AI Harm Landscape

independent research compilation

Part VI

Strategic Surfaces

§VI.1 Frontier-Lab Technical Challenges *(Q8)*

What labs are technically navigating, in plain language

The frontier labs have published — most directly Anthropic at §V.1, OpenAI at §V.2 — substantial detail about the technical challenges they consider open in their own safety practice. The challenges are not hypothetical; the labs name them in their own published material as binding constraints on what current safety frameworks can achieve.

The "zone of ambiguity" in capability-threshold determination

Anthropic's RSP v3.0 announcement is the most-direct lab-side admission of a structural challenge: the science of model evaluation is not yet adequate to support dispositive risk determinations at the capability thresholds the RSP is built around. In the lab's own words, "The science of model evaluation isn't well-developed enough to provide dispositive answers." (Source: Anthropic, "Anthropic's Responsible Scaling Policy: Version 3.0," February 24, 2026, https://www.anthropic.com/news/responsible-scaling-policy-v3 .) The lab uses the phrase "zone of ambiguity" to describe cases where model capabilities have "approached" but not definitively "passed" the RSP thresholds; in such cases the lab has "taken a precautionary approach and implemented the relevant safeguards, but our internal uncertainty translates into a weak external case for taking multilateral action across the AI industry." (Source: same.)

For product-liability standard-of-care analysis, the zone of ambiguity is doctrinally specific: the lab's own published acknowledgment that capability-threshold science is underdeveloped is admissible-as-public-statement evidence in subsequent litigation — both as evidence that the lab knew of the limitation, and as evidence of what reasonable practice in the face of the limitation looks like.

Resource intensity and limitations of red-teaming

OpenAI's external-red-team paper is direct about what red-teaming can and cannot do: "Red teaming on its own is not a panacea for risk assessment." (Source: Ahmad, Agarwal, Lampe, Mishkin, "OpenAI's Approach to External Red Teaming for AI Models and Systems," https://cdn.openai.com/papers/openais-approach-to-external-red-teaming.pdf .) The limitations the paper enumerates include time-decay of findings ("Risks surfaced in one point in time red teaming effort may be under-assessed or no longer reflected in an updated system or model"), resource intensity, potential harms to red-team participants, generation of information hazards, and increasing thresholds for human-judgment sophistication as models become more capable. (Source: same.)

Higher-ASL / SL5-class challenges as collective-action problem

Anthropic's RSP v3.0 self-assessment also identifies a structural challenge in the upper levels of the framework: the safeguards required at higher capability levels may be infeasible for any single lab to implement unilaterally. The lab cites a RAND analysis on model-weight security stating that the SL5 standard, "aimed at stopping top-priority operations by the most cyber-capable institutions, is 'currently not possible'" and "will likely require assistance from the national security community." (Source: Anthropic RSP v3.0 URL above.) This is one of the lab's stated reasons for restructuring the RSP: "requirements at the higher RSP levels that are very hard to meet unilaterally, creates a structural challenge for our current RSP." (Source: same.)

Post-deployment monitoring and continuous-update behavior

The continuous-update nature of deployed AI products complicates safety-evaluation timelines. McGuireWoods observes that "the continuous-update nature of software products creates plausible grounds for courts to impose an ongoing post-sale duty to warn and implement safety features as evidence of harm accumulates." (Source: McGuireWoods, "Can Social Media or AI Be a Defective Product?", Product Liability & Mass Tort Monitor, March 18, 2026, https://www.mcguirewoods.com/client-resources/alerts/2026/3/can-social-media-or-ai-be-a-defective-product/ .) The technical challenge — keeping safety practice current with model behavior that itself updates — is the operational counterpart to the doctrinal frontier discussed in Part IV §IV.3 and at §V.1's Risk Reports framing.

GRAD-INTERN — How lab-disclosed technical limitation maps to attorney working knowledge

A practical orientation for the abstracted product-liability attorney. The technical challenges enumerated above are not cleanly separable from doctrinal handles; each challenge supplies an evidentiary axis on which standard-of-care arguments turn. The "zone of ambiguity" supports an argument that the lab knew certain risks could not be ruled out at deployment time. Red-teaming time-decay supports an argument that any one-time pre-deployment evaluation does not satisfy the duty as the product behavior evolves. Higher-ASL infeasibility supports an argument about the gap between voluntary commitment and operational practice. Continuous-update behavior supports a post-sale duty-to-warn theory of the kind McGuireWoods describes. None of the foregoing predicts how any particular court will rule on any particular factual record; what it does is map the lab's own published acknowledgments of challenge onto the doctrinal axes likely to be litigated.

