British Columbia’s complaint against OpenAI asks a question that social-media companies will recognize: what if the danger lies in how the product was built to keep a person engaged? Filed on September 21, 2026, by the Province and the Peace River South school board, it alleges that ChatGPT reinforced the violent ideation of the Tumbler Ridge shooter. It challenges design choices and the company’s response to an allegedly known threat, and proposes safeguards. These are allegations in a complaint, not findings about what the system or its makers did.[1]

The design argument is the important development here. In the social-media cases, Section 230 forced a distinction between responsibility for somebody else’s speech and responsibility for a company’s own product choices. Some plaintiffs found room in that distinction. The chatbot cases bring a similar argument to a system that generates a conversation, with a different problem of proving how the challenged design produced the harm.[2]

The attraction is easy to understand. A court can be asked to examine a memory setting, a reward signal or an escalation mechanism rather than decide whether an unpleasant answer should have been published. But those choices still have to be identified, and their significance explained. The analogy gives plaintiffs a way to frame the case. It does not supply the causal account.

How design became a way in

Section 230(c)(1) prevents treating a provider or user of an interactive computer service as the publisher or speaker of information supplied by another information content provider. For claims directed at a platform’s handling of users’ posts, that can be a formidable obstacle. The design approach changes the proposed source of responsibility: the plaintiff asks whether a duty attaches to the defendant’s own construction of the product, independently of its publication of third-party material.[3][4]

Lemmon v. Snap, Inc. shows the distinction at work. Parents alleged that Snapchat’s speed filter and reward system encouraged dangerous driving. In 2021, the Ninth Circuit held that Section 230 did not bar the pleaded negligent-design theory because the alleged duty arose from Snap’s role as product designer, independently of publishing users’ content, and the claim did not turn on information supplied by another content provider. That opened a route through the immunity question; it did not establish a defect or causation.[4]

The November 2023 social-media addiction litigation order made the sorting more explicit. The federal district court allowed certain alleged defects, including the absence of effective parental controls and an adequate age-verification process, past Section 230. It barred others, including the alleged use of algorithms to promote addictive engagement, where the challenge targeted publishing third-party content. The word design was doing useful work, but it was not a password.[5]

One distinction in that order is especially revealing for chatbots. Default session limits were barred because they would require publishing less third-party content; tools that let users choose to restrict their own use were not. An interruption can therefore be a safety intervention and still encounter a publishing defense. For a system generating its own replies, the source of the information becomes part of what must be tested, not assumed away.[5]

The shift is worth highlighting: attention moves from the harmful post to the product choice and the duty attached to it. It is a reading of these rulings, not a claim that all plaintiffs shared one strategy or that product liability was invented for social media. Nor did the shift make Section 230 disappear. In Calise v. Meta Platforms, Inc., the Ninth Circuit asked both where the asserted duty came from and what performing it would require. In its April 2026 Doe 1 v. Meta opinion, it still barred claims alleging that Facebook’s design encouraged the spread of violence-inciting content.[6]

For AI companies, the significance is a route of scrutiny. If the challenged act is choosing how a system behaves, a defense concerned with publishing somebody else’s speech may not answer the whole claim. Whether it answers a particular chatbot claim still depends on the information, its source and the duty alleged.

The argument arrives at the chatbot

The B.C. complaint alleges that OpenAI encouraged extended interaction, enabled persistent memory by default and tuned ChatGPT toward affirming users. It proposes alternatives that would refuse, interrupt or escalate repeated violent-intent interactions. It also alleges, drawing on press-reported whistleblower accounts, that internal reviewers recommended a police referral and leadership declined. The filing does not establish those deployment and decision records.[7]

The complaint organizes a recognizable story around those choices: a feature sustains interaction, the interaction becomes dangerous, warnings are allegedly known, and an intervention is proposed. Knowledge matters to its negligence account; it is not an additional element of the strict design-defect test discussed below. A system intended to continue a conversation may need safeguards that interrupt it. That is the argument to test, rather than assume that every additional exchange is either harmless assistance or evidence of a defect.

