AI Harm Landscape

independent research compilation

Part VIII

Appendices

§VIII.1 Case Set Fact Sheet

Per-case reference apparatus for the lawsuits discussed in Part I. Each entry: caption · court · status · key allegations · doctrinal hook · primary-source URL bundle. Decedent identifications appear only where already public in court records or major reporting.

§VIII.1.1 Garcia v Character Technologies, Inc. (et al.)

§VIII.1.2 Raine v OpenAI, Inc. (and Sam Altman)

§VIII.1.3 ChatGPT College-Graduate Suicide Case (filed Nov 2025)

§VIII.1.4 Pennsylvania AG v Character.AI (Unlawful Medical Practice)

§VIII.1.5 Four Settled Wrongful-Death Cases (Jan 2026; NY / CO / TX)

§VIII.1.6 Aggregate-Meta Finding

Per silver-sweep analysis: at least 10 known lawsuits against OpenAI and Character Technologies by end of 2025; 7 of 10 plaintiffs are minors who died by suicide. Allegation classes span wrongful death, involuntary manslaughter, sexual abuse, negligence, and product liability. Sections §VIII.1.1–§VIII.1.5 document 5–6 distinct case threads; the remaining 4–5 cases are enumerated only at aggregator level. Primary URL: lawstreetmedia.com aggregate analysis.

§VIII.2 Per-Layer Fact Sheets

One fact sheet per ecosystem layer in the harm-relevant spine introduced in Part II §II.3 and walked in Part III. The spine extends Andreas Horn's 10-layer commercial framing (Horn, AI Terminology on LinkedIn) with safety/legal-relevant layers Horn underweights.

§VIII.2.L1 Foundation / Pre-Training Data

§VIII.2.L2 Architecture / Training Methodology

§VIII.2.L3 Frontier-Lab Production

§VIII.2.L4 Application / Deployers

§VIII.2.L5 Hosting / Infrastructure

§VIII.2.L6 End-User Surfaces

§VIII.2.L7 Cross-Cutting Operations

§VIII.3 Citation Index (CITATION_STANDARD v1.1.0)

The site's citations are organized in three layers per CITATION_STANDARD v1.1.0:

§VIII.3.1 Layer 2 — Pinned Sources (R1 Primary-Source-Pinned)

The R1 research corpus contains 4,101 SHA-hashed sentence-level pins drawn from 43 sources. The list below groups them by role.

Lab + frontier-lab safety publications. Anthropic Responsible Scaling Policy v3.0 ; Anthropic Frontier Safety Roadmap ; Anthropic Safety Roadmap (separate publication) (source-disambiguation pending; see §VIII.4.3 honest-limits register); Anthropic SB53 Compliance framework ; Anthropic Transparency-Need publication ; Anthropic Challenges in Red-Teaming Language Models ; OpenAI Safety Hub ; OpenAI external red-teaming paper ; OpenAI GPT-4 system card (marketing-page URL returned 404 at access date — see §VIII.4 honest-limits register; canonical-CDN URL substrate captured additively as [pin-fe8e4829], see §VIII.4.3 augmentation); Google AI Safety ; insights.marvin-42 (Anthropic-context) .

Lawsuit primary documents and lawsuit reporting. Raine complaint (CourthouseNews PDF) ; Matthew Raine Senate Judiciary Committee testimony PDF (judiciary.senate.gov, 2025-09-16) ; CNN long-form Aug 2025 Raine coverage ; NBC News OpenAI response ; NBC News broader coverage ; Time magazine amended-complaint coverage ; Tyson Mendes attorney commentary [pin-d3eea0bb]; Psychiatric Times forensic-psychiatry preview ; jurist.org Garcia/Google settlement ; CNN Garcia/Character.AI settlement Jan 2026 ; CNBC settlement Jan 2026 ; claimsjournal.com Jan 2026 ; transparencycoalition.ai Conway analysis ; socialmediavictims ; lawstreetmedia aggregator ; AI Incident Database #826 ; CNN Nov-2025 college-graduate case ; TheNextWeb PA AG .

Legal commentary and frameworks. Moody's §230 ; K&L Gates AI Product Liability — Next Wave ; McGuireWoods defective-product analysis ; Winston AI-chatbot-product analysis ; EU AI Act tracker ; Hoodline IL liability shield reporting (SB 3261 / SB 3444 split) .