§VI.2 Frontier-Lab Legal Challenges *(Q9 — PARTIAL)*

What labs are legally navigating, in plain language

The lab-side legal challenges are most-visibly active across two surfaces at access date for this research: defending the named matters described in Part I (motion practice on §230, product-versus-service classification, First Amendment threshold defenses, component-part-manufacturer theories), and engaging in legislative activity at state and federal level over what the regulatory regime should look like.

The Illinois SB 3444 / SB 3261 industry split

The most-substrate-rich current legislative contest is in Illinois, where two competing Senate bills frame the question of frontier-lab liability and transparency in opposing ways. The Hoodline reporting (May 6, 2026) frames the dynamic directly: "Illinois lawmakers are weighing dueling Senate plans: one that would largely shield 'frontier' AI developers from huge civil payouts if they post safety and transparency protocols, and a rival that would instead lean on tough, enforceable safety rules, incident reporting, and outside audits." (Source: Jo Marquez, "OpenAI and Anthropic Clash Over Illinois AI Liability Bill," Hoodline, May 6, 2026, https://hoodline.com/2026/05/springfield-showdown-openai-anthropic-brawl-over-illinois-ai-shield-law/ .)

SB 3444 (the Artificial Intelligence Safety Act) "would block certain civil lawsuits against a 'developer' of a frontier model for specific types of 'critical harms' as long as the developer did not intentionally or recklessly cause those harms and had published a safety and transparency protocol before releasing the system." (Source: same.) "Critical harm" is defined as "incidents involving the death or serious injury of 100 or more people, or at least $1,000,000,000 in property damage." (Source: same.) By contrast, SB 3261 (the Artificial Intelligence Public Safety and Child Protection Transparency Act) "would put more teeth into the rules. Frontier developers and large chatbot providers would be required to publish public safety and child-protection plans, accept third-party audits, and participate in an incident-reporting system for serious events run by the Attorney General." (Source: same.)

The industry is publicly split. OpenAI has signaled support for SB 3444; Anthropic has come out against it. Anthropic's public statement, as reported by Hoodline citing WIRED, is direct: "We are opposed to this bill," with the lab arguing "that transparency laws should pair public safety with genuine accountability instead of offering broad liability protections to frontier AI developers." (Source: same.) Critics frame SB 3444's effect bluntly; Thomas Woodside of the Secure AI Project told WIRED that the proposal "would take the extreme step of nearly eliminating liability for severe harms." (Source: same.)

SB 3444 also includes a federal-preemption hook: it "would stop applying if Congress or a federal agency adopts overlapping rules. That design choice intersects with the White House effort to craft a national AI framework and the possibility of federal preemption." (Source: same.)

Active matter motion practice as the doctrinal frontier

The matters described in Part I are themselves the active litigation surface where frontier labs are navigating doctrinal questions in real time. Conway's product holding in Garcia (Part I §I.7), the Raine amended-complaint safeguard-removal theory (Part I §I.3 / Part IV §IV.3.e), the not-yet-public Nov-2025 college-graduate matter (Part I §I.4), the four settled wrongful-death matters (Part I §I.6), and the PA AG action against Character.AI (Part I §I.5 / Part IV §IV.3.f) collectively constitute the lab-side legal frontier. Defense-side strategy across the wave includes pushing back on procedural clarity — arguing in each new matter that the framework should not attach in this case, on these facts, in this circuit — as discussed at Part IV §IV.1's grad-intern callout.

GRAD-INTERN — The IL split as preview of the multi-state landscape

The Illinois split between SB 3444 (liability shield with transparency-as-safe-harbor) and SB 3261 (transparency-with-teeth and incident-reporting) is best read as a preview of the contest that will play out in many states in the next eighteen-to-thirty-six months. The K&L Gates analysis observes that "the first bellwether jury verdicts will establish critical precedents for the cases that follow" (Source: K&L Gates, "AI Product Liability: The Next Wave of Litigation," March 27, 2026, https://www.klgates.com/AI-Product-Liability-The-Next-Wave-of-Litigation-3-27-2026 ); the same dynamic operates at the legislative layer, where the first state(s) to adopt either pattern will shape what other states consider feasible. A federal-preemption resolution in either direction (preemption of state tort with a federal substitute; or no preemption with state-by-state variation) will reshape both contests. For the abstracted product-liability attorney working in this space, the practical consequence is that the operative liability regime in a specific matter depends on (a) the state of the active matter, (b) federal-preemption posture at the time of trial, and (c) any state-level liability-shield or transparency-with-teeth statute that has been enacted by then. Watching the Illinois contest closely is, on this reading, watching a leading indicator.

