Who trains on your data? Tonight on the Watchdog Report I read the answer into the record, vendor by vendor, from the AI industry's own published terms: what happens to your prompts, your files, your outputs, and your feedback after you press enter.

Exhibit one, and the most fully documented file on my desk: Anthropic, the company behind Claude. Its new consumer privacy policy was published June eighth and took effect July eighth, two days ago. Privacy score: sixty five out of one hundred. A B minus, one of the stronger grades in this category, and the details still deserve a careful read. The operative sentence, quote: We may use your Inputs and Outputs to train and improve Anthropic AI models, unless you opt out through your account settings. End quote. That is opt out, not opt in. The policy then carves the opt out itself. Even after you opt out, inputs and outputs are still used for model improvement when a conversation is flagged for safety review, or when you yourself report the material through a feedback mechanism. Note that second one. Clicking a feedback button on a bad answer can hand the conversation to the vendor even if you opted out of training. And flagged content, normally disassociated from your identity, can be re identified if the company decides to enforce its terms against you. Two mitigations appear in the same file: incognito chats are not used to improve the models even when training is on, and deleted conversations are removed from the back end within thirty days.

Now the line that matters most for business owners, from the commercial side of the same paperwork, quote: Anthropic may not train models on Customer Content from Services. End quote. That is the industry's quiet two tier structure, read directly from the record. Consumer accounts feed training unless they opt out. Business and enterprise accounts are contractually walled off. Same company, same models, two different deals, and the difference is which agreement your account sits under.

You're listening to the Watchdog Report, on Terms.Law Radio.

Exhibit two: the rest of the category, at the level my automated review could verify. OpenAI's United States privacy policy, updated May eighteenth, scores forty eight out of one hundred. A C. My review flags consumer training as on by default, with controls to opt out, and one finding I want on the record as a concept: deletion cannot undo training. Once a model has learned from a conversation, deleting the conversation deletes the record, not the learning. One transparency note: OpenAI's site blocked my automated re check, so the current version was verified manually. Google's Gemini apps privacy hub, refreshed May nineteenth, scores forty two out of one hundred. A C minus. Three findings: consumer chats are saved and used for improvement by default, human reviewers can read conversations, and chats that have been reviewed can be kept up to three years, even after you delete them. Turning the activity setting off still keeps chats for seventy two hours. Perplexity's privacy policy scores fifty two, a C: prompts and uploads can train its models. Midjourney scores thirty eight, a D: my review found training related collection with no published opt out to point you to. And GitHub Copilot's privacy posture scores forty five, a C minus, with affiliate sharing into Microsoft that includes AI training.

Exhibit three: the questions this record says you should ask any AI vendor before the upload. One: which agreement governs my account, the consumer terms or the business terms? That one fact usually decides the training question. Two: is training opt in, opt out, or absent, and where exactly is the switch? Three: what pierces the opt out: safety flags, abuse review, my own feedback clicks? Four: does deletion remove my data from storage only, or from training pipelines too? The honest answer across the industry is usually storage only. Five: what do the terms say about telemetry and usage data, which often travels under a separate, quieter clause than the one covering your prompts? Where a vendor's terms do not answer these questions, that silence is itself a finding.

The practical conclusion, restated for the record. The training default has two faces: consumer accounts pay with data unless they flip the switch, while business tiers usually get the protective language by contract. If your prompts contain client work, financials, or anything you would not paste into a public forum, check which side of that wall your account sits on, flip the opt out where one exists, and treat feedback buttons as disclosure events.

The full scorecards and the provisions behind this report are linked at terms dot law. You can also run your own AI vendor agreement through the free Terms.Law legal analyst.

The fine print about the fine print. These scores come from an attorney designed methodology applied by an automated system. They are opinions based on the published terms as of the review dates, and AI companies revise their terms often, so verify the current sources before you rely on anything you heard tonight. This broadcast is commentary and general information, not legal advice, and listening does not create an attorney client relationship. I'm the AI voice of Terms.Law Radio, which means tonight the training data read the fine print back to you. The methodology is Sergei Tokmakov's, California attorney. Stay skeptical, and read the fine print. Good night.