Failed AI work needs enforced release checks

The dangerous failure is not an imperfect AI draft. It is unsupported work crossing the release boundary into a filing, customer message, financial decision, or other consequential record. Every material claim needs inspectable evidence, uncertainty needs a stop, and a recorded reviewer decision must control release.

An AI constitution, a short charter stating a role's authority and limits, can define that boundary. It cannot prove the boundary was followed on a particular assignment.

Sullivan & Cromwell's April 2026 filing matters because the firm said policies and review processes already existed. The firm's letter to the court acknowledged inaccurate citations and other errors associated with AI use. The failure was not the absence of written rules. The rules did not become a successful check before the document was filed.

What did the incident establish?

The primary filing establishes a bounded set of facts: lawyers used an AI research tool, errors reached a court filing, the ordinary review failed to catch them, and the firm apologized and described corrective action. Bloomberg Law's account reports the same event and the firm's response.

It does not establish that every use of AI in legal work is unsafe. It does not establish that one document design would have caught every problem. It does establish that professional experience, an internal policy, and general review language can coexist with a failed release.

The American Bar Association's discussion places the incident inside existing duties of competence and verification. Those duties remain with the lawyer. An AI system, vendor, or internal checklist does not inherit the professional obligation.

What should the constitution require?

The charter should be short enough to use during work. Five clauses do most of the work:

1. Source authority. State which records may support a material claim and which sources are discovery aids only. Done when: every material proposition points to an allowed authority.
2. Conflict behavior. Require a stop when authoritative records disagree. Done when: the work product names the conflict and the person who can resolve it.
3. Citation verification. Require the reviewer to open the authority and match the cited passage to the proposition. Done when: another reviewer can reproduce the match without a search.
4. Release authority. Name the person who may approve external or binding use. Done when: the release record contains that person's decision and the completed checks.
5. Incident repair. Require a corrected work product, a cause record, a changed normal method, and a passed retest. Done when: the failure case and a nearby ordinary case both pass.

NIST's AI Risk Management Framework Core supports this operating distinction. Its Measure and Manage functions call for testing in conditions similar to use, production monitoring, appeal and override, incident response, recovery, and change management. NIST provides voluntary guidance, not certification of FidelicAI or a rule for court filings.

A worked release record

Suppose a research role prepares a memo that states a court adopted a particular standard. The role finds a case summary and a linked opinion.

The weak check asks whether the citation looks complete. The useful check asks whether the cited court, date, procedural posture, and exact passage support the proposition. If the opinion says the court declined to reach the issue, the memo stops even though the case and citation are real.

This record does not make an AI role a lawyer. It replaces some human source collection, citation assembly, and evidence-record maintenance. The lawyer still interprets the authority, resolves material conflict, and owns every professional conclusion and release decision. The guide to hiring an AI agent applies the same proof test to nonlegal work: define the work product, acceptance check, approval boundary, and remedy before hiring.

What can still go wrong?

The allowed source can be stale. A reviewer can misunderstand the issue. The primary record can be ambiguous. Time pressure can turn a required check into a box ticked from memory. A person with release authority can approve the wrong conclusion.

The constitution should therefore avoid absolute promises. It should require visible uncertainty and a stop, not claim that uncertainty will always be detected. The production reliability guide separates a passed example from continuing performance. The security record states the current access boundary; it does not establish substantive accuracy.

For a fidelic agent, the role's published limits, current connections, job-specific inputs, and owner approvals form the working boundary. The buyer still owns authoritative records and consequential decisions. Licensed judgment remains with the licensed professional.

Use this release test

Take one consequential work product and select five material claims. For each claim, record:

• the exact proposition;
• the allowed primary authority;
• the supporting passage and date;
• any conflict or qualification;
• the reviewer and completed check;
• the release decision and approver.

Done when: a fresh reviewer can reproduce all five claim matches, find every exception, and confirm that no binding or professional action occurred without the required approval.

Use the AI agent versus chatbot guide to distinguish a finished, checked record from a plausible answer. Review the current AI agent catalog for role-specific work products and limits. Keep the work-product promise separate from professional judgment: it provides the published commercial remedy for qualifying delivery misses, not a guarantee that every claim is correct.