Who Is Responsible When It Is Wrong

There is a version of this that has already played out publicly in several organisations: a document goes out with an invented citation, a fabricated figure, or a legal claim that is not true. In every case, the answer to 'who is responsible' has been the same — the person who sent it, and the organisation they sent it for.

No regulator, client or court has yet accepted the tool as the responsible party, and planning on the assumption that one will is not a plan.

Review proportionate to consequence

Checking everything to the same depth is how review collapses — it is expensive, it feels pointless on low-stakes work, and people stop doing it everywhere including where it mattered. Scale it deliberately.

• Reversible and internal — a draft, a summary for yourself, a brainstorm. Read it. That is the review
• External but low-stakes — a social post, an internal announcement. Read carefully, check any name or figure
• External and consequential — a client deliverable, published content, anything with a price. Verify every factual claim against a source, and have a second person read it
• Legal, financial, medical or safety-related — a qualified human owns the output. AI may assist in drafting; it does not carry the judgement, and this is not a place to be clever about efficiency

Keep a record for consequential work

For anything a client pays for or a regulator could ask about, note which tool was used, for what part, and who verified it. This takes a line in the document properties or the project file.

It is not bureaucracy for its own sake. When a question arrives six months later, the difference between a two-minute answer and a two-day reconstruction is whether that line exists.

When something wrong gets out

1. Correct the substance first, before working out how it happened. The wrong information is the live problem
2. Tell whoever relied on it, plainly, without hedging. Late correction is much worse than fast correction
3. Establish what actually failed: was it not checked, checked badly, or checked by someone with no way to know?
4. Fix the process, not the person. 'Be more careful' changes nothing; 'figures get traced to a source before sending' changes something
5. Write down the case. Real incidents are the most persuasive training material any organisation has

Disclosure — the practical line

There is no universal rule yet, and pretending there is helps nobody. A workable position: disclose when someone's decision would change if they knew.

A client paying for expert analysis should know how it was produced. A reader of a how-to article generally does not need a note that the draft was tidied by a model. Never claim first-hand experience, original research or professional judgement that did not happen — that is the line that does not move, and it is a claim about you, not about the tool.

The thing to keep in view

These tools raise the floor of how fast work can be produced. They do not raise the floor of how carefully it is checked, and the gap between those two is where the incidents live.

The organisations getting real value are not the ones using AI most. They are the ones who decided early where a human stays in the loop, and then actually kept them there.