Blamefall

Blamefall describes the downward transfer of risk that occurs when an organisation mandates an artificial intelligence system, forcing a senior professional to absorb the consequences of a tool they neither chose nor can opt out of.

Because all such systems are inherently ‘unfinished’, the residual risk is pushed directly onto the named person who must sign off on the work, making them the fall guy in the most precise sense of the phrase. Consequently, the familiar disclaimer that these systems are used at the user's own risk is a convenient fiction that ignores the complete absence of choice.

OpenAI's terms of use state that "any use of Outputs from our Service is at your sole risk", and tell the user not to treat output as the only source of truth or as a replacement for professional advice.

The professional is not choosing to take on the risk; they are being directed to absorb it by an institution that will hold them responsible when the unfinished technology inevitably fails.

In policy adopted in November 2024, the American Medical Association recognised that doctors employed by hospitals may have no real choice about using AI once an employer requires it or builds it into the medical record. This is the premise of Blamefall. The liability concentrates on the one person whose name is on the output, while the organisation that mandated the tool and the vendor that built it remain clear of it.

The Three Conditions: false output creates the visible fall; weak output creates delay for the expert and liability for the non-expert

Blamefall has three possible conditions:

A governed AI tool produces failure in more than one way, although only the first is easily visible.

First, the tool's output is false. The untrue output passes the organisation's controls, appears in work issued under someone's name, and that person carries the error. This is the visible case because courts produce public records. In Mata v. Avianca, Ayinde, and Al-Haroun, AI-generated false material appeared in formal work, and the consequence attached to the professionals responsible for that work. Damien Charlotin’s database now records more than 1,600 examples, with penalties in some cases reaching hundreds of thousands.

But these visible failures are the rare ones. They are visible because they have reached a court or another formal record. Most instances of Blamefall will not reach that point, so they leave no public trace.

Second, the output is weak, and the person has the expertise to see it. They recognise that the work is inadequate, so they redo it themselves before signing, escaping the liability, and paying for that escape in time. Inside the organisation, their judgement can then be read as delay or obstruction. This is the cost of refusing the fall.

Third, the output is weak, and the person does not have the expertise or time to recognise the problem. They sign the work because they do not see the fault. The liability then attaches to a defect they never perceived.

The Governance Paradox: The controls designed to prevent false output help produce weak output

To stop the tool from being wrong, organisations often deploy strict governance rules to restrict what the LLM can produce. A tool that is restricted in what it may say is less likely to make an obvious false claim. But governance is a double-edged sword, and an LLM with enterprise-level governance controls in place is also less likely to produce specific, useful judgement. The output effectively becomes safer because there is less in it that can be directly shown to be wrong.

That safety has a cost. The work may become more cautious, more general, and less useful for the decision it is meant to support. It can pass through a process without being good enough for the task.

The person who signs the work did not set those limits. They receive output produced under them. That person never weighed the obvious failure against the weaker one. The organisation weighed it when it governed the tool, then handed the result over to be signed. The choosing and the signing are done by different people.

That is the governance paradox. The controls reduce the visible failure and leave the weaker failure in place. The professional is then exposed either to delay or to the fall.

Where Existing Names Stop: existing accounts explain blame after wrong output; Blamefall isolates the split created by weak output

Others have named neighbouring structures. Madeleine Clare Elish's "moral crumple zone" describes the human who absorbs blame for an automated system's failure. The "liability sink" describes professionals who carry legal exposure for errors they had little control over. Researchers in Nature Machine Intelligence described a credit and blame asymmetry, where the human earns less credit for shared work. Every one of these stops at the case where the tool was actively wrong. None isolates the split inside the weak output, where the same inadequacy falls on the expert as delay and on the non-expert as risk. That is the ground Blamefall claims.

Where the Liability Falls: vendor terms put output risk on the user, while mandate concentrates the consequence on the signatory

Vendor terms set one boundary around responsibility. OpenAI places use of outputs at the user’s sole risk. Google and Anthropic use similar protection through “as is” and no-reliance language. In Walters v. OpenAI, those disclaimers helped OpenAI defeat the claim. In Moffatt v. Air Canada, the deploying company could not disown its chatbot’s answer, and the supplier was never made a party.

The vendor has legal language pointing responsibility away from itself. The organisation may have issued the mandate. The person whose name appears on the work remains closest to the formal consequence.

The American Medical Association has said that mandate should matter. If a professional is required to use a system, and that requirement prevents them from mitigating the risk, the AMA says liability should sit with whoever issued the mandate.

The false-output cases cited earlier do not do that. Mata v. Avianca, Ayinde, and Al-Haroun show consequences attaching to professionals responsible for formal work containing AI-generated false material. They do not show liability moving back to whoever required the tool to be used.

That leaves the professional exposed. The mandate may explain why the tool had to be used. The consequence can still attach to the person responsible for the work, because the cited cases do not treat mandate as the reason to move liability away from the signatory.