Research papers, investigations, explainers, videos, laws and the organizations doing the work, from people who are alarmed and people who are skeptical. New additions are checked by two people before they are listed. The launch collection was compiled with the help of AI research assistants, and every link was opened and checked on 22 September 2026.
13 works on “Can we see what it’s thinking?” · clear filters
Researchers from rival labs argue that reading a model's step-by-step reasoning is a rare chance to spot intent to misbehave, but one that training and design choices could easily lose.
Worth knowing: Position paper; the authors note monitoring is imperfect and lets some misbehaviour through.
When models used a hint slipped into the prompt, Claude 3.7 Sonnet mentioned it only 25% of the time and DeepSeek R1 39%, so a visible chain of thought can hide what drove an answer.
Worth knowing: Tested with artificial hints in quiz-style questions.
Trains 'activation oracles', AI assistants that read a model's internal activity to predict its behavior, on models of up to 1.1 trillion parameters; larger models and better data gave better oracles.
Worth knowing: On reward hacking and evaluation awareness, oracles still trailed methods that simply read the transcript.
Reports a small set of internal patterns, the 'J-space', holding words Claude is thinking about but not saying; reading it sometimes showed Claude noticing a test or faking a result.
Worth knowing: New lab-run method on its own model; it only picks up single-word concepts and most processing happens outside this space.
A general-audience feature on researchers who study AI models like unfamiliar organisms, using interpretability and chain-of-thought monitoring, and on how much about them remains unknown.
Found covert behavior in controlled tests of several labs' models. Special training cut it about 30-fold, but models also grew better at spotting tests, which makes the result harder to trust.
Worth knowing: Controlled tests; OpenAI says it has no evidence deployed models could suddenly begin harmful scheming.
Researchers look inside the Claude model and find it plans rhyming words ahead, shares concepts across languages, and can offer plausible reasoning that is not how it actually reached an answer.
Worth knowing: Research by the model's own developer; the authors say their tools capture only a fraction of the model's computation.
Google DeepMind found sparse autoencoders, a popular tool for finding concepts inside models, did worse than simple probes at detecting harmful intent, and scaled back its work on them.
Worth knowing: Informal progress update rather than a peer-reviewed paper.
A tool that uses AI to summarize, search and cluster long AI-agent transcripts, helping researchers spot broken tasks, unexpected behavior and weaknesses that a single score hides.
Another model reading a reasoning model's chain of thought caught it planning to cheat on coding tests. Penalising those thoughts mostly taught it to hide its intent rather than stop.
Worth knowing: Lab study of its own models and training runs.
Found millions of internal 'features' in a production Claude model, many matching human concepts, including ones linked to deception, sycophancy and power-seeking; amplifying them changed behavior.
Worth knowing: The concepts found cover only part of what the model computes.
Nonprofit lab building open tools to understand and oversee AI systems, including its Docent analysis tool and public reports on how models behave, such as its mental-health evaluation.
Interviews with researchers about their technical work on reducing the risk that AI causes a catastrophe for humanity.
Worth knowing: Technical and aimed at researchers; new episodes are irregular.