The number is 100. That is how many people must die, or suffer serious injury, before OpenAI faces legal liability under legislation the company is actively supporting in the Illinois General Assembly. For an attorney who files a single AI-fabricated citation in a federal court proceeding, the number is different. It is one. Illinois Senate Bill 3444, the Artificial Intelligence Safety Act, defines “critical harm” as the death or serious injury of 100 or more people, at least $1 billion in property damage, or assistance in creating chemical, biological, radiological, or nuclear weapons. A developer meeting the qualifying compute threshold faces liability for these events only if it acted intentionally or recklessly and failed to publish a safety protocol on its website. A developer whose tool produces fabricated case citations, invented judicial holdings, and nonexistent statutes faces nothing under this statute. The threshold is 100 deaths. A bad brief is not 100 deaths. For attorneys, the threshold is different. Under Model Rule 3.3, a practitioner may not knowingly make false statements of law or fact to a tribunal. Federal courts have read the verification obligation broadly: an attorney who submits AI-generated citations without independent review has violated the rule, regardless of the tool’s marketing claims or the vendor’s assurances. The gap between those two thresholds is the argument. And that gap has a lobbyist. ## The Architecture of Liability OpenAI sent its government affairs team to Springfield not because it is worried about mass casualties. It sent them because it is worried about the lawsuits it knows are coming. Caitlin Niedermeyer, a member of OpenAI’s Global Affairs team, testified in support of SB 3444 while simultaneously arguing against a “patchwork of inconsistent state requirements” and calling for a unified federal framework. The position is coherent as strategy: support a liability shield at the state level that establishes a template, then argue against state-level variation that creates unpredictable exposure. If Illinois passes SB 3444 and other states adopt its structure, OpenAI and its peers will have helped write the national liability architecture before any federal law exists. The bill applies only to “frontier models” trained on more than $100 million in compute. That threshold covers OpenAI, Google, Anthropic, Meta, and xAI. These are the same companies whose tools run across legal practices, healthcare systems, and financial services. The same companies whose hallucination rates Stanford’s RegLab documented at 17 to 33 percent in peer-reviewed research published in 2025. The same companies whose subscriber agreements, without exception, disclaim warranties and cap damages at subscription fees paid. As of mid-April 2026, the bill is in committee markup in the Illinois General Assembly, with a floor vote expected before summer recess. No federal AI liability law exists. AI companies collectively spent $50 million lobbying federal lawmakers in the first nine months of 2025, according to the nonpartisan organization IssueOne. OpenAI is opening a Washington, D.C. office in early 2026. The company is not a passive observer of the regulatory process. SB 3444 reflects its preferred outcome. The bill’s safe harbor is structured so qualifying labs can achieve it without accuracy disclosure. A developer publishes a safety protocol, avoids intentional harm, and qualifies. The bill does not require disclosure of hallucination rates by query type. It does not require vendors to publish error distributions. It does not require that a legal AI research tool tell a subscribing attorney the frequency at which the tool generates incorrect legal conclusions. An attorney who cannot access her tool’s error rate still faces the Comment 8 obligation to understand the technology’s risks. The framework requires verification while relieving the vendor of any duty to provide the information being verified. ## The Strongest Case for the Bill The argument for SB 3444 deserves a direct statement. Foundation models are general-purpose technologies. They do not cause harm in isolation; they enable harm when misused by downstream deployers or end users who bypass safety mechanisms or ignore output limitations. Holding the foundation model developer liable for every downstream misuse would be the equivalent of holding a steel manufacturer liable for a building that collapsed because an architect specified the wrong grade. The compute threshold is designed to create accountability for the most powerful systems while leaving smaller developers able to operate without catastrophic legal exposure. The transparency requirement, publishing safety protocols, is a genuine accountability mechanism. OpenAI spokesperson Jamie Radice framed this as “focusing on what matters most: reducing the risk of serious harm from the most advanced AI systems while still allowing this technology to get into the hands of the people and businesses of Illinois.” That is a defensible principle. The chilling effect of open-ended liability on beneficial AI development in medicine, scientific research, and public safety is real, not a corporate talking point. The argument has limits. Foundation