The Center for Artificial Intelligence and Digital Policy asked the US Federal Trade Commission to investigate OpenAI and stop further commercial deployment of GPT-4. The complaint attempted to connect emerging AI risks with an existing enforcement mechanism: the prohibition on unfair or deceptive business practices.
CAIDP described GPT-4 as biased, deceptive and a risk to privacy and public safety. Its filing argued that the model had reached the market without an independent assessment of its risks and called for guardrails that could protect consumers, businesses and the commercial marketplace.
The Complaint Used OpenAI’s Own Risk Disclosures
CAIDP cited the GPT-4 System Card, in which OpenAI acknowledged that the model could reproduce biases, harmful stereotypes and demeaning associations. The document also discussed broader risks from systems that can reinforce worldviews, including false or harmful ones.
Other identified outcomes included content encouraging self-harm, graphic material, harassment, hateful responses, assistance with violence and directions toward illegal content. Listing these possibilities did not establish that every user would encounter them, but it showed that foreseeable harms were known when the product was released.
The complaint’s logic was that disclosure alone was not enough. If a company knows a system can produce harmful results, regulators can ask whether reasonable protections were built into the product and whether marketing accurately described its limitations. Transparency is a starting point, not a substitute for mitigation.
The FTC Emphasised Durable Safeguards
The source cited FTC guidance warning companies that telling users about misuse or requesting disclosures would not adequately deter bad actors. The agency said protections should be durable, built-in features rather than optional additions that third parties could remove or bypass.
That standard matters for general-purpose models because the provider does not control every downstream application. Documentation and terms of service may guide responsible developers, but they cannot prevent a poorly designed interface from exposing users to unsafe behaviour.
Durability requires several layers: model evaluation, access controls, monitoring and product-level restrictions. It also requires testing whether safeguards survive ordinary attempts to work around them. A protection that functions only under cooperative prompts offers little assurance in a commercial system.
Independent Assessment Became the Core Demand
CAIDP asked the FTC to examine OpenAI and other operators of powerful AI systems, prevent further GPT-4 releases and require safeguards before commercial expansion. The request reflected concern that companies developing a model should not be the only parties judging whether it was ready.
Independent review can challenge internal incentives and test claims against common standards. It also raises practical questions about access to models, confidential information and evaluator expertise. A credible process needs enough visibility to identify weaknesses without requiring a company to publish security-sensitive details.
The timing of the complaint added urgency. Technology leaders had separately called for a pause on systems more powerful than GPT-4, while other figures argued that progress should continue. Enforcement offered a different path from a voluntary industry-wide pause: evaluate specific conduct under existing consumer-protection authority.
Global Regulation Was Already Diverging
The complaint arrived as Italy blocked ChatGPT over privacy concerns and European institutions developed a stricter regulatory environment. UNESCO encouraged countries to apply its ethics framework, while the UK favoured a lighter, sector-based approach intended to support innovation.
Different models of oversight reflected the same balancing problem. Too little regulation can expose the public to avoidable harm, while poorly designed requirements can impose costs without improving safety. A sector-based system may use specialised expertise, but a fragmented approach can leave gaps when one model crosses many industries.
Existing agencies such as the FTC bring enforcement experience and legal authority. They can address deceptive claims and consumer harm without waiting for a complete new AI law. However, general rules still need technical interpretation so companies know what evidence, testing and controls are expected.
The CAIDP filing was significant even before any regulatory outcome because it reframed AI safety as an enforceable product issue. Questions about bias, deception and public risk moved from research discussions into the language of market conduct and consumer protection.
For developers, the practical message is that self-reported limitations must lead to action. A system card can demonstrate candour, but it also creates a record of known hazards. Teams should connect each material risk to a mitigation, an evaluation and a decision about acceptable deployment.
For regulators, the challenge is to require evidence without freezing standards around one model generation. Outcome-focused rules can ask whether protections are effective, independently tested and maintained as products change. That approach gives enforcement room to evolve alongside capability.
The complaint captured an early attempt to define accountability for powerful general-purpose AI. It did not resolve the dispute between speed and caution, but it set a clear expectation: commercial deployment carries responsibilities that cannot be met by innovation claims or warning labels alone.
Remedies also need proportionality. A complete stop can prevent harm, but narrower restrictions, staged access or required testing may address a specific failure while preserving legitimate use. Regulators should be able to escalate when companies ignore evidence, yet their decisions should identify the conduct being corrected and the conditions for returning to market. That clarity helps consumers, developers and competitors understand the standard. It also turns enforcement from a one-off confrontation into a repeatable signal about what responsible commercial AI requires.
