Legal teams are increasingly deploying AI assistants for drafting and research, but treating generated text as verified authority is a critical infrastructure failure. Isabelle Dubuis, writing for DEV.to on October 22, 2026, argues that an AI draft is not evidence. The core recommendation is to replace vague "human in the loop" disclaimers with a rigorous, auditable "claim-to-authority register" that independently verifies every material proposition before release.

The Four Failure Modes of AI Citations

The article identifies four distinct ways an AI-generated citation can fail: it may not exist at all, it may identify the wrong case or statute version, it may be correctly identified but fail to support the specific proposition, or it may be valid in one jurisdiction but inapplicable to the current matter. Standard link validation only addresses a fraction of these issues. A URL returning HTTP 200 does not prove the content supports the legal claim, especially if the publisher has updated the page or the model hallucinated the context.

Building a Claim-to-Authority Register

Dubuis proposes a structured register that treats review as a sequence of independent checks rather than a single glance. Each material claim requires a stable ID, the specific jurisdiction and date, the exact pinpoint location in the primary source, and a record of independent access outside the AI interface. The register must distinguish between primary authority and commentary, ensuring that a blog post or vendor page never silently becomes the sole basis for a client-specific legal conclusion.

Executable Preflight Checks

To operationalize this, the article provides a Python-based preflight checker that enforces mechanical release conditions. The script validates that every material claim has a corresponding reviewer, a verified source ID, and a confirmation that the primary source was opened independently. While the code cannot determine legal validity, it prevents unreviewed claims from slipping through the document pipeline. It ensures that the final artifact matches the reviewed version, blocking releases if the draft version number changes or if the support decision is not marked as "supported."

Key Takeaways

  • Treat AI output as provisional draft text, not verified authority, until it passes a primary-source check.
  • Implement a claim-to-authority register that tracks jurisdiction, date, pinpoint location, and reviewer sign-off for every material proposition.
  • Use automated preflight scripts to enforce version consistency and ensure independent verification before release.
  • Recognize that link validation is insufficient; a valid URL does not prove the content supports the specific legal claim.

The Bottom Line

Automated tools cannot replace legal judgment, but they can enforce discipline. If your workflow allows a citation to pass from AI draft to final filing without an independent, human-verified link to primary authority, you are not using AI responsibly; you are gambling with professional liability.