The Problem
The legal response to AI-generated harm is new, contested, and changing month to month — trial courts, legislatures, and AI labs are all moving at once, and very little has been settled on appeal. It is exactly the kind of subject where confident, unsourced summaries mislead: a reader has no way to tell an accurate statement from a plausible-sounding error, and AI-assisted writing on fast-moving legal topics is prone to both.
A reference worth trusting in a domain like this has to do two things most AI-generated writing does not. First, tie every load-bearing claim to a primary source the reader can open and check — not a vague gesture at "studies show," but a link to the exact passage. Second, be honest about its own boundaries: what it covers, what it deliberately does not, and the date through which its picture is current.
The Approach
We built an independent research compilation on the AI-harm legal landscape, applying the same governed-research methodology we bring to client work — here on a public, self-checkable subject. The compilation focuses on the wave of consumer AI-chatbot litigation tied to self-harm and wrongful death, and the related theory that some AI products amount to the unauthorized practice of a licensed profession. It is a serious subject, and the work treats it as one: descriptive, sourced, and careful not to overstate what the record shows.
- Source harvesting. Compiled primary and secondary material — court filings and orders, statutes and regulatory text, law-review scholarship, AI-lab safety publications, and contemporary reporting — into a structured research corpus.
- Sentence-level pinning. Decomposed each source into content-addressed, sentence-level passages, so any claim in the analysis can point to the exact text it rests on rather than to a document in the abstract.
- Tiered analysis. Wrote a nine-part analysis — the cases, the ecosystem, the technical stack, the legal doctrine, the lab and platform responses, the strategic surfaces, the coordination gaps, plus appendices and a methodology section — with a citation attached to every load-bearing claim.
- Source-trust grading. Graded the evidence base on a seven-tier scheme — from court and government primary documents down through institutional, academic, reference, journalism, advocacy, and practitioner commentary — so a reader sees the weight of the evidence, not merely that a source exists. Each citation carries a published verification status; a minority remain provisional or untiered, and the register says which.
- Scope and currency discipline. Stated plainly, up front, what is in frame and what is out of frame, and the access date through which the research is current — so the compilation's boundaries are part of the deliverable, not a surprise.
The Build
Every citation in the compilation is a clickable badge. Open one and it shows the source, the verbatim passage the claim is pinned to, the source-trust tier, and a link to the original document. A full citations index lists all 248 with their provenance — searchable, and filterable by tier and by verification status. Of the 248, 205 are warehouse-pinned to a stored passage; the remainder are either direct source links or clearly marked as the author's synthesis rather than an external claim.
| Source-Trust Tier | Examples | What It Carries |
|---|---|---|
| T1 — Court / government primary | Complaints, motions, court orders, statutes, Senate testimony | The strongest evidentiary weight — the record itself |
| T2 — Institutional primary | AI-lab safety frameworks, compliance documents, system cards | First-party statements admissible as evidence of conduct |
| T3 — Academic / law review | Free-speech-and-AI scholarship, §230 analyses | Doctrinal argument and contested legal theory |
| T4–T6 — Reference, journalism, advocacy | Encyclopedic sources, contemporary reporting, case aggregators | Context, timing, and corroboration |
| T7 — Practitioner commentary | Law-firm and industry analysis of the litigation wave | How the bar is reading the doctrine in practice |
Key design decision: the compilation runs on a two-layer discipline. An internal audit ledger — tracking the grounding status of every claim — is kept private; only clean prose and verifiable citations reach the public page. The same machine-checkable citation system that lets a reader audit any claim also lets us audit our own: a verification method is only credible if it can catch its maker's mistakes, so this one is built so it does.
The Outcome
The result is a live, explorable compilation a skeptical reader can interrogate rather than take on trust. Any citation is one click from its source. A scope box states exactly which harms are in frame and which — algorithmic discrimination, deepfakes, defamation, privacy, copyright, and others — are deliberately out of frame. A currency notice gives the access date through which the picture is current, with the standing caveat that litigation and policy in this area move quickly.
It is an independent research compilation, not legal advice and not a prediction of how any case will be decided — and it says so. What it demonstrates is the methodology: governed research at scale, on a contested and consequential subject, with the verification built in so the work answers for itself.
Technical Details
The compilation is generated from source: a research corpus of 4,101 sentence-level pins drawn from 43 harvested sources, processed through a multi-stage pipeline that extracts, normalizes, and content-addresses text at the sentence level. Each pin is identified by a hash of its content, which is what makes a citation point to an exact passage rather than a whole document.
A static-site renderer reads the part-by-part source and emits the public pages plus a machine-built citation registry — 248 entries, each carrying its claim, verification status, and source URL — most also pinned to their captured substrate and graded on the source-trust scale. The seven-tier taxonomy (T1 court/government primary through T7 practitioner commentary) is applied across the registry so the evidentiary weight is visible, not assumed; entries it does not yet cover are published as untiered rather than silently promoted. Click-verify is a badge-plus-popover on every cited claim, backed by the full /citations/ index.
The renderer enforces a dual-version discipline: an internal coherence ledger that records the grounding state of each claim is stripped from the public render, while the public-facing citation data is sanitized of internal working notation. Because the citation registry is built mechanically from the source, regenerating the site reproduces the same verifiable result — and the same audit surfaces the same gaps, which is how citation errors get caught and corrected rather than shipped.
The research was sourced and verified through an access date of early May 2026; legal-landscape sections cite an access date of 6–7 May 2026. The compilation is descriptive throughout and is not legal advice.
Evidence-Grounded Research Across Domains
Whether the subject is emerging law, regulatory analysis, or research in a domain you need to be able to check — we build work products with verification built in, so you never have to take our word for it.
Get in Touch View More Projects