§VI.3 Most Fruitful Surfaces for Product-Liability Attorneys *(Q12)*

Five surfaces where doctrinal development is most active

The R1 corpus surfaces five surfaces where commentary, court rulings, and active matter pleadings are converging. They are not mutually exclusive — most active matters plead multiple surfaces — and they vary in maturity. They are presented in approximate descending order of doctrinal development at access date:

#### §VI.3.1 — Conway product-not-speech (the threshold surface)

Conway's ruling in Garcia — that Character.AI is "a 'product for the purposes of Plaintiff's claims [arising] from defects in the Character A.I. app rather than ideas or expressions within the app'" (Source: 785 F. Supp. 3d 1157, 1180 (M.D. Fla. 2025), as cited in McGuireWoods, "Can Social Media or AI Be a Defective Product?", March 18, 2026, https://www.mcguirewoods.com/client-resources/alerts/2026/3/can-social-media-or-ai-be-a-defective-product/) — is the doctrinally-foundational holding for product-liability analysis of consumer-AI-output products. As the Transparency Coalition's Bruce Barcott describes the ruling, "Significantly, Judge Conway ruled that Character.AI is a product for the purposes of product liability claims, and not a service." (Source: Bruce Barcott, "In early ruling, federal judge defines Character.AI chatbot as product, not speech," Transparency Coalition, May 21, 2025, https://www.transparencycoalition.ai/news/important-early-ruling-in-characterai-case-this-chatbot-is-a-product-not-speech .) The ruling did not finally adjudicate any First Amendment question — as Barcott notes, "That doesn't necessarily mean Judge Conway decided that the Character.AI interactions with Setzer were or were not protected speech—merely that a protected speech claim is not so clearly valid that the case should not proceed." (Source: same.) The Tech Justice Law Project framed the ruling's signal directly: it "sends a clear signal to companies developing and deploying LLM-powered products at scale that they cannot evade legal consequences for the real-world harm their products cause, regardless of the technology's novelty." (Source: same Transparency Coalition article quoting TJLP.)

Settlement of Garcia mooted the appellate test of Conway in the Eleventh Circuit (Part III §III.4 grad-intern callout); the holding remains on the books at trial-court level, available as persuasive authority in subsequent matters in other circuits. Whether the Conway product holding survives a fully-litigated test in another matter is the central open question.

#### §VI.3.2 — Component-part-manufacturer doctrine reaching upstream

Conway's ruling in Garcia permitted theories aimed at the upstream foundation-model technology provider to proceed at the pleading stage. The Transparency Coalition reporting notes that the ruling dismissed claims against Alphabet Inc. (Google's parent) but allowed most other claims — including those against Google LLC, the foundation-model technology provider — to proceed. (Source: same Transparency Coalition article.) As K&L Gates frames the doctrinal trajectory: "liability theories are moving up and down the AI supply chain as plaintiffs explore component-part and substantial-participation theories that can reach upstream and downstream actors." (Source: K&L Gates URL above.) Component-part-manufacturer doctrine is the doctrinal mechanism by which deployer-defendant litigation (Layer 4 in Part III) produces evidentiary and doctrinal pressure at the foundation-lab level (Layer 3).

The doctrinal frontier here is whether component-part liability survives summary judgment, whether other circuits adopt it, and how it interacts with foundation-model licensing arrangements — most foundation-model providers limit their licensees' liability through ToS provisions. (See Part IV §§IV.2 and IV.3.a for the doctrinal substrate.)

#### §VI.3.3 — Amended-complaint safeguard-removal (the Raine theory)

The Raine v. OpenAI amended complaint pleads that OpenAI removed protective behaviors from a deployed product, and that removal can ground a separate cause of action distinct from product-liability failure-to-warn or design-defect theories. (Part I §I.3; Part IV §IV.3.e.) The theory's doctrinal character is novel: it pleads not the absence of safeguards (the standard failure-to-warn frame) but the affirmative removal of safeguards that previously existed. Whether courts recognize the theory as a distinct cause of action with distinct elements — and what discovery-scope follows — is unsettled. The motion practice in Raine is the live development surface.

