The Signature Fiction: Liability and the AI-drafted document

When a company builds an autonomous car, the team does not need a philosophical debate to take accountability seriously. The crash is a statistical certainty; people turn up each day knowing the courtroom is inevitable.

Swap the road for the board pack, and the accountability problem hasn’t changed. The surface is different, but the mechanics of failure are identical: unclear authorship, hard reconstruction, invisible omissions, drifting behaviour, and a liability that still demands a signature.

Accountability exists in both places. One team knows that the tide goes out. The other behaves as if it never will.

The governance bargain, and the new fiction

Corporate governance rests on a specific bargain: that a signature binds a named person to a set of claims and judgements and makes them answerable for the consequences.

The assumption underneath is simple. If challenged, the organisation can defend how it got there.

Model-first drafting changes how that chain is formed, but it does not change what the signature means.

When an LLM builds the first draft, it sets out the logic and the structure. A human usually adjusts the prose, but the AI's core choices often remain untouched because they read well. Documents have always carried reasoning the signer didn't originate, from templates, precedent libraries, and junior drafters, but those earlier versions arrived rough enough to invite reconstruction. A model's draft does not. It arrives fluent, structured, and complete, which is precisely why it survives review intact. That is what Residual Logic means: inherited reasoning that persists because it never looked like it needed questioning.

Ask who authored it, and you will hear: "I shaped the brief." "I worked with the model." This isn't modesty; it is a defensive reflex. The human is pre-emptively distancing themselves from the logic because the foundational reasoning was generated rather than authored.

The production method has shifted while the ritual has stayed the same.

The editorial retreat

“I tidied it up” describes what the work has become: surface responsibility for a document whose shape arrived pre-formed. A model-first draft makes it easy to improve prose without ever revisiting the frame, the sequence, or the omissions, and that is how Residual Logic persists in documents that look polished and complete. The edits can be genuine and extensive, while the reasoning still follows the path that arrived first, including what it kept out of view.

A board pack can feel finished in the meeting where it is approved, because it reads well and the process completes. The shift happens later, when the same document is read back as evidence of judgement rather than as a coherent piece of work.

Style becomes irrelevant. Structure becomes the record. The paper fixes what was treated as material, what was treated as causal, and what never entered scope in the first place. Absences do not announce themselves at drafting time. They only become legible when consequences force the missing question into view.

The tide goes out when someone outside the workflow reads the document and assumes the signature means the reasoning can be defended. That is the bargain the signature has always represented, and it is the bargain AI-assisted drafting is currently weakening.

A routine signature can sit on reasoning that was made presentable without being rebuilt, because reconstruction is not demanded until the document meets scrutiny it was, arguably, never designed to withstand.

The system works because everyone believes someone else could do it if needed.

The crash is assumed in the car world. It is discounted in the document world, even though it is just as predictable in a large enough organisation.

What it looks like when the tide goes out

In 2025, Deloitte Australia used GPT-4o to help produce a 237-page independent review for the Department of Employment and Workplace Relations. The final document was found to contain fabricated academic citations and non-existent court references that Deloitte's own review process did not catch, according to subsequent reporting. The firm was forced to refund part of the AU$440,000 contract and disclose its use of AI, which had not been mentioned in the original version.

Deloitte's post-hoc position was that the core findings remained valid, and that response is worth pausing on because it is the same instinct the piece has been describing. The document read well, the process completed, but the reasoning beneath it had not been rebuilt by anyone whose name was attached to the work. The belief that the surface can be defended even when the underlying construction was never owned is exactly how Residual Logic survives institutional review.

This was a Big Four firm delivering a formal government review with institutional quality assurance in place, and the failure mode was editorial retreat operating at scale, where a review process checked for coherence and presentation while the sourcing and logic went unexamined. A year earlier, in Mata v Avianca, lawyers submitted a court filing built on fabricated case citations generated by a language model, and the fabrications survived the review process intact, which tells you what kind of review it was.

Two professions with different review cultures produced the same structural failure, because a plausible document passes review when the review has become a plausibility check rather than a defence of the underlying reasoning. In both cases, consequences followed. The signature held its punitive weight. But the governance bargain is not only about what happens after the failure. It is about the assumption that the organisation can defend how it got there when challenged, and in both cases that assumption collapsed before anyone outside the workflow tested it. Neither the legal system nor the regulator cares that the workflow has changed. They only care that the signature exists.