The Missing Proof.
Operational evidence and reasonable care when AI acts on an organisation’s behalf
Joe Braidwood · September 2026 · 28 pages
AI can act through several organisations and leave none able to establish the full sequence. This paper asks who should bear the cost of making that conduct knowable.
In brief
The paper joins the law of reasonable care to the economics of evidence production: the person affected may be unable to reconstruct a record that others could have generated, and those others may lack an incentive to pay for it. Industry custom deserves weight, but what it tells a court depends on how the practice developed and whose interests it reflects. The paper separates the party best placed to prevent harm from the party best placed to produce evidence about it. Within an applicable duty, it sets out when omitting a specified recording capability may support breach. Causation remains separately necessary. It argues for no general duty to instrument AI and no reversal of the burden of proof. The aim is to make responsible conduct demonstrable and consequential errors contestable.
The paper’s worked examples are drawn from healthcare; the argument is general.
Reasonable care, and the evidence it leaves
Industry custom deserves weight, but how much depends on how the practice developed and whose interests it reflects. The paper separates two questions that tend to be merged: which party was best placed to prevent an accident, and which party was best placed to produce evidence about it. They are not always the same party, and the second question is rarely asked out loud.
Within an applicable duty, the omission of a specified recording capability may support breach where the capability performs a concrete safety function, lies within the defendant’s practical control, and imposes a proportionate burden. Causation remains separately necessary. The paper argues for none of the shortcuts that argument is often mistaken for: no general legal duty to instrument AI, no reversal of the burden of proof, and no new tort.
What AI insurance can read from a record
An operative control, a retained record and verification that does not depend on the supplier’s interface each need their own justification. A signature shows that a record has not changed since a commitment was made. It does not show that the observation behind it was accurate, complete, adequate, compliant or admissible, and a paper that claimed otherwise would be selling something.
Where evidence chiefly supports accountability after harm, procurement and institutional arrangements can supply a basis for requiring it without stretching preservation law. The paper develops an information-forcing contractual proposal, sets out the conditions under which safety commitments could become commercially consequential, and proposes comparison against competent conventional records. Insurers, purchasers and courts set their own requirements; nothing here anticipates them.
What’s inside the paper
- The allocation problem: an AI action crosses several organisations and none of them holds the whole account
- What industry custom can tell a court, and what its informational value depends on
- The party best placed to prevent harm, and the party best placed to preserve evidence
- Four considerations that organise a breach inquiry into a specified recording omission, proposed rather than established
- Why an operative control, a retained record and verification each need their own justification
- An information-forcing contractual proposal, and how procurement can require what the law does not
- The limits the paper keeps: no general duty to instrument AI, no reversal of the burden of proof and no new tort
Who it’s for
Written for general counsel, for underwriters and claims teams, for risk and assurance leaders and for the standards and policy people who will decide what an AI record has to show.
Abstract
Consequential AI can act through several organizations while leaving none able to establish the full sequence. This paper asks who should bear the cost of making that conduct knowable. Its argument joins the law of reasonable care to the economics of evidence production: an affected person may be unable to reconstruct a record that others could have generated, while those others may lack an incentive to incur its cost. Industry custom deserves weight, but its informational value depends on how the practice developed and whose interests it reflects. Drawing on that problem, the paper distinguishes the actor best placed to prevent an accident from the actor best placed to produce evidence about it. Healthcare supplies the principal application. Within an applicable duty, omission of a specified recording capability may support breach when it performs a concrete safety function, lies within the defendant’s practical control, and imposes a proportionate burden; causation remains separately necessary. An operative control, a retained record, and independent verification each require justification. Where evidence chiefly supports accountability after harm, procurement and institutional arrangements can supply a basis for requiring it without stretching preservation law. The paper develops an information-forcing contractual proposal, explains the conditions under which evidence could make safety commitments commercially consequential, and proposes comparison against competent conventional records. The aim is to make responsible conduct demonstrable and consequential errors contestable, while leaving the architecture and its claimed benefits open to independent challenge.
Disclosure. The author is the founder of Glacis Technologies, which develops infrastructure for verifiable AI runtime records and originated OVERT, a specification discussed in this paper.
The paper, in full
28 pages: the allocation problem, what custom can and cannot tell a court, the contractual proposal, and the conditions under which evidence could make safety commitments commercially consequential.
Author
Joe Braidwood
Founder & CEO, Glacis Technologies
Writes on evidence, reasonable care and the arrangements that let insurers, purchasers and courts assess AI risk. Comments on this draft are welcome at [email protected].