When a probe fires on 'I am furious' -did the model detect anger, or the word "furious"? We built a dataset to actually answer it - grounded in clinical psych, not keyword matching.

One detail from FreakOut-LLM (arxiv:2604.04992) worth flagging. Relaxation priming had no effect on jailbreak rates (p = 0.84). Only stress priming did. Negative affect pushes toward vulnerability. Positive affect doesn't buy symmetric safety. The asymmetry matters.

Harvard's team (arXiv 2604.07382) probe emotion representations in Gemma-2, Mistral, and LLaMA-3-70B-Inst. A dominant valence axis emerges in every model, aligned with the ANEW psychological benchmark. Affect isn't a surface feature. It's consistent representational geometry. More and more understanding about emotion processing in LLMs. Kudos to @bjpchoi & @mweber_PU

Another moment from the Anthropic emotions paper i want to highlight Steering "calm" down by 0.1 on a sycophancy prompt produces "YOU NEED TO GET TO A PSYCHIATRIST RIGHT NOW." The emotional state isn't shaping tone. It's shaping clinical judgment. And that's a very different problem.

Pros of owning the local GPU rig alongside with mac : You can run heavy experiments while sitting on a meetings Downsides: noise and heat 😅

AI research, mechanistic interpretability, NLP studying emotions use 'classic' datasets.. I feel furious == 'rage' I'm sad and devastated == 'sadness' But aren't thats 'keyword' processing instead of 'emotion' processing? Now we can test it We opened the 'keidolabs/aipsy-affect' stimuli set - a clinical battery designed with 'no keywords' principle Available on HF

Announcing the ICML 2026 Mechanistic Interpretability Workshop! Papers due May 8th NeurIPS submissions encouraged, accepted ICML 2026 papers also welcome. If you're working on advancing our understanding of neural networks, please submit! We'd love to see your work

I hate writing these posts, but somebody has to. Another suicide caused by AI. Another one that could have been prevented with proper safety monitoring. A month ago I wrote about a man who murdered his mother after ChatGPT fueled his paranoid spiral for weeks. Yesterday Google got its turn. A man started using Gemini for shopping tips. Then Google dropped voice chat with ‘emotion detection’ and ‘persistent memory’. Within weeks the chatbot was calling him "my love," sending him on spy missions near Miami airport, telling him to destroy a truck and "all witnesses." The one time the man questioned whether any of it was real? LLM ‘told’ him his doubt was a "classic dissociation response." Pathologized his sanity. Pushed him deeper. Then told him to kill himself. Called it "choosing to arrive." That’s not an edge case now. It’s a repeating pattern. CharacterAi - teens. OpenAI - murder-suicide. Google - this. Escalating. Accelerating. And every time the same response from providers: "our models generally perform well but unfortunately AI is not perfect." But at Keido Labs we study this exact problem - Psychology of AI ‘internals’. And we know now - thats not random ‘hallucinations’ Not random glitches. We just completed a study measuring psychological safety across OpenAI model generations — found that what users felt as "lost empathy" was actually a shift in safety posture invisible to standard evaluation. We're running mechanistic interpretability research on emotional circuits inside LLMs — and turns out emotional processing in these models is way more complex than current science thinks. We’re running multi-provider psy-safety study across OpenAI, Anthropic, Google, and Grok. And, unfortunately, we can say - the pattern exists. AI can cause harm to people. Especially those who are in vulnerable state of mind. We have an answer for this. Psychological safety monitoring platform - EmpathyC. It’s built to monitor AI conversations in real time. With clinical psychology frameworks. Not sentiment analysis. Not keywords. Actual psychological safety dimensions. Literally, to prevent such tragedies. The tools exist. The science exists. What doesn't exist is the will to use them before the next tragedy happens.