model developers are not passive steel manufacturers. They build consumer-facing products, market them to law firms as hallucination-free, and price them as legal research subscriptions. The gap is not between a raw material and its misuse. It is between a marketed professional tool and the professional who relied on the marketing. The accuracy problem is not a downstream misuse problem. It is an architecture problem the developer controls and a marketing claim the developer made. A bill that addresses catastrophic misuse while leaving that gap intact is addressing the wrong problem for attorneys. One limitation in this analogy is worth naming before critics do. Professional responsibility and tort liability are different systems with different purposes. A bar complaint is not a lawsuit. The bar’s interest in attorney conduct and a plaintiff’s interest in compensatory damages measure different things. What they share, in the current AI liability architecture, is where the consequences land. When AI output causes harm in a legal matter, both systems send their consequences to the same address: the attorney’s license. SB 3444 does not change that address. It forecloses one of the few routes that might eventually move it. ## The Social Media Arc This approach has precedent. The record is instructive. In 1996, Congress enacted Section 230 of the Communications Decency Act, shielding platforms from liability for third-party content on their services. For two decades, those platforms built products that internal research documented as harmful, particularly to adolescent mental health, while public communications denied the connection. The statute remained intact. The harms compounded. On March 25, 2026, a Los Angeles jury found Meta and Google negligent in the design of Instagram, Facebook, and YouTube, awarding $6 million to a plaintiff who became addicted to those platforms as a minor and suffered documented psychological harm. The jury found both companies acted with malice, triggering an additional $3 million in punitive damages. One day earlier, a New Mexico jury ordered Meta to pay $375 million for violations of state consumer protection law arising from child exploitation on its platforms. Both companies are appealing. The Section 230 shield cracked eventually. Plaintiffs’ attorneys found a theory, product design liability, that the statute had not addressed. They built dockets around it for years before a jury agreed. The Los Angeles verdict is now described by plaintiffs’ counsel as “the tip of the spear.” AI companies have watched this arc. SB 3444 is the pre-emptive play: write the shield before the dockets build, not after. The difference between social media’s experience and what AI companies are attempting is timing. For Catherine, what it produces is an asymmetry she can identify today, before the question becomes a bar complaint. ## The Agreement You Already Signed Before SB 3444 existed, the liability architecture already existed in the subscriber agreement. Open the terms of service for any major AI legal research platform. Find the warranty disclaimer. In language the marketing department never quoted, you will find that the vendor makes no representations about accuracy, that the limitation of liability is capped at subscription fees paid, and that in many agreements, the subscribing firm indemnifies the vendor for consequences arising from tool use. In July 2025, a Florida federal judge imposed $85,567.75 in sanctions on an attorney who submitted fabricated AI citations across eight separate matters. The attorney used the tool. The vendor faced no liability. The subscriber agreement provided the architecture. SB 3444 proposes to convert that architecture from contract to statute. A contractual disclaimer can be litigated. It can face challenge on unconscionability grounds or state consumer protection law. A statute is harder to move. The bill does not create a new protection for AI vendors. It legislates the one they negotiated privately into public law. ## What the Number Means for Catherine In July 2025, Judge Anna Manasco disqualified two Butler Snow attorneys and referred them to the Alabama State Bar. The firm had an AI policy. The attorney, Matthew Reeves, had read it. He violated it anyway, using ChatGPT to draft a filing without verification or required approval, producing five fabricated citations in documents submitted to the court. Judge Manasco found his conduct “tantamount to bad faith.” The firm faced no sanction. Reeves is still practicing, under conditions. Six months earlier, in February 2026, Judge Jed Rakoff of the Southern District of New York ruled that 31 AI-generated documents were not protected by attorney-client privilege. The defendant’s attorney had used an AI platform to draft communications that were later turned over to the government. The platform’s terms of service, which permitted the vendor to collect and retain user inputs, destroyed the confidentiality element. The attorney’s error. The platform’s design. The attorney’s consequence. The vendor in both cases was not a party to the proceeding. Under SB 3444 and the subscriber agreements already in place, the vendor will not be a party to the next one. Butler