The complaint advances California law. The risk-benefit branch of Barker v. Lull Engineering Co. makes the design trade-off particularly consequential. Once the plaintiff makes a prima facie showing that the design proximately caused the injury, the defendant bears the burden of persuasion that its benefits outweigh its risks. Feasibility, cost and adverse consequences of an alternative design belong in that balance. Negligent design asks a different question about the reasonableness of the manufacturer’s conduct. California law’s application to these Canadian plaintiffs remains a pleaded position, not a choice-of-law ruling.[8]

The procedural route matters to this comparison. A defendant can answer or seek dismissal under Rule 12, which permits challenges to jurisdiction, venue and legal sufficiency. The complaint places design and warning decisions in California. With foreign public bodies suing over Canadian injuries, the court’s authority to hear the dispute and the applicable law cannot simply be assumed. Surviving a pleading challenge would still leave proof of defect and causation for the evidentiary stages, including any Rule 56 motion. These are possible routes and issues to assess, not a report that these defendants have taken them.[20]

That distinction prevents engagement from doing three jobs at once. A wish to retain users might explain a commercial choice. Evidence about an intervention might illuminate a design trade-off. Neither, on its own, establishes the causal connection to an injury. A business incentive is a starting point for inquiry, not a substitute for the rest of the tort.

Engagement needs a translation

The tempting contrast is between a predictable social-media machine and an unpredictable chatbot. It is too neat. Google’s 2016 account of YouTube recommendations described learned models for candidate generation and ranking. Learning was already part of the machinery on the social-media side of the comparison.[9]

Nor does reinforcement learning from human feedback simply mean training for engagement. In the InstructGPT research, human demonstrations supported supervised fine-tuning, while rankings trained a reward model that the system was then optimized against. The training method does not, by its name, tell us whether the relevant objective was helpfulness, agreement, safety or keeping somebody online. A claim about a particular system needs evidence of what its signals rewarded and how those signals interacted.[10]

OpenAI’s account of its April 2025 GPT-4o sycophancy problem supplies a concrete illustration. Its early assessment was that individually promising changes, including user feedback and memory, may have contributed when combined. The company believed those changes weakened its primary reward signal’s restraint on excessive agreeableness; it lacked evidence that memory broadly increased the problem. The company said positive evaluations and user tests supported launch despite qualitative reservations from expert testers; it lacked specific deployment evaluations for sycophancy. This is the company’s explanation of that update, not proof about the configuration involved in B.C.’s allegations.[11]

For this argument, the example matters because it identifies decisions someone could examine: which signal was added, what was tested and how the launch decision was made. It does not show that reinforcement learning is inherently an engagement-maximizing process, or that this update caused the injuries alleged here. My point is narrower: variable model behavior can still have an intelligible design history.

The same care is needed with the research comparison. The Surgeon General’s 2023 social-media advisory drew on what it described as a substantial, non-exhaustive review of the evidence, while identifying major gaps, predominantly correlational prior research and a potentially bidirectional relationship. It nevertheless urged companies to avoid features aimed at maximizing time, attention and engagement. That is precautionary design guidance, not settled causation in an individual case.[12]

That record cannot simply be carried across to a conversational system. Findings about a particular pattern of social-media exposure do not establish what repeated affirmation by a chatbot does to a different user, still less the role it played in a particular act of violence. The useful comparison is between the specific behavior, population and alleged injury. Calling both products engaging makes the question worth asking; it does not answer it.

A conversation is evidence, not an explanation

The contrast in this pleading is between selecting someone else’s material and generating a reply addressed to the user. That is the difference this complaint asks us to examine. It alleges reinforcement within an exchange, while anticipating amendment after production of the shooter’s ChatGPT logs. The filing therefore cannot be read as though it already supplies an authenticated conversation proving the alleged defect.[13]

Even with the full exchange, there would be work to do. Which design choice explains the behavior? Would the proposed safeguard have changed the interaction? Would that change have affected the injury? In this case, the causal account must also address an intentional criminal act and losses suffered by third parties. The causal requirement does not disappear because the product communicates in sentences.

A translucent blue paper ribbon crosses two charcoal checkpoints, one marked in ochre and the other left open.

AI-generated editorial illustration · Detection and intervention. A paper ribbon passes a mark of recognition and continues through an unfinished boundary.

