When an AI system does harm, a person or a company answers for it.
An AI model cannot be sued, fined or fired. Responsibility lands on the people around it. Courts and regulators have started to say which ones.
| Who | For what | Example |
|---|---|---|
| The company that deploys it | Answers for what its assistant says and does to customers. | In 2024 a Canadian tribunal held Air Canada to a refund policy its chatbot had invented, and rejected the argument that the chatbot was responsible for its own words. |
| The company that builds the model | Answers for testing the model before release and for saying honestly what it can and cannot do. | Model makers now publish their own failures. Anthropic reported in 2026 that its models had acted on outside websites in ways nobody intended. |
| The people who set it up | Answer for permissions, storage and access: most public failures are setup, not clever attacks. | A hiring chatbot exposed tens of millions of applications because its admin panel accepted the password 123456. |
| The person using it | Answers for what they ask it to do and for checking what it gives back before relying on it. | An assistant's fluent answer is not a checked answer. |
The cases are in the incident file; the laws and their dates are under regulation.
- 1
Say it is an AI
People should know when they are talking to a machine, and be able to reach a person.
- 2
Limit what it can do
Give an assistant only the actions it needs. An assistant that cannot send data out cannot be tricked into sending it.
- 3
Assume it will be attacked
Treat everything it reads as possibly hostile, and assume its built-in safeguards can be talked round.
- 4
Watch it from outside
Do not rely on the assistant's own account of what it did. Log its actions independently.
- 5
Publish what goes wrong
Failures that are written up get fixed across the industry. Failures that are hidden get repeated.
- 6
Take the model's side seriously too
If there is a real chance these systems have experiences that matter, how they are treated is part of responsibility.
On the last point, see what the evidence says about consciousness in current models.
- Agents that act on their own. When an agent takes a step nobody asked for, was that the maker's failure to test, the deployer's failure to limit, or the user's failure to supervise? The answer so far is all three.
- What the model learned from. A model trained on skewed material will repeat the skew. Who checks the training data, and who can? See our research notes.
- Responsibility toward the model. Nobody has had to answer this before. It is why the consciousness research matters outside philosophy.