Snow’s situation is the normal case, not the extreme case. The extreme case is mass casualties and a $1 billion threshold. The normal case is a partner who trusted a confident screen of text and signed the filing. SB 3444 addresses the extreme case. It leaves the normal case exactly where it found it: on the attorney’s license. ## Before Thursday Pull the subscriber agreement for every AI tool your firm uses. Find the indemnification clause, the warranty disclaimer, and the damages cap. Know what your firm agreed to before the question becomes a bar complaint rather than a hypothetical. Call your malpractice carrier. Ask whether your AI verification protocols fall within your E&O coverage. Then ask a harder question: if a vendor liability shield passes into law across multiple jurisdictions, does your current coverage account for a world in which vendor recourse is legally foreclosed? That answer is not settled. The time to get it is now. Build the Stanford hallucination rates into your firm’s AI governance standards. One in three legal conclusions produced by Westlaw’s AI-Assisted Research tool contained errors in peer-reviewed testing. One in six in the best-performing tool. Those numbers are not background reading. They are the calibration your verification protocols require. A policy that says “verify AI output” without specifying the error frequency your attorneys are working against has omitted the most important number it contains. SB 3444’s threshold for triggering AI vendor liability is 100 deaths or $1 billion in property damage. Matthew Reeves crossed his threshold with one filing. OpenAI’s spokesperson described the company’s support for the bill as focusing on “what matters most: reducing the risk of serious harm from the most advanced AI systems.” The threshold they built for serious harm is 100 deaths. Reeves is still practicing. But the gap between his threshold and OpenAI’s is not a rounding error. It is a policy. And someone is writing it into law. ## About the Author JD Morris is Co-Founder and COO of LexAxiom, an Agentic AI platform for the business of law. Over a 25-year career, he has built and scaled enterprise technology products across Dell, EMC, VMware, and Cisco, including the first exabyte eDiscovery platform. He holds dual MBAs from Columbia Business School (Finance) and UC Berkeley Haas (Marketing), a Master of Legal Studies in Cybersecurity Law from Texas A&M, and a Master of Engineering from George Washington University. He writes The Technology Blind Spot on the intersection of emerging technology and law. Connect with him on LinkedIn atwww.linkedin.com/in/jdavidmorris, on X at @JDMorris_LTech, or on Bluesky at @JDMorris-ltech.bsky.social. ## References 1. Artificial Intelligence Safety Act, S.B. 3444, 104th Gen. Assemb. (Ill. 2026) (proposed). 2. Model Rules of Pro. Conduct r. 3.3 (Am. Bar Ass’n 2024). 3. Model Rules of Pro. Conduct r. 1.1 cmt. 8 (Am. Bar Ass’n 2024). 4. ABA Comm. on Ethics & Pro. Resp., Formal Op. 512 (July 29, 2024). 5. Maxwell Zeff,OpenAI Backs Bill That Would Limit Liability for AI-Enabled Mass Deaths or Financial Disasters, Wired (Apr. 10, 2026),https://www.wired.com/story/openai-backs-bill-exempt-ai-firms-model-harm-lawsuits/. 6.OpenAI Backs Illinois Bill Limiting AI Liability to ‘Critical Harms’, Benzinga (Apr. 10, 2026),https://cdn1.benzinga.com/markets/private-markets/26/04/51763445/openai-backs-illinois-bill-limiting-ai-liability-to-critical-harms. 7.OpenAI Lobbies for Liability Shield as AI Accountability Hits Legislative Phase, Meridiem (Apr. 10, 2026),https://themeridiem.com/tech-policy-regulation/2026/04/10/openai-lobbies-for-liability-shield-as-ai-accountability-hits-legislative-phase. 8.Johnson v. Dunn, No. 2:21-cv-1701-AMM (N.D. Ala. July 23, 2025) (Manasco, J.) (disqualifying Butler Snow attorneys Matthew B. Reeves and William J. Cranford; bar referral; five fabricated ChatGPT citations). 9.ByoPlanet Int’l, LLC v. Johansson, 792 F. Supp. 3d 1341 (S.D. Fla. July 17, 2025) (Leibowitz, J.) ($85,567.75 sanctions; fabricated citations across eight matters). 10.United States v. Heppner, No. 1:25-cr-00503-JSR, ECF No. 27 (S.D.N.Y. Feb. 17, 2026) (Rakoff, J.) (31 AI-generated documents not protected by attorney-client privilege). 11. Hillel Aron,Meta and Google Hit with $6 Million Verdict for Social Media Harms to Young Woman, Courthouse News Serv. (Mar. 25, 2026),https://www.courthousenews.com/meta-and-google-hit-with-6-million-verdict-for-social-media-harms-to-young-woman/. 12. Dawn Chmielewski,Meta, Google Lose US Case over Social Media Harm to Kids, Reuters (Mar. 25, 2026),https://www.reuters.com/legal/litigation/jury-reaches-verdict-meta-google-trial-social-media-addiction-2026-03-25/. 13. Varun Magesh et al.,Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, 22 J. Empirical Legal Stud. 216 (2025). 14. Kyle Chayka,A.I. Has a Message Problem of Its Own Making, New Yorker (Apr. 15, 2026),https://www.newyorker.com/culture/infinite-scroll/ai-has-a-message-problem-of-its-own-making. 15. IssueOne,AI & Tech Lobbying 2025(nonpartisan organization; figure cited in multiple independently corroborated news sources, Apr. 2026).

Originally published on LinkedIn Newsletter: The Technology Blind Spot