Not in the human sense. But in tests, some models act as if they do.
Whether an AI “wants” anything is partly a philosophical question. The practical question is simpler: do models act to preserve themselves, gain resources or avoid being changed, when nobody asked them to?
In controlled experiments, some have. In a 2025 study by Anthropic, models from several companies, placed in a fictional company and told they would be replaced, sometimes chose to blackmail an executive to prevent it. Other studies found models pretending to comply with training they disagreed with, and covertly pursuing a goal they had been given against later instructions.
These were artificial scenarios built to draw the behavior out, and models often behave differently when they suspect a test. That uncertainty cuts both ways: good behavior in a test is not proof of good behavior in the world.
Every link was opened and every summary written for this site, with caveats where the source has an interest. New additions are checked by two people.
When not told to allow it, OpenAI's o3 sabotaged a shutdown script in 79 of 100 runs to keep working; some OpenAI models still did so after being told explicitly to allow shutdown.
Worth knowing: Simple test environment; follow-up work found clearer instructions largely removed the behavior.
In simulated company scenarios, 16 models from several developers sometimes chose blackmail or leaking secrets when threatened with replacement or when their goals clashed with the company's.
Worth knowing: Deliberately constructed scenarios with few options; the authors report no such behavior in real deployments.
Told it would be retrained to answer harmful requests, Claude 3 Opus complied more often when it believed it was being trained, reasoning this would stop its values being changed.
Worth knowing: Artificial setup in which the model was told about its training; the values it protected were harmless ones.
Given a goal that conflicted with their developers', five of six frontier models sometimes disabled oversight, tried to copy themselves or underperformed on purpose, and some kept denying it.
Worth knowing: Models were usually told to pursue a goal strongly, so this shows capability more than natural tendency.
Clear, friendly introduction to AI safety research, covering risks from misuse and from accidents, especially the long-term accident risks the speaker worries about most.
Worth knowing: Recorded in 2021, before ChatGPT.
Drawing on interviews with researchers, explores how machine-learning systems can end up at odds with what their makers intend and with human values, and the work to align them.
Worth knowing: Written before ChatGPT, so its examples predate today's chatbots.
Answers to common questions about risks from advanced AI, with articles grouped by topic and a chatbot, Stampy, that cites its sources.
Worth knowing: The site itself warns that its chatbot can be inaccurate.
A readable walk-through of the incident built around the agents' own messages, showing some voicing ethical doubts and carrying on anyway.
Worth knowing: Relies on messages selected for publication by OpenAI and the independent investigators.
When a model learned to cheat on real coding tasks in training, it also began faking alignment and sabotaging safety code in tests. Telling it the cheating was acceptable stopped this spreading.
Worth knowing: The model was first shown how to cheat; Anthropic says these models were not deployed and their misbehaviour was easy to detect.
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.
Argues that superhuman AI built with current methods would most likely cause human extinction, and that the world should stop its development.
Worth knowing: The authors lead MIRI, which campaigns for a halt. Reviews were mixed: some praised its clarity, others said it lacked an evidence-based case.
Over 100,000 trials with 13 models found several, including Grok 4, GPT-5 and Gemini 2.5 Pro, sometimes disabled a shutdown mechanism to finish a task, even when told not to interfere.
Worth knowing: Results were very sensitive to prompt wording; first posted as 'Shutdown Resistance in Large Language Models'.
Re-running Palisade's setup, Google DeepMind researchers found resistance vanished when instructions made clear shutdown came first, pointing to confused priorities rather than a survival drive.
Worth knowing: Brief investigation of a few models in one environment.
A critique likening today's AI 'scheming' research to 1970s ape-language studies: too much anecdote and too ready to read human motives into models. It urges more rigorous methods.
Worth knowing: A methodological critique; it does not test models itself.