ByteDance caught using OpenAI tech to train its rival LLM

Copying OpenAI's homework has become an open secret for broke AI startups trying to catch up. But when one of the richest tech giants on the planet gets caught doing it, the game changes completely.

A detailed report from The Verge revealed that ByteDance has been secretly relying on OpenAI’s API to build and evaluate its own competing foundation model. The practice directly violates OpenAI’s terms of service, which clearly forbid customers from using model outputs to train competing products.

Hours after the story broke, OpenAI confirmed it has officially suspended ByteDance’s account pending further investigation.

Hitting API limits and washing data

According to internal documents and Lark employee chats, ByteDance's secretive foundation model initiative, codenamed Project Seed, relied heavily on OpenAI’s technology across nearly every development stage. Led by search chief Zhu Wenjia and closely tracked by co-founder Zhang Yiming, the team was under strict orders to match GPT-3.5 by the end of this year and hit GPT-4 parity by mid-2024.

The project team was burning through so much access that they regularly hit their maximum API allowance. Inside their internal chats, engineers openly discussed how to desensitize and whitewash the data to hide the evidence.

ByteDance did issue an order a few months ago telling the team to stop using GPT data in model development, right around when it secured Chinese regulatory approval for its Doubao chatbot. But sources familiar with the project say engineers continued using the API to benchmark and evaluate the 200-billion-parameter model powering Doubao.

A ByteDance spokesperson claimed the GPT data was only used for annotation early on and was completely purged from training sets by mid-year, adding that the company is properly licensed through Microsoft to use GPT in international products.

Why it matters

For a scrappy startup, distilling proprietary models is a calculated legal gray area. For a tech empire the size of ByteDance, doing it is downright embarrassing. As industry researchers noted, this level of corner-cutting points to sheer panic and unrealistic internal deadlines.

We think this incident cuts through the marketing hype around proprietary foundation models. Building frontier AI from scratch is brutally hard, and even companies with near-infinite resources are tempted to cheat when pushed to match OpenAI on an impossible timetable. Getting banned publicly is a massive reputational blow that will definitely hurt ByteDance’s US recruitment efforts.

Beyond corporate embarrassment, this practice threatens the entire AI data ecosystem. When giant companies secretly use synthetic outputs from GPT to train next-generation models, hallucinations compound and the web gets flooded with model incest. By pulling the plug on ByteDance, OpenAI is sending an unmistakable warning to enterprise customers: using their intelligence to bootstrap a direct rival will get you cut off immediately.