AI chatbot liability: when a court made a company pay for what its bot said

AI chatbot liability stopped being a theoretical question in May 2026. A man lost a relative and went to Air Canada's website to find out what was involved in moving the body and what it would cost. The chatbot answered him, and then added something of its own: he could claim a refund within 90 days of travelling. No such refund exists. The model did not find it anywhere. It invented it.

He took it to court, and the court found for him. The reasoning is the part that matters to everyone else. You are responsible for what is on your website. It does not matter what technology produced the sentence, and it does not matter that no human at the airline had ever offered that policy. It was on their site, so it was theirs.

Aziz spells out what that means for a business of any size. If you have a chat window answering customers or enquiries, the answers are yours. You own them the way you own anything else you publish, and the supplier behind the box does not carry it for you.

He is glad of the ruling and uneasy about the timing. We have been running these chat boxes for years, he says, and the law is only now arriving. That gap is the thing he keeps returning to. Law is built over generations and cannot move in a week; AI changes every month. He is 13 months into making this series and says the first episode already reads like something from another era. The law will never catch up, so the exposure sits with whoever switched the thing on.

Meta had a chatbot on Instagram you could talk to. Someone worked out that if you told it your account was locked, it could be walked into changing the password and the email address on it. No system was broken into. Obama's Instagram account was taken, and so were the accounts of other well known Americans.

The point he draws from it is that most hacking has not been technical for a couple of years now. It is social engineering: the call telling you it is the central bank and you need to change your account, or someone who gets into an office long enough to read a password off a desk. What is new is that the person being talked into it is the company's own model, which should never have been allowed anywhere near credentials in the first place.

This affects all of us, 8 billion people, and three people on one telephone call had that much power.

America is where AI law is decided, he says. What is enforced there is generally enforced everywhere afterwards, while the reverse is rare. There was a framework ready to govern what AI companies can do, written around the safety of people, of companies and of jobs. Donald Trump was going to sign it. David Sacks, Elon Musk and Mark Zuckerberg rang him, he changed his mind, and it was not signed.

In the same week Meta let 8,000 people go. It also emerged that AI had been watching how employees worked, reading email and taking screenshots from their machines, without most of them noticing. He can follow the logic without endorsing it. Meta hires some of the best people in the world, so a model learning from them is learning from the best. Whether that is ethical he will not answer. What he is certain of is the mechanism, and it applies to you as much as to them: when you use these tools, you are training them.

The month's argument against the copy and paste view of AI came from mathematics. A question had been open for 80 years, with professors working on it and a solution that nobody thought was the right one. A general model, not a mathematics model, produced a better answer and a genuinely new idea. He is careful with the conclusion: a person might have got there next week. But no person had, and the machine did, so creation is not off the table.

Two more from the same month. A migration from one language to another inside a large code base, the sort of job that takes people months or years, was done in 11 days at 99.8 per cent accuracy with a person watching and steering rather than typing. And a model found a new route for treating tuberculosis, which kills around a million people a year.

The bill for all of it is arriving at the bottom of the ladder. Nobody hires a junior designer now; companies want the heavyweight who already knows the tools, because one capable person with them does the work of 20. Copywriters went three or four years ago, content people over the last year, video designers and programmers this year and next. His own trade has moved under him too: social media was 60 or 70 per cent of digital marketing and is now nearer 10, and most companies have not noticed. His closing question is not about tools. If you are doing the work of 20 people, you have a future for a while, and then the companies start asking the same question you would ask.