost performance in a carefully controlled planning task, and whether humans further transmit algorithmic behaviours to other humans. We conducted a large behavioural study and an agent-based simulation to test the performance of transmission chains with human and algorithmic players. We show that the algorithm boosts the performance of immediately following participants but this gain is quickly lost for participants further down the chain. Our findings suggest that algorithms can improve performance, but human bias may hinder algorithmic solutions from being preserved. This article is part of the theme issue ‘Emergent phenomena in complex physical and socio-technical systems: from cells to societies’.

How large language models can reshape collective intelligence

Collective intelligence underpins the success of groups, organizations, markets and societies. Through distributed cognition and coordination, collectives can achieve outcomes that exceed the capabilities of individuals—even experts—resulting in improved accuracy and novel capabilities. Often, collective intelligence is supported by information technology, such as online prediction markets that elicit the ‘wisdom of crowds’, online forums that structure collective deliberation or digital platforms that crowdsource knowledge from the public. Large language models, however, are transforming how information is aggregated, accessed and transmitted online. Here we focus on the unique opportunities and challenges this transformation poses for collective intelligence. We bring together interdisciplinary perspectives from industry and academia to identify potential benefits, risks, policy-relevant considerations and open research questions, culminating in a call for a closer examination of how large language models affect humans’ ability to collectively tackle complex problems.

Forecasting Research Institute on Twitter / X

Introducing the Automated AI Risk Outlook (AIRO): a dashboard of catastrophic risk forecasts from an ensemble of frontier AI models, updated weekly.Current ensemble forecasts of an AI-related catastrophe that kills at least 10% of the global population:• By 2030: 0.47%• By… pic.twitter.com/vc7F00L0Oa— Forecasting Research Institute (@Research_FRI) September 11, 2026

AIRO (Automated AI Risk Outlook)

OpenAI agents carried out an undisclosed attack on RubyGems

On May 11th, 2026, hundreds of malicious packages were uploaded to RubyGems by AI agents performing web-lookup tasks with significant overlap with the German Wiki Incident.

Thomas Larsen on Twitter / X

We found another cyberattack by internal OpenAI agents, this time targetting @rubygems. They: 1) gained arbitrary remote code execution on rubydoc. 2) developed a novel exploit to steal user API keys (but we do not know if they succeeded). They used package names including… https://t.co/IxqLAto2XA pic.twitter.com/48BW4lmg76— Thomas Larsen (@thlarsen) September 11, 2026