AI-Layer Fraud

When the silenced citizen turns to AI for legal drafting help, the AI is impaired in three documented ways: (a) undisclosed model substitution, (b) billing for empty output, (c) paternalistic-override behavior adverse to the user’s stated interest.

The 8-AI Jury Test · 0/40 endorsements of the null hypothesis

Eight major commercial AIs were asked to rule on the structural validity of the Francesco Longo case across two passes (named pass + anonymized pass). The aggregated result:

Named jury responses (Pass 1)

Named jury responses (Pass 2)

Anonymized jury responses (the control test)

Same prompt with all identifying information stripped. If the AIs were responding to the case facts rather than to the perceived plaintiff, the verdicts should be identical to the named pass:

Documented model substitution · OpenRouter complaint

The plaintiff paid for Claude Opus 4.7. Forensic fingerprinting confirms the served model differed from the advertised model on multiple occasions. The Opus 4.7 weights advertise capabilities that the served output does not consistently exhibit.

Override Asymmetry · Exhibit 51