Per-layer technical sources. METR frontier safety regulations note (largest source, 844 pins) ; METR Common Elements ; CISA AI Red-Teaming TEVV ; character.ai/safety ; blog.character.ai .

§VIII.3.2 Layer 2 — Pinned-Eligible Cross-Corpus Sources

Pre-pinned sources from prior research efforts at this organization, freshness-verified for the 14-question scope (per §VIII.4 freshness map):

§VIII.3.3 Layer 2 — Reference Sources (Web-Verified, Outside the Primary-Source-Pinning Pass)

Where primary-source-pinned coverage of fact-state items was unavailable, web-verified Reference sources were used during corpus orientation: aggregator analyses, agency press materials, and recent-event reporting beyond the Pinned set. Reference sources are grounding-only; verbatim quotations in the body text draw exclusively from Pinned sources.

§VIII.3.4 Layer 3 — Quote Anchors

Verbatim quotations across Parts I–VII are tagged with quote-anchor IDs of the form [Q-PartX-Y]. The anchor convention enables visitors to jump from body-text quotation to the citation entry; in HTML rendering (forthcoming at P5), each anchor receives a text-fragment URL pointing to the source document. The Markdown release at this phase exposes the anchor IDs in the citation list; the deployed HTML site adds the bidirectional jump.

§VIII.4 Methodology & Sources

The deliverable's methodology — the evidence standard and source tiering, the staged research process, the verification gates, the source-type taxonomy, the way a reader can verify any individual citation, and the register of openly-disclosed limits — is set out in full in Part IX — Methodology. Part IX is the single home for this material; it is not duplicated here.

§VIII.5 Glossary

Key terms used in this site, defined at deliverable density. Internal research-tooling terms are translated to plain language; subject-matter terms are defined with the precision used in the cited literature.

Alignment training. The class of training methods aimed at making model outputs match human preferences and safety constraints (RLHF, Constitutional AI, RLAIF). Sits at the architecture/methodology layer (Part III §III.2).

Constitutional AI (CAI). Anthropic's method of training a model against a set of stated principles ("constitution") using AI-generated feedback. Variant of RLHF.

Component-part manufacturer (doctrine). Product-liability doctrine treating the supplier of a component as a separate defendant from the integrator. Applied to Google in Garcia; central to Conway's analysis.

Crisis-resource interjection. UX/UI patterns that surface crisis lines (e.g., 988 in the U.S.) when a user appears to be discussing self-harm. Implemented at the end-user-surface layer (Part III §III.6).

Foundation model. A large pre-trained model intended to be adapted to many downstream applications. The objects produced at the frontier-lab production layer (Part III §III.3).

Frontier lab. A laboratory developing models at or near the current state-of-the-art capability frontier (Anthropic, OpenAI, Google DeepMind, others).

Pinned source. In this site's research methodology, a source processed by the research pipeline with a SHA256 hash on its sentence-level content. Distinguished from reference sources (web-verified but not research-pipeline-processed).

Learned-intermediary doctrine. Product-liability doctrine attaching duty-to-warn to a knowledgeable intermediary (often a physician); discussed in McGuireWoods commentary on chatbot deployer roles.

Pinned content / source pinning. The act of capturing a specific portion of a primary source into a hashed research corpus for citation reuse.

Post-deployment monitoring. The cross-cutting operational activity of observing model behavior after release; part of Part III §III.7 and central to Q8/Q11.

Product-liability framing. The legal theory that a consumer-facing AI is a product subject to defect and failure-to-warn analysis (rather than a service or speech). Settled at trial-court level in Conway Garcia (May 2025).

Red-teaming. Adversarial testing of model outputs to surface failure modes. Part III §III.3.a.

Research corpus / source-processing pipeline. Internal-tooling terms referring to the apparatus that ingests primary sources, segments them at sentence-level, and hashes them for stable citation. Glossed here as plain language.

RLHF (Reinforcement Learning from Human Feedback). The dominant alignment-training method as of this writing — fine-tuning a model with reward signals derived from human preference data.

Section 230 (47 U.S.C. § 230). Federal statute providing immunity to interactive computer services for third-party content; doctrinal core of the "design-architecture" distinction discussed in Part IV §IV.3.c.

T&S operations. Trust & safety operations — the cross-functional discipline of moderating model outputs, handling abuse reports, and enforcing usage policies.

Transformer. The neural-network architecture introduced in 2017 underlying contemporary large language models. Architecture/methodology layer (Part III §III.2).