Paste your API key into a free ChatGPT account and start asking it about your production incident. Describe the architecture, the error logs, the customer impact. You just handed OpenAI a detailed map of your system, and unless you dug into Settings > Data Controls and flipped the right toggle, that conversation may inform future model training. Most people never flip it.
The actual policy breakdown is less dramatic than privacy advocates claim and more consequential than the AI companies advertise. OpenAI and Anthropic both guarantee zero training on inputs for enterprise and business tiers. That protection is contractual, not aspirational. When OpenAI launched its Health feature in January 2026, they were explicit: health conversations, memory, and connected files "are not used to train our foundation models." That's a real engineering commitment, not a press release flourish. The problem is that commitment applies to a specific product tier and a specific feature, not to the 987 million regular users on free and consumer plans.
The Opt-Out Default Is a Policy Choice, Not a Technical Constraint
OpenAI could default everyone to opt-out. They don't. That's a decision, and it benefits them. Consumer conversations at scale are genuinely valuable training signal. I understand the engineering argument: diverse real-world prompts improve models in ways synthetic data can't replicate. Fair point. But "we need your data to build better products" is not a consent framework. It's a rationalization for a default that serves the company.
Meta AI is worse. Meta confirmed it uses chat interactions to personalize ads. If you're using Meta AI to draft anything sensitive, you're not using a productivity tool; you're using an ad-targeting pipeline with a chat interface bolted on. That's a documented behavior, not speculation.
DeepSeek sits at the far end of this spectrum. Multiple security researchers have flagged data disclosure risks serious enough that the guidance is simply: don't use it for sensitive data. Full stop. The $11.8 billion chatbot market in 2026 includes a lot of products with very different threat models, and "AI chatbot" as a category obscures that spread entirely.
What Builders Should Actually Do
If you're shipping anything that touches user data, legal strategy, financial models, or health information, the free tier of any major chatbot is the wrong tool. ChatGPT Team or Enterprise, Claude for Work, these tiers exist precisely because the enterprise contracts prohibit training on your inputs. The cost difference is real but so is the liability difference.
For personal use, the fix is tedious but takes 90 seconds. In ChatGPT: Settings > Data Controls > toggle off "Improve the model for everyone." Anthropic's Claude doesn't train on conversations by default for paid users; check your plan. For Meta AI, the answer is simpler: don't use it for anything you wouldn't say in a Facebook post, because the data pipeline is essentially the same.
The deeper issue is that this tier system shouldn't require a settings archaeology expedition to understand. Regulators in the EU are already pushing toward mandatory opt-in for training data collection. The U.S. has no equivalent requirement, which means the default stays wherever it benefits the company. Until that changes, the burden falls on users to read the fine print, find the toggle, and understand that "free" in this context means something specific about whose interests the product serves.
You opted in the moment you hit Enter. The question is whether you knew that.
