AI Training Data Ethics: How to Vet Any AI Tool

A practical playbook for vetting AI tools on training data provenance, retention, and transparency before you commit. Includes our vendor picks.

Anthropic says it caught someone using millions of Claude exchanges to train competing models, and the story points to labs in China. Read that sentence again and sit with it for a second. The tool you are probably using to draft emails, debug a function, or summarize a PDF is also, in aggregate, a training corpus. Every prompt you type is a data point. The question is who gets to learn from it, and whether the vendor you pay is honest about that. This is not a China story. It is a vendor-vetting story. If a competitor can quietly hoover up millions of exchanges from a commercial AI product, then your own data hygiene depends entirely on how transparent your vendor is willing to be. Here is the playbook we use before we let any AI tool near our work. ## Why the Claude story matters beyond Anthropic The mechanics are mundane. A model provider logs prompts and completions. Someone with access, or someone who found a gap in access controls, extracts a large volume of that traffic. That traffic gets used as training data for a rival model. No malware, no dramatic breach. Just a pipeline. What makes it newsworthy is the scale and the intent. Millions of exchanges is not a rounding error. It is a deliberate harvest. And the target was a frontier model, which means the harvested data was high quality by definition. You cannot get that signal from scraping Reddit. For anyone choosing AI tools, the lesson is blunt. Training data provenance is a security property, not a marketing footnote. If a vendor cannot tell you where its training data came from, you cannot assess your exposure. Full stop. ## The three questions that actually matter Most AI vendors publish a policy page that says something like "we may use your data to improve our services." That sentence is worthless. It tells you nothing about scope, retention, or whether humans review your inputs. We ask three questions instead. They are harder to dodge. 1. What is the default? Is my data used for training unless I opt out, or excluded unless I opt in? 2. What is the retention window? Thirty days, ninety days, indefinitely, or until I delete it? 3. Who can see it? Automated systems only, or human reviewers, or third parties? If a vendor cannot answer all three in plain language on a public page, treat that as an answer too. ### Question one: opt out or opt in Opt-out is the industry default and it is a trap. You have to find a setting, often buried two menus deep, and toggle it before your first prompt. Most people never do. Opt-in flips the burden. Your data stays out unless you explicitly hand it over. We weight this heavily. A vendor that defaults to opt-in is telling you something about how it thinks about users. ### Question two: retention Retention windows are where vendors get slippery. "We retain data as long as necessary" means nothing. A real answer looks like "thirty days for abuse monitoring, then deleted." Vague retention plus training rights is the worst combination. You have no idea how long your prompts live or where they end up. ### Question three: human review This one surprises people. Many vendors use human reviewers to grade model outputs. That means a person can read your prompts. If you paste a client contract, a medical note, or a proprietary codebase into a chat window, a contractor might see it. Ask directly. Vendors that use human review usually disclose it in a subprocessor list or a trust center. Vendors that hide it are the ones to worry about. ## A quick scorecard for any AI tool We run every candidate through the same grid before it touches real work. It takes ten minutes and it has saved us from at least two bad contracts. | Signal | Green | Yellow | Red | |---|---|---|---| | Training default | Opt-in only | Opt-out, easy to find | Opt-out, buried or absent | | Retention policy | Specific window published | Vague but stated | "As long as necessary" | | Human review | Disclosed, with controls | Mentioned in passing | Not mentioned | | Subprocessors | Public list | On request | Refused | | Incident history | Disclosed promptly | Delayed disclosure | Denied or hidden | | Data residency | Region selectable | Stated | Unknown | Two greens and no reds is our floor. One red and we walk, unless the tool is doing something we genuinely cannot replace and we can isolate it from sensitive data. ## Our take: who we trust and why We are not going to pretend every vendor is equally bad. They are not. For general work, we default to Claude and [ChatGPT](https://chat.openai.com/) with training toggled off and retention set to the shortest window the plan allows. Both publish subprocessor lists and both offer zero-retention options on business tiers. Anthropic's handling of this incident, going public with the scale and the target, is the behavior we want from a vendor. Silence would have been worse. For code, we lean on GitHub Copilot with the business plan, which excludes your code from training by default, and we keep anything sensitive in a local model. Ollama running a mid-size open-weight model on a laptop is not frontier quality, but it never leaves the machine. That tradeoff is worth it for client work. For anything touching regulated data, we do not use hosted chat at all. Self-hosted or nothing. The pattern is simple. Pay for the business tier, read the trust center, turn off training, and keep a local fallback for the sensitive stuff. None of this is exotic. It is just work most people skip. ## What to do this week You do not need a security team to act on this. You need an afternoon. - Audit your AI subscriptions. List every tool that sees your prompts. - Check each one's training default. Toggle it off if you can. - Move anything sensitive to a local model or a zero-retention tier. - Write a one-page internal rule for what never goes into a hosted chat. - Re-check quarterly. Policies change quietly. The last item matters most. Vendors update their terms. A tool that was clean last year may have new training rights this year. Put a calendar reminder on it. ## FAQ ### Did Anthropic confirm Chinese labs trained on Claude data? Anthropic has publicly stated that it detected large-scale use of Claude exchanges to train competing models and linked the activity to labs in China. The company disclosed the finding rather than staying quiet, which is the transparency signal we look for. ### Can I stop my prompts from being used for AI training? Usually yes, but it depends on the vendor and your plan. Business and enterprise tiers often exclude your data from training by default, while free and consumer tiers typically opt you in. Check the settings page and the trust center, not the marketing site. ### How do I vet an AI tool's training data ethics quickly? Ask three things: whether training is opt-in or opt-out, how long data is retained, and whether humans review your inputs. If a vendor answers all three clearly on a public page, that is a good sign. If it dodges, assume the worst and keep sensitive data out.

AI Training Data Ethics: How to Vet Any AI Tool

Frequently asked questions

Related articles