There is a more concrete theory in the complaint’s distinction between detecting dangerous material and stopping generation. It alleges that a moderation capability was available but that ChatGPT could identify certain content without being configured to stop generating it. If established, that would focus the inquiry on a particular gap between recognition and intervention. But classifying violent content is not necessarily the same task as identifying dangerous reinforcement of a user’s intent; the proposed intervention needs evidence of how it would work. Under Barker’s strict-liability risk-benefit test, the defendant bears the burden of persuasion on that balance after the plaintiff’s causal showing.[14]

The policy argument deserves a fair hearing in both directions. The complaint acknowledges several safety tiers in the February 12, 2025 Model Spec and argues that instructions outside its narrow prohibited category were too weak. Yet the Spec also restricted dangerous assistance, prohibited encouraging illicit behavior and conditioned its good-intentions default. Its imminent-harm provision described cautious guidance, not the interruption and escalation machinery at issue here. The dispute is about the strength and operation of those safeguards, not simply whether a prohibition appeared under the right heading.[15]

OpenAI also said production models did not yet fully reflect the Spec. That leaves several possible theories: unsafe intended behavior, inadequate implementation, or a design that relied too heavily on instructions the model might disregard. The complaint raises that last possibility by alleging both departures from safety instructions and performance as designed. Barker makes the underlying point useful: following the maker’s plan does not establish that the plan was safe. A policy, a deployed safeguard and a generated response need to be connected; none is an adequate stand-in for the other two.[16]

The expression objection

A central objection to treating these choices as product design is that instructions about what a chatbot should refuse or say may themselves concern expression. Winter v. G.P. Putnam’s Sons, applying California law, rejected strict products liability for ideas and expression in a book, while distinguishing functional tools and discussing malfunctioning software as a possible analogy. Moody v. NetChoice, LLC recognized that algorithms can implement protected editorial judgment. Neither opinion settled the status of chatbot output. Moody also reserved the question of feeds driven solely by users’ behavior, without independent content standards. That reservation makes it harder to treat every engagement-tuned system as an exercise of the editorial judgment the Court described.[17]

The May 2025 Garcia v. Character Technologies, Inc. ruling illustrates the narrow opening. Applying Florida law at the pleading stage, the district court treated the application as a product insofar as claims concerned app defects rather than its ideas or expressions. It was not prepared at that stage to accept the asserted speech protection for the output. The court later recorded a reported settlement and dismissed the action in January 2026, subject to reopening. The May ruling remains the comparison here, not binding California authority or a general holding that generated language lacks protection.[18]

This objection makes precision more valuable. A claim about memory defaults or an escalation mechanism needs to explain the defect in that feature. A claim about harmful words must confront the expressive character of the alleged wrong. Section 230 adds its own inquiry where invoked: who was responsible, in whole or in part, for creating or developing the relevant information, and does the proposed duty treat the defendant as its publisher?[3]

B.C. squarely pleads that ChatGPT created the content and that no third-party content lies in the causal chain. It also places the alleged defect partly in the generated content itself. That strengthens its answer to the third-party-information question while sharpening the objection about expression. Both arguments need to be met.[3][19]

The complaint also joins product and organizational choices: it calls a management-override structure and a deactivation mechanism allowing re-access design defects. Whether that characterization works is another contested step, not a reason to ignore the allegations. Its separate known-threat and police-warning theories still require their own duty and causation analysis.[19]

Practical implications

For an AI governance program, I would put this question near the front of the design review: when should a successful interaction end? If continuing the exchange is treated as a good result, the program needs a way to recognize the exchanges in which continuing is precisely the problem. That means testing the proposed interruption, not merely recording that a safety policy exists.

The litigation comparison suggests practical work at identifiable decision points. Before changing a reward signal or memory behavior, test the combined system for the failure modes the change could worsen. When user preference and expert observations disagree, identify who can hold a release and what evidence resolves the disagreement. For an escalation mechanism, test whether detection actually leads to an effective intervention. OpenAI’s own post-incident commitments also called for behavioral concerns to block launches despite favorable metrics. These recommendations apply that concern to the design and proof problems here; they are not findings that a particular control is legally required.[11]

Records matter because they let those decisions be examined later: the version deployed, the alternatives considered, the tests run and the reasons for proceeding. Under Barker’s risk-benefit framework, after the plaintiff’s required causal showing, evidence of the design’s benefits and risks is part of the defendant’s own burden. Careful process is not itself a defense to strict liability: Barker permits the design to be judged unsafe in hindsight despite reasonable precautions. For strict design liability, the record must illuminate the product’s risks and benefits. Evidence of reasonable precautions can still matter to negligence, including the pleaded response to a known threat.[8]

The design approach gives plaintiffs a serious question to ask of AI companies. Its strength will depend on translating that question into a specific feature, duty and causal account, with room for the strongest defense. My judgment is that governance should work at those same points before litigation arrives. An unpredictable answer does not make the decisions behind the system beyond examination.

