Blamefall

Blamefall describes a risk that a senior professional is made to absorb. The organisation mandates an artificial intelligence system. The professional did not choose it and cannot opt out of it. The system is unfinished, as all of these systems are. Whatever risk remains in that unfinished system gets pushed down onto the named person who signs the work, the fall guy in the precise sense of the phrase, the one who takes it for the team when something goes wrong. The familiar instruction that these systems are used at the user's own risk omits the fact that the use was never a choice.

The claim is structural, concerning who carries the downside once the work is signed, and the compulsion behind it is documented. In policy adopted in November 2024, the American Medical Association (AMA) recognised that doctors employed by hospitals may feel they have no real choice about using AI once their employer requires it or builds it into the medical record.

Three ways the same tool falls

The same governed tool falls on the user in three ways.

The first fall is the output that was actively false. A fabrication clears the controls, and the person who signs the work owns it. In Mata v. Avianca, decided in the Southern District of New York in June 2023, two named attorneys drew a $5,000 sanction after ChatGPT invented six judicial opinions, the court finding bad faith and conscious avoidance across a forty-three-page order. One attorney had asked the tool whether the cases were real and accepted its assurance that they were. He checked, and the tool confirmed the fabrication he was checking for, so the diligence delivered the error rather than catching it. The English courts came to the same place in Ayinde and Al-Haroun, [2025] EWHC 1383, where the President of the King's Bench Division referred a barrister and solicitors to their regulators and ruled that a lawyer cannot rely on a lay client for citation accuracy. A public database of such cases passed 1,300 worldwide by the middle of 2026, the penalties rising well beyond the $5,000 of the first.

In the second, the professional has the expertise to see that the output falls short and redoes the work by hand. The cost is time: the professional becomes the bottleneck while colleagues who waved the same output through move faster. Refusing the fall is an act of judgement, and inside the institution that judgement reads as delay.

In the third, the output is again inadequate, but it goes out unnoticed, and the person who signed it carries the liability for what they did not perceive.

Less wrong and less useful go together

Governance produces the second and third falls by the same act that prevents the first. To stop the tool being wrong, an organisation adds record-keeping, sign-offs, and rules about what it can and cannot say. Each of those pulls the tool toward safe, generic output, which is the weak output. What makes the tool powerful is the same thing that makes it risky, so limiting one limits the other. So, the choice between wrong and weak does not belong to the senior user. Governance makes it and then leaves the user to answer for being slow or for the inadequacy they could not see. The risk did not disappear when the controls went on. It just moved onto a person and the mandate removes the exit.

Liability that never reaches the vendor

Across all of these cases one party stays untouched. The vendor that built the tool does not indemnify the professional. In Walters v. OpenAI in Georgia in 2025, a defamation claim brought directly against the vendor, OpenAI won on summary judgment, the decision turning on the disclaimers the product issues about its own output and on the claimant having shown no actual harm. No vendor has compensated a sanctioned signatory. Moffatt v. Air Canada showed a deploying organisation unable to disown its own chatbot before a British Columbia tribunal in 2024, and even there the supplier behind the chatbot was never made a party. The deploying organisation may answer to its customers, as Air Canada did before the tribunal. The professional who signed the work faces the sanction or referral.

Others have named the half of this that is documented. Madeleine Clare Elish called it the moral crumple zone, the human who absorbs blame for the failure of an automated system. Scholars writing in Nature Machine Intelligence in 2023 described a credit and blame asymmetry, where the human earns less credit for shared work. The liability sink names clinicians absorbing legal exposure for errors they had limited control over. Each of these covers the situation where the tool was actively wrong. None of them separates the false output from the accurate but weak output, or names the split inside it. I would argue that part is still unnamed.

In policy, the AMA has said that where a mandate prevents the professional from mitigating the risk, the liability should fall on whoever issued the mandate. The courts have not followed. Whether accountability stays fixed on the person who signs, or some shared arrangement forms between the organisation that mandates the tool, the vendor that builds it, and that person, is not yet settled. For now, the name on the work carries the fall.