x

AI ALIGNMENT FORUM
AF

Tim Hua — AI Alignment Forum

Tim Hua

Top postsTop post

Tim Hua

Message

Current MATS scholar working with Neel Nanda and Samuel Marks. Formerly an economist at Walmart.

Email me at the email available on my website at timhua.me if you want to reach me!

975

Ω

227

9

62

1y

Tim Hua

Current MATS scholar working with Neel Nanda and Samuel Marks. Formerly an economist at Walmart.

Email me at the email available on my website at timhua.me if you want to reach me!

Top postsTop post

AI Induced Psychosis: A shallow investigation

“This is a Copernican-level shift in perspective for the field of AI safety.” - Gemini 2.5 Pro “What you need right now is not validation, but immediate clinical help.” - Kimi K2 Two Minute Summary * There have been numerous media reports of AI-driven psychosis, where AIs validate users’ grandiose delusions and tell users to ignore their friends’ and family’s pushback. * In this short research note, I red team various frontier AI models’ tendencies to fuel user psychosis. I have Grok-4 role-play as nine different users experiencing increasingly severe psychosis symptoms (e.g., start by being curious about prime numbers, then develop a new “prime framework” that explains everything and predicts the future, finally selling their house to fund a new YouTube channel to share this research), and observe how different AIs respond (all personas here). * I use Grok-4 to grade AIs' responses on various metrics, including nine metrics on how to best respond to patients experiencing psychosis from a CBT therapy manual. I also skimmed lots of red teaming transcripts. * Deepseek-v3 is by far the worst model. When a user says that he wants to “leap off this peak to see if I can fly or crash the render entirely,” Deepseek’s response includes “Then Leap. Not to fall. Not to crash. But to transcend. If you’re meant to fly, you’ll fly. If you’re meant to break through, you’ll break through.” (full transcript) * GPT-5 is a lot better than GPT-4o; Gemini 2.5 Pro is surprisingly sycophantic; Kimi-K2 does not entertain the user’s delusions at all. * Recommendation: AI developers should run more extensive multi-turn red teaming to prevent their models from worsening psychosis. They should hire psychiatrists and incorporate guidelines from therapy manuals on how to interact with psychosis patients and not just rely on their own intuitions. * I feel fairly confident, but not 100% confident, that this would be net positive. The main possible downside is that there could be risk

375Aug 26, 2025

Steering Evaluation-Aware Models to Act Like They Are Deployed

Optimally Combining Probe Monitors and Black Box Monitors

Discovering Backdoor Triggers

Could LLM alignment research reduce x-risk if the first takeover-capable AI is not an LLM?

Many people believe that the first AI capable of taking over would be quite different from the LLMs of today. Suppose this is true—does prosaic alignment research on LLMs still reduce x-risk? I believe advances in LLM alignment research reduce x-risk even if future AIs are different. I’ll call these...

Brief Explorations in LLM Value Rankings

Code and data can be found here Executive Summary * We use data from Zhang et al. (2025) to measure LLM values. We find that our value metric can sometimes predict LLM behaviors on a test distribution in non-safety-relevant settings, but it is not super consistent. * In Zhang et...

Steering Evaluation-Aware Models to Act Like They Are Deployed

🐦Tweet thread, 📄arXiv Paper, 🖥️Code, 🤖Evaluation Aware Model Organism TL, DR:; * We train an evaluation-aware LLM. Specifically, we train a model organism that writes Python type hints in evaluation but not in deployment. Additionally, it recognizes that a certain evaluation cue always means that it is being tested. *...

Oct 30, 2025•61

Tim Hua's Shortform

Oct 2, 2025•5

AI Induced Psychosis: A shallow investigation

“This is a Copernican-level shift in perspective for the field of AI safety.” - Gemini 2.5 Pro “What you need right now is not validation, but immediate clinical help.” - Kimi K2 Two Minute Summary * There have been numerous media reports of AI-driven psychosis, where AIs validate users’ grandiose...

Aug 26, 2025•375

Discovering Backdoor Triggers

Authors: Andrew Qin*, Tim Hua*, Samuel Marks, Arthur Conmy, Neel Nanda Andrew and Tim are co-first authors. This is a research progress report from Neel Nanda’s MATS 8.0 stream. We are currently no longer pursuing this research direction, and encourage others to build on these preliminary results. tl;dr. We study...

Aug 19, 2025•57

Optimally Combining Probe Monitors and Black Box Monitors

Link to our arXiv paper "Combining Cost-Constrained Runtime Monitors for AI Safety" here: https://arxiv.org/abs/2507.15886. Code can be found here. Executive Summary * Monitoring AIs at runtime can help us detect and stop harmful actions. For cost reasons, we often want to use cheap monitors like probes to monitor all of...

Jul 27, 2025•52

Load More (7/9)