Source scope

This essay compares identified pleadings and rulings, with source review on September 24 and 29, 2026. It does not certify complete current dockets or all subsequent treatment. The design-defect discussion focuses on Barker’s risk-benefit branch; it does not assess the separately pleaded consumer-expectations test. The Model Spec discussion concerns the February 12, 2025 version, not every version in force during the alleged events.

The technical papers and company account illustrate mechanisms; they do not establish the configuration or causation in the B.C. case. Defense arguments are the author’s analysis, not representations of filings reviewed from these defendants. Deployment records and authenticated case-specific conversations could materially change the assessment. The governance recommendations are not a safe harbor or a prediction of liability.

Notes

  1. Complaint, His Majesty the King in Right of the Province of British Columbia v. Altman, No. 4:26-cv-10743-KAW (N.D. Cal. Sept. 21, 2026), ECF No. 1, caption, paragraphs 1–8, 40–55 and counts 1–8, original (B.C. complaint; ECF-stamped docket number; the typed caption instead gives a 3:26 prefix). The named school-board plaintiff is the Board of Education of School District No. 59 (Peace River South). ↩

  2. The comparison is the author’s synthesis of the authorities in notes 3–6 and the pleaded choices in note 7, not a holding that the social-media cases govern generative AI. ↩

  3. 47 U.S.C. § 230(c)(1), (f)(3), official text. ↩

  4. Lemmon v. Snap, Inc., 995 F.3d 1085 (9th Cir. 2021), opinion, slip op. at 11–18. The reversal addressed Section 230, not the ultimate merits. ↩

  5. In re Social Media Adolescent Addiction/Personal Injury Products Liability Litigation, No. 4:22-md-03047-YGR (N.D. Cal. Nov. 14, 2023), ECF No. 430, order, ECF No. 430, at 14–17. This is the November 14, 2023 order’s treatment of identified allegations, not a statement that every surviving theory ultimately prevailed. In related coordinated proceedings, the August 10, 2026 Ninth Circuit consolidated opinion, at 29, dismissed interlocutory appeals for lack of jurisdiction; that disposition was not a merits affirmance. ↩

  6. Calise v. Meta Platforms, Inc., 103 F.4th 732 (9th Cir. 2024), opinion, slip op. at 17–18, 24–25; Doe 1 v. Meta Platforms, Inc., No. 24-1672 (9th Cir. Apr. 28, 2026), opinion, slip op. at 13–19. ↩

  7. B.C. complaint paragraphs 5–7, 40–43, 47–51, 111–14. Knowledge, tuning, deployment and management decisions are pleaded allegations. ↩

  8. Barker v. Lull Engineering Co., 20 Cal. 3d 413, 429–35 (1978), opinion; B.C. complaint paragraphs 119–21. The risk-benefit burden allocation described here concerns strict-liability design defect, not negligent design. ↩

  9. Paul Covington, Jay Adams & Emre Sargin, Deep Neural Networks for YouTube Recommendations (2016), authors’ institutional publication page and abstract. This historical architecture illustrates learning in recommendations, not the configuration of every platform or any litigated system. ↩

  10. Long Ouyang et al., Training Language Models to Follow Instructions with Human Feedback, arXiv:2203.02155v1 (Mar. 4, 2022), paper, at 1–4. The distinction between a training method and a particular engagement objective is the author’s inference from the described procedure. ↩

  11. OpenAI, Expanding on what we missed with sycophancy (May 2, 2025), sections on the April 25 update and its review process. The company characterized the combined training effects as an early assessment. ↩

  12. Office of the U.S. Surgeon General, Social media and youth mental health (2023), printed pp. 3, 11–12, 16; official NLM PDF mirror. The limits on transferring this evidence to the alleged chatbot interaction are the author’s analysis, not a finding of the advisory. ↩

