The Challenge
Product liability defense counsel facing AI-related claims need to answer a deceptively simple question: is AI output a “product” under strict liability doctrine? The answer draws on decades of case law spanning aeronautical charts, software licensing, information products, and emerging AI-specific jurisprudence—a body of authority that no single researcher can hold in working memory.
AI-assisted research can synthesize these threads rapidly. But the legal profession cannot afford the kind of errors that language models characteristically introduce: a dropped qualifier that transforms a hedged judicial observation into settled law, a misattributed quotation that assigns one scholar’s position to another, or a subtle word substitution that shifts the meaning of a doctrinal provision.
Our challenge was to produce a comprehensive non-product defense analysis—spanning the Restatement foundation, the information products doctrine, software-as-product jurisprudence, AI output classification, the EU counter-narrative, and practical defense strategy—where every quotation and source attribution could be independently verified against the original material.
Our Approach: Evidence Classification
AI language models generate text from statistical patterns learned during training. This means even well-constructed outputs can contain subtle inaccuracies: merged sources, shifted qualifiers, or plausible-sounding text that doesn’t quite match the original. For professional research, this is unacceptable.
Core principle: We classify every piece of evidence by its provenance—how it was obtained and how it was verified. Not all evidence is equal. Content extracted directly from a verified source archive is treated differently from content surfaced through web search, which is treated differently from an AI model’s training knowledge. This classification is enforced throughout the research process, so the final deliverable contains only material with the highest level of source verification.
The result is a transparency guarantee: for every claim in the final document, a reader can trace backward through the evidence chain to the original source. This doesn’t replace professional judgment—it supports it by making the evidentiary foundation visible and independently checkable.
The Evidence Pipeline
The non-product defense synthesis required processing a substantially larger and more heterogeneous corpus than our previous projects. The source material spanned law review articles, court opinions, Restatement provisions, regulatory documents, and EU directives—a mix of formats, access restrictions, and quality levels.
We built a structured, searchable evidence archive from these 41 sources. Each record in the archive preserves the original text, its source location, and integrity verification data—so any claim in the final synthesis can be traced back to a specific passage in a specific document.
Processing a mixed corpus at this scale surfaced practical challenges. Not all sources are equally accessible: some academic repositories block automated retrieval, requiring alternative approaches. And not all downloaded content is genuine—we developed quality gates to identify and remove non-content artifacts (bot-protection pages, cookie consent screens) before they could contaminate the evidence archive.
Citation Verification
Raw evidence counts matter less than how that evidence connects to the prose a reader sees. We implemented a multi-level citation system designed for readers who will check sources, verify quotations, and hold the document to professional standards.
The system operates at three levels of specificity. At the broadest level, every substantive claim links to a numbered source reference with full bibliographic information. At the middle level, each source reference includes a direct link to the original document. At the most granular level, direct quotations are linked to the exact passage in the source—so a reader can click through and see the quoted text in its original context.
Each level works independently. If a source link breaks due to URL changes, the bibliographic reference and quote-level verification remain intact. No single point of failure can compromise the entire citation apparatus.
What We Discovered
Building the citation system at this scale revealed two systematic error patterns that had not surfaced in smaller projects. Both were architectural rather than incidental—the kind of issues that would recur in any future project using similar techniques unless explicitly addressed.
Automated quote-linking produced unexpected structural damage to the document. The linking process matched text in locations where it shouldn’t have—inside navigation elements and metadata—creating broken page structure and content artifacts visible to readers. The fix required a fundamentally different approach to how the linking process identifies and modifies quotation text.
When a quotation couldn’t be matched directly to its source, a fallback mechanism assigned the nearest available source. This produced a subtle but dangerous error: a quotation from one scholar was attributed to a different source simply because of document proximity rather than actual provenance. The fix introduced a secondary verification requirement: no source assignment without positive confirmation that the quote actually appears in the proposed source.
Both error patterns were formally documented with detection criteria and prevention protocols. They now apply to all future projects—turning project-specific incidents into system-wide safeguards.
Why This Matters
The non-product defense is not an academic exercise. It is the argument that may determine whether AI companies face strict liability for their systems’ outputs—and the legal authority supporting that argument spans five decades of evolving doctrine. The underlying research must be precise, traceable, and transparent enough for professionals to verify independently.
Source Traceability
Every quotation in the final document links back to its original source. Professionals reviewing the work can verify any claim without relying on the research process itself—the evidence trail is open and independently checkable.
Doctrinal Precision
The boundary between “product” and “non-product” turns on precise language from Restatement provisions, judicial holdings, and scholarly analysis. A dropped word or shifted qualifier can misstate the doctrine. Source-verified methodology prevents this at the point of synthesis.
Scalable Rigor
41 sources and 8,039 evidence records represent a corpus too large for manual verification but too important for unverified synthesis. Our methodology bridges this gap—scaling rigorous verification to the volume that complex legal questions require.
The question is not whether AI should assist professional research—it should. The question is whether the methodology behind that assistance is transparent and rigorous enough to trust. Our approach produces research where every claim is traceable, every quotation is verifiable, and every source attribution is independently checkable—supporting the professional judgment of those who build on it.
Changelog
Record of the Non-Product Defense research module’s development.