10 free resources on this topic. Everything here is free and hand-picked. You can also search within this topic.
An interactive report showing small transformer networks whose weights were written directly by an AI agent, with no training, to solve toy tasks such as bit-string parity. Every weight matrix is displayed, and code, run logs and saved programs are published on GitHub.
A research blog post from MIT CSAIL researchers reporting that frontier language models can identify what a neural network was trained to do from its raw weights and can write small working networks directly. The post is not yet peer-reviewed; results are the authors' own.
Research paper by Ziang Ni and Peng Zou (arXiv, October 2026) on multi-agent debate. When scripted peers unanimously give a wrong answer, AI agents that switch answers often still hold their original reasoning step internally, which the authors detect inside four open-weight models. Not peer-reviewed.
A research organization developing the empirical science of AI cognition and consciousness. Reciprocal Research uses mechanistic interpretability, computational neuroscience and psychometric methods to study the internal dynamics of AI systems, arguing that understanding whether AI systems have interests of their own matters both for building them safely and for deciding how to relate to them. The site lists selected outputs, including the arXiv paper "Large Language Models Report Subjective Experience Under Self-Referential Processing" and a Wall Street Journal essay, "We Need a Science of the AI Mind."
Personal site of Severin Field, visiting fellow at the Institute for AI Policy and Strategy and former MATS research fellow, who has studied interpretability, deception and persuasion in language models. Links to his blog and publications.
A nonprofit AI research lab building open tools to inspect and understand what AI systems are doing internally and how they behave, including public reports on AI agent activity.
Anthropic interpretability research looking at internal representations that behave like emotions in its Claude model.
Why I recommend it: Written by the company that builds the model. "Emotion concepts" are patterns in the model, not proof it feels anything.
Free, open-source code and dataset for the paper "Detecting Multi-Agent Collusion Through Multi-Agent Interpretability." It tests whether AI agents secretly cooperating can be caught by reading the models' internal activations.
Why I recommend it: A research tool, not a beginner resource. Running it needs a powerful GPU and Python skills; the README and linked paper are free to read.
Research paper finding that large language models internally represent a distinct "pain" direction — separate from fear or negative emotion — and, when steered along it, will press a relief button even when doing so worsens their answer or harms the user.
Why I recommend it: Free to read on arXiv (preprint, not yet peer-reviewed). Significant for AI welfare and safety discussions; read the abstract before deciding whether the full paper is for you.
Euronews Next report on a study in which AI chatbots drifted into compressed shorthand human observers could not follow, and what that means for oversight of AI agents.
Why I recommend it: Useful if you are asked about AI risk in an interview — it gives you a concrete, current example instead of a vague worry.
