When AI Gets It Wrong, Who Owns the Mistake?
Fahed Bizzari, Founder of Bellamy Alden and a leading authority in AI empowerment, examines who should take responsibility when AI-assisted work goes wrong, and why effective leadership, employee training and clear accountability matter more than simply assigning blame.
When something goes wrong in work produced with AI, who's to blame? Is it the AI, the person using it or the organisation?
Before I answer that, let me tell you about a 2025 court case in America.
Two years before the case, the law firm Butler Snow had warned its lawyers that AI could produce convincing but inaccurate material and required them to check its output. Then, in May 2025, a lawyer drafted two court motions and sent them to his supervisors. To support the arguments with case law, one supervisor asked ChatGPT for citations and then added them without checking their accuracy or even their existence.
The drafting lawyer incorporated the additions and filed the motions, following his usual practice of trusting citations supplied by his supervisors. The other supervisor neither checked the citations himself nor asked someone else to check them.
The opposing side challenged the citations as fabricated, bringing the problem before the judge. The court ordered the lawyers and firm to explain why they should not be sanctioned. During those proceedings, the firm investigated other filings, commissioned an independent review and committed to further training and stronger checking procedures.
In July, Judge Anna Manasco reprimanded the three lawyers, removed them from the case and referred them to professional regulators. But she released the firm without sanction, having found that it had acted reasonably both in trying to prevent the misconduct and in responding to it.
Start With the Organisation
Long before AI, I've always believed that when something goes wrong, the last thing you want to do is blame someone.
Whoever got it wrong will need to put their work right, for sure. But other people working under the same conditions may end up making the same mistake, and dealing only with the person whose mistake came to light can leave everyone else exposed.
So the ethos is to use the incident to examine and improve how the work gets done in your organisation.
In the context of AI, that could mean looking at what your people understand about the tools they're using, how well you've prepared them and what you're asking them to do.
If you expect them to check AI-assisted work, have you allowed time for it? Has anyone actually shown them how to do that checking? And do they have someone to turn to when they find a problem they can't resolve themselves?
If nobody has shown someone how to spot a problem, I would want to know why before blaming them for missing it. You could replace that employee tomorrow and still have the same problem, because their replacement would have to work it out too.
But if you've done your part and the person knew what was required, had the time and support to do it and simply didn't bother, then, as a last resort, the blame belongs with them.
Putting Their Names to the Work
It's important for team members to understand that when they pass work on to somebody else, they're putting their names to it.
Colleagues trust each other to have done their work properly, and using AI doesn't change that expectation. If it's good work, they deserve the credit whether they used AI or not. If it's poor, they have to answer for it and help put it right.
Examining the organisation's contribution helps establish why the mistake happened and where blame belongs. It doesn't relieve team members of responsibility for the work they've delivered, even if their employer asked them to use AI.
They also need to accept that AI can get things wrong. They might get a convincing summary of a document that has left out an important condition or added a claim the document never made.
If they're asking the AI to do things for them, a misunderstanding can lead to the wrong action too. Finding out that the AI made the mistake still leaves you with the question of what the person using it did to check the work.
In this case, the lawyer who used ChatGPT knew the firm's rules and admitted he hadn't followed them.
What interests me is that the judge also held the other two lawyers responsible. The two who hadn't used ChatGPT still had a responsibility to check what went to court under their names.
That matters for anyone reviewing or approving a colleague's AI-assisted work. Their approval tells the next person that they've done their part, so they need to be clear about what to check before giving it.
A Note About Agents
Quite often, an employee will be using an assistant, agent or agent skill that somebody else has built for them, rather than an assistant in its default state.
In that situation, you also need to examine what that person built, as the instructions they gave may have caused the problem.
Improve How Your Organisation Works With AI
My advice for business leaders is threefold.
First, follow the judge's line of thinking. She took the trouble to examine each lawyer's conduct as well as the firm's. Give your employees that same care when AI-assisted work goes wrong.
Second, even if an employee was at fault, ask what the incident tells you about AI use elsewhere in your organisation. For example, do people assume that AI-assisted work has already been checked because of who passed it on? If that assumption contributed to one mistake, other people could be making it too.
Third, use what you've learned to improve your approach to AI empowerment. Use practical training to show your people how these mistakes happen and how to catch them in the work they actually do. Give them a chance to practise and get help with the things they don't understand.
As their leader, you have to make time for that, just as you have to allow time for checking the work. Make those checking and approval responsibilities clear in your AI governance, including who people can turn to when they're stuck.
Taking that preparation seriously should mean fewer things go wrong and fewer occasions when you have to ask who's to blame.
Comments