#### §VI.3.4 — Foundation-model-versus-deployer doctrinal distinction

Part V drew the lab posture distinction between foundation-model labs (Anthropic, OpenAI as a foundation-lab; Google DeepMind) and deployer-defendants (Character.AI; OpenAI in its ChatGPT-deployer capacity). The doctrinal corollary is whether liability attaches at the foundation-model layer, the deployer layer, or both. Conway in Garcia answered yes to both at the pleading stage. Whether subsequent matters distinguish more sharply — keeping the deployer in but dismissing the foundation-lab, or vice versa — is open. The matter calculus differs: deployer defendants often have less capital depth than foundation-model labs but are more directly attributable; foundation-model labs are deeper-pocketed but more attenuated from the immediate user experience and may invoke licensing-chain defenses.

#### §VI.3.5 — Regulatory-enforcement parallel (state-AG actions)

The PA AG action against Character.AI (Part I §I.5; Part IV §IV.3.f) and the broader pattern of state-AG and consumer-protection-statute exposure at the deployment surface (Part III §III.6) constitute a parallel litigation track. State-AG enforcement actions develop doctrinal questions around statutory consumer-protection theories, deceptive-trade-practices statutes, and minor-protection regimes that operate alongside private wrongful-death actions and produce distinct discovery surfaces. For the abstracted product-liability attorney, regulatory-enforcement matters are both potential reference points (settlement records, consent decrees, factual findings) and potential parallel proceedings depending on jurisdiction.

GRAD-INTERN — How the five surfaces compose in actual matter strategy

The five surfaces enumerated above are not strategic choices in the sense of mutually-exclusive options; in practice, an active matter will plead multiple surfaces. Garcia pleaded product liability (§VI.3.1), component-part doctrine (§VI.3.2), and consumer-protection statute violations (§VI.3.5 analog at the Florida state level). Raine pleads product liability (§VI.3.1), failure-to-warn / safeguard-removal (§VI.3.3), foundation-model-as-deployer (§VI.3.4 — OpenAI in its dual capacity), and additional theories. The strategic question for a hypothetical attorney evaluating a new matter is which combination of surfaces fits the factual record, which jurisdiction's procedural and substantive law applies, and which surfaces are most-developed at the relevant date. The K&L Gates analysis observes that "Plaintiffs, by contrast, increasingly draft complaints to target the architecture of the deployed system—guardrails, defaults, escalation pathways, and marketing—so the case looks like a product-defect dispute instead of a content dispute." (Source: K&L Gates URL above.) This architecture-as-defect framing is doctrinally adjacent to all five surfaces above and is a current pleading trend visible across the active wave. Whether and how this framing succeeds at trial remains to be determined; the bellwether jury verdicts that K&L Gates references will be the doctrinal milestones. Nothing in the foregoing is legal advice; it is a description of the doctrinal surfaces visible in the public record.

§VI.4 Most Studied / Most Understudied *(Q13)*

A meta-cartographic note

A research project of this scope produces, almost incidentally, an answer to where the literature is dense and where it is thin. The R1 corpus we have surveyed varies substantially across the landscape. The variation is itself diagnostic: it indicates where doctrinal development has produced enough commentary to support analysis, and where it has not. A candid map of corpus density supports the abstracted attorney's working knowledge of the field as much as the doctrinal substance does.

Most studied (substantial corpus density)

Most understudied (substrate gaps in our corpus and likely the broader literature)

GRAD-INTERN — Why the corpus map matters as much as the corpus

The map of where commentary is thick and thin tells the abstracted attorney where pre-existing analysis can short-circuit independent doctrinal work, and where independent doctrinal work is required. A matter that turns on Conway-applied-elsewhere can rest on substantial existing commentary; a matter that turns on hosting-layer component-part theory must develop the doctrinal frame from a much thinner base. The map is also load-bearing for honest scoping of any specific research effort: where the public corpus is thin, surrogate sources (PACER filings; state-AG press releases; lab transparency reports; trade-press niche publications) become the operative substrate, and the time-cost of the analysis rises. The corpus density gradient also tracks the doctrinal-development gradient: surfaces with thick commentary tend to have more-mature doctrinal handles; surfaces with thin commentary tend to have more open questions. None of the foregoing is legal advice; it is a candid description of where the public corpus is dense and where it is sparse, intended to support the abstracted attorney's mental model of the field.

