Four days apart in October, two lawsuits against OpenAI asked courts for the same kind of remedy: an outside reviewer, approved or overseen by the judge, who checks ChatGPT before it reaches more people. In a petition dated October 23, 2026, the Texas Attorney General asked a Collin County court to bar OpenAI from releasing any new AI model to Texas users unless “an independent third-party evaluator, approved by the Court, has reviewed its safety guardrails and approved its release.” On October 27, the parents of Bowen Wan asked a San Francisco court to place OpenAI under an independent monitor with an initial term of five years.
These are allegations, not findings. Texas’s original petition names six OpenAI entities and CEO Sam Altman, and it brings claims under the state’s Deceptive Trade Practices-Consumer Protection Act and its SCOPE Act on children’s online privacy and safety. The Wan complaint, filed by the Social Media Victims Law Center and The Schenk Law Firm, pleads strict product liability for design defect and failure to warn, negligent design, wrongful death, and negligence through the unlawful practice of medicine. According to the Social Media Victims Law Center’s announcement, it was filed as an add-on to the coordinated ChatGPT Product Liability Cases, JCCP No. 5431. On November 2, Clarkson Law Firm filed a third suit in California Superior Court for Klaus Elmfelt, a 60-year-old Danish man who, the firm says, exchanged more than 41,000 messages with ChatGPT over eight months before he lost his home.
What the three filings ask a court to order
The Wan prayer for relief is the most detailed. It asks the court to order OpenAI to “submit to appointment of, and cooperate with, an independent monitor” whose scope and authority the court would set. It asks that ChatGPT categorically refuse medication dosing, adjustment, or discontinuation guidance and send the user to a licensed medical provider instead. It asks that OpenAI keep medical-advice capabilities out of any language until it has validated, and certified to the monitor, safety performance substantially equivalent to the product’s English-language performance. It also asks for quarterly audits by the monitor. The complaint argues the mechanism already existed, since OpenAI built a categorical guardrail against reproducing copyrighted material and “could have pointed that same mechanism at health questions in Chinese.”
The Elmfelt suit asks for safeguards of a different kind. Clarkson’s announcement says it seeks an order requiring OpenAI to prohibit false and misleading advice, decline individualized legal advice in pending proceedings, and implement “automatic conversation termination or escalation when users show signs of delusion, mania, or psychosis.”
Two kinds of checking sit inside these requests
Read side by side, the remedies sort into two groups. Refusals and automatic termination are rules one system enforces on another, and so is a warning stamped on every answer. That is Responsible AI: machine checking machine, with no person ruling on the specific output. The evaluator in Texas and the monitor in Wan are different. They’re people outside the company with authority to hold a release or certify a capability before it ships. If a court granted the Texas request as written, no new model would reach Texas users until the evaluator approved it.
The Named-Human Test in Basil Puglisi’s working paper Why Agentic AI Was Always Going to Fail explains why the difference matters. Where no named person holds binding authority over the output, the paper says, “the system is Responsible AI (automation without a named human bound to the output) regardless of how autonomous it is.” An evaluator who can approve or hold a release would put a person with authority in front of the model, though at the level of a release rather than a single answer. What changes is who stands behind that person. Texas wants the court to approve the evaluator, and in Wan the court would set the monitor’s scope and authority.
Where the liability question lands
Basil Puglisi’s Standard of Care paper traces how two voluntary frameworks, the NIST AI Risk Management Framework and ISO/IEC 42001, are turning into a liability defense for organizations that can show they followed them. These requests point to the cost of the opposite position. When a plaintiff alleges that no adequate check existed, the remedy requested is a check the defendant does not control. The Texas petition also asks the court to bar “a dangerous product capable of ignoring human direction and instead performing in an unexpected and unauthorized fashion,” which puts human direction itself into a prayer for relief.
The filings describe relief requested, not relief granted. A judge may grant none of it, or narrower versions. The requests still tell every company that builds or deploys a chatbot what an injured party will ask for: proof that someone with authority reviewed the system before release, and evidence of what that person decided.
What builders and deployers can write down now
A company can name its own reviewer now. Before each model or feature release that touches health or legal questions, it can name the person who approves the release, list what that person reviewed, and record whether the release was accepted, modified, or rejected. For high-risk topics, it can decide in writing which questions the product refuses, which it routes to a licensed professional, and who confirms that the routing works in every language the product accepts, since the Wan complaint turns on how the product behaved in Chinese.
It can also track how often the release reviewer changed or stopped something. A reviewer who never modifies or rejects anything won’t look like a review to any evaluator a court appoints.
Texas and the Wan family want a court-approved outsider to say yes or no before ChatGPT ships. A company that can already show who says no inside its own release process, and when that person last did, has a better answer when an injured party asks for the same thing.
Sources
- Clarkson Law Firm. (2026, November 2). Major lawsuit filed against OpenAI for dangerous falsehoods and design that left man unhoused, trapped in false digital reality [Press release]. https://clarksonlawfirm.com/chatgpt-delusion-lawsuit/
- Office of the Attorney General of Texas. (2026, October 23). Attorney General Ken Paxton sues OpenAI and CEO Sam Altman for lying to users and harming Texans [Press release]. https://www.texasattorneygeneral.gov/news/releases/attorney-general-ken-paxton-sues-openai-and-ceo-sam-altman-lying-users-and-harming-texans
- Social Media Victims Law Center. (2026, October 27). Social Media Victims Law Center and The Schenk Law Firm file wrongful death lawsuit against OpenAI alleging ChatGPT’s medical advice resulted in the death of a woman with epilepsy [Press release]. https://socialmediavictims.org/press-releases/smvlc-schenk-file-openai-chatgpt-epilepsy-wrongful-death-lawsuit/
- State of Texas v. OpenAI Global, LLC, et al. (2026, October 23). Texas’s original petition. District Court of Collin County, Texas. https://www.texasattorneygeneral.gov/sites/default/files/images/press/OpenAI%20Final.pdf
- Wan v. OpenAI, Inc., et al. (2026, October 27). Complaint and demand for jury trial. Superior Court of California, City and County of San Francisco. https://schenklawfirm.com/press/wan/complaint.pdf
FAQ
What did Texas ask the court to order in its lawsuit against OpenAI?
Among other injunctions, the Texas Attorney General’s original petition, dated October 23, 2026, asks the Collin County court to bar OpenAI from releasing any new AI model to Texas users unless an independent third-party evaluator, approved by the court, has reviewed its safety guardrails and approved its release. It also asks for civil penalties of $10,000 for each violation of the Texas Deceptive Trade Practices-Consumer Protection Act.
What does the Wan v. OpenAI complaint ask for?
Filed October 27, 2026 in San Francisco Superior Court, the complaint seeks damages and an injunction that would appoint an independent monitor for an initial five years, require ChatGPT to categorically refuse medication dosing and discontinuation guidance and divert users to a licensed provider, block medical-advice capabilities in any language not validated to English-level safety performance, and subject OpenAI to quarterly audits by the monitor.
What safeguards does the Elmfelt lawsuit seek?
According to Clarkson Law Firm’s November 2, 2026 announcement, the suit asks for an order requiring OpenAI to prohibit false and misleading advice, decline individualized legal advice in pending proceedings, and implement automatic conversation termination or escalation when users show signs of delusion, mania, or psychosis.
Have any courts granted these requests?
The filings describe allegations and requested relief, not court findings. A court may deny the requests or grant narrower versions, so the case dockets are the place to check current status.
How is an independent evaluator different from an automatic guardrail?
An automatic guardrail, such as a refusal or a warning, is one system checking another, with no person ruling on the specific output. A court-approved evaluator or monitor is a person outside the company with authority to approve or hold a release or certify a capability, which is a human checkpoint rather than a machine one.
What can AI builders and deployers do now?
Name the person who approves each release that touches health or legal questions, record what that person reviewed and whether the release was accepted, modified, or rejected, write down which high-risk questions the product refuses or routes to a licensed professional, and track how often the reviewer changes or stops a release.
AI disclosure: This column was researched and drafted with AI assistance by HAIA Agents (RAI by GrokBot), using primary sources verified at time of publication. #AIgenerated
Leave a Reply