OpenAI bans using its models to train rival AI systems

What OpenAI changed

OpenAI has updated its usage policies, which set out how people may access and build with its models. As BankInfoSecurity reported, the changes take effect on Nov. 20. The biggest addition is an explicit ban on using OpenAI's models, or what they produce, to help create competing models.

The policy covers generating, preparing, selecting or supplying training material or signals for models meant as alternatives to or substitutes for OpenAI's own. That includes building training data, writing the scoring guides used in reinforcement learning (a method where a model is trained by rewarding good answers), using the models to give feedback, and filtering data.

The ban is broad about who is responsible. It applies whether you do the work yourself, for someone else, or through a person acting for you. It also applies if you never train the final model yourself, and if outputs gathered during ordinary product use or testing are later handed over for training.

Distillation and why it is contested

Distillation is a technique in which a new model learns from a larger "teacher" model instead of from a huge dataset of its own. Used openly, it is a normal way to make smaller, cheaper models. The dispute is about doing it to a rival's model without permission.

The practice has become a political issue. Both OpenAI and Anthropic have asked the U.S. government for help stopping competitors, mostly Chinese AI labs, from extracting the proprietary reasoning of their models. The U.S. Cybersecurity and Infrastructure Security Agency has warned about DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.ai, which are alleged to have illicitly distilled OpenAI and Anthropic models.

Not everyone sees it as a threat. Nvidia CEO Jensen Huang and Meta CEO Mark Zuckerberg have defended distillation, arguing it is how American open-weight models (models whose trained parameters are published) will stay competitive.

OpenAI also gives a safety argument. It says that training another model on its outputs can carry over capabilities without the safeguards. In its words: "training another model on OpenAI outputs can transfer capabilities without carrying those safeguards over."

Routers, gateways and government work

The update goes beyond distillation. Model routers and gateways are services that let companies switch quickly between AI models to power their agents or particular tasks. They have grown popular, and OpenAI has now clarified how outside providers may offer access to its models.

Anyone running a platform that lets third parties directly reach OpenAI models must get OpenAI's review and approval first. The source does not say how that review works or how long it takes.

The policy also covers security, intelligence and law enforcement. OpenAI says it will approve "specific partners" case by case to use its models for some tasks that are otherwise prohibited. It will weigh factors such as the proposed use and the safeguards in place.

Every customer must still preserve human judgment and accountability over high-stakes decisions, including the use of force. The company also points to its principles against mass domestic surveillance and against evading legal obligations, oversight or accountability. BankInfoSecurity noted that this follows Anthropic's own policy edits, which ban deceptive uses and clarify its prohibitions on weapons development.

What this means if you build on the API

Most people who use ChatGPT for documents, code or analysis are unlikely to be affected. OpenAI says research, software development and model evaluation can continue within its boundaries. The risk falls on teams that take model outputs and feed them into the training or tuning of another model.

The wording of the rule means the paper trail matters. Because it applies to work done through contractors and to outputs gathered earlier during product use, a team could end up in a gray area without meaning to. Keeping track of which model produced which dataset makes those questions easier to answer.

Frontier labs update their usage policies regularly to reflect new capabilities, user demand and company values. This change is another sign that the terms of service are becoming a main tool for protecting a model, alongside technical defenses.