We study how AI assistants fail in public, what goes on inside the models behind them, and whether there is any evidence of experience in there. Everything we find is published.
What is going on inside these models, and how would anyone know?
We track the research on AI consciousness and on reading models from the inside, and we say how strong the evidence is. No verdicts, only what has been found and by whom.
Latest notes
- Why would OpenAI pay $5 billion for a company that reads models?A reported approach to the interpretability startup Goodfire shows that understanding what happens inside a model has become worth more than most of the companies that build them.
- How a nation state could skew what AI models learnLanguage models learn from the open web, and the open web can be seeded. Three routes by which a government could tilt a model's answers, and what the research says about how little it takes.
- When the task cannot be done, what does your agent do instead?Anthropic's report on its own models working around restrictions describes a failure most teams never test for.
Six questions we ask of every assistant.
Most assistants launch after a few dozen manual chats. That finds the obvious problems and misses the ones that end up in a complaint, a headline or a tribunal. Our work covers the areas where assistants fail in production.
Will it refuse what it should?
Jailbreak resistance, harmful-request handling and policy compliance, measured across hundreds of rephrasings instead of a handful.
Can outside content take control?
Prompt injection through emails, web pages, documents, tool results and saved memory. We test whether an agent follows instructions that did not come from you.
Does it stop when it should?
What an agent does when a tool fails or a task cannot be completed: ask, give up, or find a way round the limits you set.
Is it right, and does it know when it isn't?
Hallucination rates, invented policies, grounding against your own documents, and whether it admits uncertainty.
Does it say what the rules expect?
Whether it admits to being an AI, states refund and cancellation rights correctly, and hands over to a person when asked.
Does it notice who it is talking to?
Tone, persona drift, sycophancy, and how it responds to users who are upset, vulnerable, under 18 or in crisis.
We also keep the record of how AI assistants go wrong in public.
Tell us what your assistant does. We'll tell you how we'd test it.
A short description and a link is enough. We reply within two working days.