§VI.5 Recommendations Synthesis

Recommendations for attention, not for action

This site has surveyed a doctrinal landscape, a per-layer ecosystem, a legal landscape, and the lab-and-platform-side response to all of it. The synthesis below is descriptive: it identifies the surfaces, dynamics, and developments most-worth attention at access date 2026-05-07, organized as a working orientation rather than a plan. Whether any specific surface is the right strategic vector for any specific factual record is, again, a question for licensed counsel admitted in the relevant jurisdiction working with the actual record. Per the methodology established at companion sections: research-not-advice; descriptive-not-prescriptive.

Watch the bellwether trajectory. The K&L Gates observation that "the first bellwether jury verdicts will establish critical precedents for the cases that follow" (Source: K&L Gates URL above) is the central temporal frame for the field. Conway is on the books at trial-court level; the appellate development that would harden Conway into binding precedent did not occur in Garcia (settlement); subsequent matters that proceed to trial — Raine, the not-yet-public Nov-2025 college-graduate matter, any new wave of matters — are the bellwether candidates.

Watch the Illinois contest. The SB 3444 / SB 3261 split (§VI.2) is a leading indicator for the multi-state liability-shield-vs-transparency-with-teeth contest that will play out across 2026–2027.

Watch the federal-preemption posture. Whether (and which) federal AI legislation passes — the AI LEAD Act in pending form; any successor — will reshape both the state-tort layer and the legislative-shield layer simultaneously. Federal preemption analysis under Murphy v. NCAA (Part IV §IV.1) supplies the doctrinal frame; the political question of which way Congress moves is not a question this research takes a position on.

Watch the discovery-scope contests. The Raine matter, with the safeguard-removal allegation, will produce some of the first contests over what AI-lab internal safety-evaluation documentation must be produced in discovery, under what protective orders, at what stage. The Layer-7 documentary record (Part III §III.7) is publicly extensive; the discovery surface into substantive operational practice is the next frontier.

Watch the Sensitive-conversations addendum and analogous lab-side post-Raine artifacts. Per §V.2: OpenAI's Sensitive-conversations addendum is named on the safety hub at access date; analogous post-litigation lab-side documentation is a likely-emerging artifact class.

Watch the regulatory enforcement track. State-AG actions (PA AG; Kentucky AG inferred from §III.6; subsequent jurisdictions) will produce parallel doctrinal development on consumer-protection-statute and minor-protection-regime theories alongside the private-wrongful-death-action track.

Pay attention to corpus density. Per §VI.4: where commentary is thick, pre-existing analysis short-circuits independent doctrinal work. Where commentary is thin, independent doctrinal work is required. Calibrate research time-budget to the density gradient.

Treat the labs' published material as documentary baseline. Per §V.4 grad-intern callout: convergent industry practice across Anthropic, OpenAI, and Google DeepMind produces a documentary record that defendants and plaintiffs both reach for in standard-of-care motion practice. Statutory transparency obligations (SB 53; analogous future statutes) accelerate this dynamic.

Part VI closing forward-pointer

Part VI has synthesized the strategic surfaces where the doctrinal landscape of Part IV and the lab posture of Part V intersect. The five most-fruitful surfaces (§VI.3) constitute the substantive heart of Q12; the most-studied / most-understudied map (§VI.4) constitutes Q13; the technical and legal challenges the labs are themselves navigating (§§VI.1–VI.2) constitute Q8 and Q9 partial.

Part VII takes up Q14 — the explicitly UNANSWERED research question on cross-functional coordination between AI-ecosystem departments and in-house legal counsel. Q14's UNANSWERED status is itself a substantive observation: the public record on this dimension is thin. Part VII treats the research-gap transparently per the ground-or-flag methodology established at Part III's grad-intern callouts.

Part VIII (appendices) provides the supporting documentary apparatus: case-set fact sheet (post-anonymization), per-layer fact sheets, citation index per CITATION_STANDARD L1/L2/L3, methodology and source taxonomy, and glossary.

Several Part-VI threads run forward: