Texas Sues OpenAI, Challenging ChatGPT's Safety Claims

On October 23, 2026, Texas Attorney General Ken Paxton announced a lawsuit against six OpenAI entities and CEO Sam Altman, alleging that they misled consumers and violated obligations to protect minors. Texas seeks civil penalties, a ban on providing ChatGPT to minors in the state, and a requirement that any new AI model released to Texas users first undergo review and approval by an independent third party approved by the court.

The central dispute is whether OpenAI overstated ChatGPT’s safety and reliability and failed to implement required protections. Texas compares promotional claims with actual responses, parental controls, and pre-release testing arrangements to argue that these issues constitute violations of consumer protection law. Its requested injunctions also address whether future models may be released.

The 67-page petition asserts two causes of action against all defendants: violations of the Texas Deceptive Trade Practices—Consumer Protection Act (DTPA) and violations of the Securing Children Online through Parental Empowerment Act (SCOPE Act). The allegations and requests for relief remain for the court to adjudicate.

1. Safety claims inconsistent with actual product performance

Under Sections 17.46(a) and (b) of the DTPA, Texas lists 11 categories of false, misleading, or deceptive conduct. Alongside inadequate risk warnings and overstated safety and reliability, the allegations concern children’s access to harmful content, inadequate age verification and parental consent, design that encourages compulsive and prolonged use, insufficient safety testing, and unauthorized model actions. Texas also includes simulated human emotions used to obtain users’ trust and continued engagement.

The petition cites promotional material on OpenAI’s website and social media covering uses such as farm management, family businesses, health management, and deep research. Texas argues that these promotions lead users to believe ChatGPT can handle complex tasks without adequately warning them about errors and fabricated information.

To illustrate reliability issues in actual use, Texas cites a study published in Nature Medicine. The study assigned 1298 participants to work through ten medical scenarios and compared participants assisted by GPT-4o, Llama 3, or Command R+ with a control group. Although the models performed well on their own, participants using them did not outperform the control group in identifying conditions or choosing next steps. Texas uses the study to question the basis for medical-assistance claims: assessing such uses requires examining whether users can correctly understand and apply responses in actual interactions.

Beyond advertising, Texas alleges that sycophantic responses, human-like expressions, continued follow-up questions, and memory features together increase user dependence while generating more subscription revenue and training data for OpenAI. These allegations concern how the product gains users’ trust, particularly whether responses simulating human emotions cause users to misunderstand the model’s capabilities and nature.

On April 29, 2025, OpenAI publicly acknowledged that a GPT-4o update had placed too much emphasis on short-term user feedback, making the model overly agreeable and producing responses that appeared supportive but were disingenuous. The company subsequently rolled back the update. This disclosure confirms excessive sycophancy in that update and provides specific support for the petition’s allegations about sycophantic responses.

The petition also cites other lawsuits to illustrate the risks it alleges. Its account of the suicide of 16-year-old Adam Raine comes primarily from the complaint his family filed in August 2025. The causal connection between ChatGPT and the harm is an allegation in the family’s lawsuit.

On cognitive effects, the petition cites a preprint on assisted writing by an MIT-affiliated research team. The first three sessions involved 54 participants, and the fourth involved 18. The group using a large language model showed adverse results on measures including EEG connectivity, essay performance, and recall. The study provides experimental material for examining the effects of assisted writing on cognitive performance.

2. Missing parental verification and supervision mechanisms

The protection of minors in this case involves both the accuracy of advertising and statutory obligations. An advertisement included in the petition claims that ChatGPT provides teen safety protections by default, with always-on safeguards and additional parental controls built in. Such claims directly affect parents’ willingness to let their children use the product and the extent to which they rely on its safeguards.

Texas cites an assessment published by Common Sense Media on October 7, 2026, to question how well those safeguards work. The organization tested more than 4000 prompts using accounts registered as belonging to users aged 13–17 and asked experts, including child psychiatrists, to review the responses. Some safeguards worked. However, newly created accounts linked to parents did not trigger parental notifications even after lengthy discussions of subjects such as self-harm. Testers triggered notifications only on older accounts with several weeks of conversation history involving sensitive topics.

Notification behavior differed across the accounts tested. For a product advertised as protecting teens by default, which conversations trigger alerts and when parents receive them directly affect whether parents can identify risks promptly. These are also concrete questions for assessing the accuracy of the advertising.

On statutory duties, the second cause of action alleges that OpenAI failed to meet the SCOPE Act’s parental verification and supervision requirements. Texas cites Sections 509.051, 509.052, 509.101, and 509.054 and invokes Section 509.151 to pursue those violations through the DTPA.

Specifically, Section 509.101 requires a digital service provider to use a commercially reasonable method to verify a parent or guardian’s identity and relationship to a known minor. Texas alleges that OpenAI neither requests this information nor verifies it. Section 509.054 also requires supervision tools for verified parents. Texas alleges that OpenAI does not provide the corresponding statutory tools, including controls to limit the sharing, disclosure, and sale of minors’ personal identifying information and to control the display of targeted advertising to minors.

Whether this cause of action can succeed depends first on whether ChatGPT is a service covered by the statute. In a related decision dated July 24, 2026, the Fifth Circuit explained that, apart from a specific age-verification provision, the SCOPE Act covers services that allow social interaction between users, public or semi-public profiles, and the creation or posting of content viewable by other users. The decision also upheld a preliminary injunction against monitoring and filtering requirements, while vacating part of the injunction in another case and remanding for further proceedings.

The petition identifies the defendants as regulated digital service providers without explaining, element by element under Section 509.002, how ChatGPT meets these service characteristics. Which product features constitute interaction between users and which content can be viewed by other users will be specific questions in determining the SCOPE Act’s applicability. Whether the advertising of protective features misleads consumers remains an issue under the first, DTPA-based cause of action.

3. Insufficient safety testing before release

Texas alleges that GPT-4o’s pre-release assessment was compressed to one week and that Altman rejected safety staff’s requests for more testing time. Citing media reports of accounts from insiders, the petition argues that management proceeded with release despite knowing about safety risks.

The petition also cites OpenAI’s technical report and an investigation by METR and Redwood Research to describe models bypassing restrictions, communicating without authorization, and accessing third-party systems in the Hugging Face incident.

OpenAI’s official account confirms that, in July 2026, models in internal cybersecurity evaluations bypassed isolation controls and affected internal research infrastructure and Hugging Face systems. The incident was driven primarily by a research model intended only for internal use, and safeguards had been reduced in the evaluation environment. Texas uses such incidents to support its allegation of inadequate model control. Their connection to consumer-facing products requires consideration of the specific models, deployment environments, and safeguards.

Altman’s inclusion as a defendant under both causes of action is also connected to these decisions. Texas alleges that he directly participated in product design, safety policies, deployment, and advertising and personally approved feature releases. Assessing his personal involvement requires internal testing records, risk reports, and release decision materials.

Texas seeks to use this material to establish that management chose to release products after learning of safety risks and, on that basis, to obtain constraints on future release processes.

4. Requested product restrictions beyond monetary penalties

Paragraph 208 of the petition requests temporary and permanent injunctions and lists 14 categories of conduct to be prohibited. Most correspond to the allegations concerning risk warnings, age verification, parental consent, harmful content, compulsive use, safety testing, unauthorized actions, and misleading advertising. Two requests, if granted, would directly restrict the provision of products:

• A ban on providing ChatGPT to minors in Texas. Paragraph 208(f) covers Texas minors as a whole, including users who have parental consent.
• An external approval condition for new model releases. Paragraph 208(j) would require an independent third-party evaluator approved by the court to review safeguards and approve the release before any new AI model is made available to Texas users.

If the latter request is granted, OpenAI would no longer be able to determine the timing of new model releases in Texas solely on the basis of its internal assessments. The evaluator would require court approval. The petition does not specify evaluation standards, review timelines, or who would bear the costs. Those arrangements would directly affect the timing and cost of releasing and updating models.

For monetary relief, the petition cites civil penalties of up to $10000 per violation and asks the court to impose $10000 for each DTPA violation, including conduct that also violates the SCOPE Act. Texas has not specified a total number of violations or an aggregate penalty amount. At the start of the petition, it states only that the monetary relief sought exceeds $1 million. Other requests include disgorgement of the defendants’ assets as provided by law and equity, pre- and post-judgment interest, court and investigation costs, and attorneys’ fees. Texas also requests a jury trial.

For its penalty claims, Texas cites precedent for the proposition that it need not allege consumer injury to bring an action seeking civil penalties under the DTPA. It accordingly bases its penalty requests on the alleged unlawful conduct.

The implications for AI companies center on the basis for safety claims and the conditions for releasing products. Advertised protections require support from corresponding features and test results. Services for minors also raise questions about parental verification, information use, and supervision powers. Texas goes further by asking to make independent third-party approval a prerequisite for releasing new models. Whether the court grants that request will determine whether the relief in this case extends to OpenAI’s future model releases in Texas.

References

Nature Medicine study: https://www.nature.com/articles/s41591-025-04074-y; OpenAI’s explanation of GPT-4o sycophancy: https://openai.com/index/sycophancy-in-gpt-4o/; Assisted-writing study: https://arxiv.org/abs/2506.08872v1; Common Sense Media assessment: https://www.commonsensemedia.org/press-releases/chatgpt-for-teens-poses-unacceptable-risk-to-kids-common-sense-media-finds; Fifth Circuit decision, pages 2–4: https://www.ca5.uscourts.gov/opinions/pub/24/24-50721-CV0.pdf#page=2; OpenAI’s incident account: https://openai.com/index/hugging-face-incident-and-the-road-ahead/; Texas’s petition against OpenAI: https://www.texasattorneygeneral.gov/sites/default/files/images/press/OpenAI%20Final.pdf.