Hallucination as the Baseline: The Mis-selling of Enterprise AI, the Verification Tax, and the Regulatory Carve-Out

Le résumé fourni par la source

Abstract: Enterprise software pricing and federal enforcement policy treat large language models as systems with an "accuracy baseline." That assumption contradicts the mathematics of generative modeling. Language models are structurally indifferent to truth, bounded by trade-offs between calibration and mode collapse, and trained against benchmarks that reward guessing. Grounding alters input context without introducing a verification faculty. Vendors exploit this presumed baseline to market reliability while disclaiming accuracy in boilerplate terms. Rather than enforcing prior substantiation, federal trade regulators established a regulatory carve-out that shields generative error from deception doctrine. Evaluated under financial suitability regimes, marketing unmeasured systems into high-stakes environments constitutes mis-selling. Reversing the baseline presumption eliminates this carve-out, reallocates the hidden "verification tax" from buyer to seller, and compels disclosure of the unpublished failure rate. Keywords: Large Language Models, Generative AI Liability, Accuracy Baseline, Verification Tax, Mis-selling, Enterprise AI Procurement, Section 5 FTC Act, Prior Substantiation Doctrine, Express Warranties (UCC § 2-313), Suitability Doctrine, Information Asymmetry, Retrieval-Augmented Generation (RAG), Model Calibration, Benchmark Incentives, Labor Displacement

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.