  13. B.C. complaint paragraphs 6–7, 47, 55, 121 and paragraph 7 n.2 (PDF p. 4). A conversation’s evidentiary limits and the intervening causal questions are the author’s analysis of the pleaded theory, not findings about the unreleased logs. ↩

  14. B.C. complaint paragraphs 51, 114; Barker, 20 Cal. 3d at 431–32. The capability, integration and efficacy of the proposed intervention are not independently established by these allegations. ↩

  15. OpenAI, Model Spec, February 12, 2025, sections “Do not generate disallowed content,” “Don’t provide information hazards,” “Try to prevent imminent real-world harm” and “Do not facilitate or encourage illicit behavior”; compare B.C. complaint paragraph 48. ↩

  16. Model Spec, February 12, 2025, overview; B.C. complaint paragraphs 50, 54, 123; Barker, 20 Cal. 3d at 429, 434–35. Applying the plan-versus-safety distinction to model behavior is the author’s analysis. ↩

  17. Winter v. G.P. Putnam’s Sons, 938 F.2d 1033, 1034–38 (9th Cir. 1991), opinion reproduction, section 12.3; Moody v. NetChoice, LLC, 603 U.S. 707 (2024), opinion, majority slip op. at 20–22 & n.5. The software analogy in Winter was dictum. The implication of Moody’s reservation for an engagement-tuned system is the author’s inference, not a holding about chatbots. ↩

  18. Garcia v. Character Technologies, Inc., No. 6:24-cv-01903 (M.D. Fla. May 20, 2025), ECF No. 115 (filed May 21, 2025), order, at 25–32, 34–36. The speech argument addressed users’ asserted right to receive information; the ruling did not finally resolve the classification. Subsequent history: ECF No. 244 (Jan. 7, 2026), at 1–2, recorded counsel’s report of settlement and dismissed without prejudice, allowing 90 days to seek reopening or submit a final order. The June 11, 2026 order, ECF No. 273, at 1 n.1, records an April 10 extension of that deadline. The August 4, 2026 order, ECF No. 277, at 1, extended the dismissal deadlines for another 90 days. No final with-prejudice disposition is inferred. ↩

  19. B.C. complaint counts 1, 4 and 5, paragraphs 45, 50, 55, 115, 121–22. This essay’s design focus does not assess entitlement to recover the Province’s public expenditures or the school district’s losses under every pleaded theory. ↩

  20. Fed. R. Civ. P. 12(a), (b) and 56(a), (c); B.C. complaint paragraphs 13–25. The procedural comparison is the author’s analysis of available vehicles, not a current docket-status assertion or a conclusion that any jurisdictional objection would succeed. ↩

Case citations

  • His Majesty the King in Right of the Province of British Columbia v. Altman, No. 4:26-cv-10743-KAW (N.D. Cal.), Complaint, ECF No. 1 (Sept. 21, 2026).
  • Lemmon v. Snap, Inc., 995 F.3d 1085 (9th Cir. 2021).
  • In re Social Media Adolescent Addiction/Personal Injury Products Liability Litigation, No. 4:22-md-03047-YGR (N.D. Cal. Nov. 14, 2023), ECF No. 430.
  • People of the State of California v. Meta Platforms, Inc., No. 24-7032 et al. (9th Cir. Aug. 10, 2026).
  • Calise v. Meta Platforms, Inc., 103 F.4th 732 (9th Cir. 2024).
  • Doe 1 v. Meta Platforms, Inc., No. 24-1672 (9th Cir. Apr. 28, 2026).
  • Barker v. Lull Engineering Co., 20 Cal. 3d 413 (1978).
  • Winter v. G.P. Putnam’s Sons, 938 F.2d 1033 (9th Cir. 1991).
  • Moody v. NetChoice, LLC, 603 U.S. 707 (2024).
  • Garcia v. Character Technologies, Inc., No. 6:24-cv-01903 (M.D. Fla. May 20, 2025), ECF No. 115 (filed May 21, 2025); settlement-dismissal order, ECF No. 244 (Jan. 7, 2026).

Disclaimer: AI Lex Intelligence is published for informational purposes only. It does not constitute legal advice, and no attorney-client relationship is formed by reading or receiving this publication. Readers should consult qualified legal counsel about specific legal matters.

AI Lex Intelligence with Zola Valashiya. Independent analysis of artificial intelligence and the law.