# Responsible AI

By [DYLIT Chronicles](https://dylit.info/user/dylitmediabuzz)

[Everything AI - beyond the hype](https://dylit.info/pr/everything-ai-beyond-the-hype/6a9efac02e92664f4d50cf9d) > [Responsible AI](https://dylit.info/ch/responsible-ai/6ab3685f99638e119f592dae)

Assume Anything You Paste In Could Be Read by Someone Else Adopt one default and most AI privacy incidents stop happening. Assume anything you paste into a tool could be stored, reviewed by a human, used to improve the service, or surfaced in a breach. Sometimes that assumption will be stricter than the reality. That costs you very little. The reverse assumption, made once, on the wrong document, costs a great deal. Why the assumption is reasonable Vendor practices differ widely and change without much fanfare. Many services retain conversation history by default. Many use inputs to improve their models unless you are on a tier or setting that excludes it, and those defaults commonly differ between free and paid versions of the same product. Many have human review processes for safety and quality, which means a person may read what you wrote. None of this is sinister. It is ordinary software operation. But it means the mental model of "a private conversation with a computer" is wrong, and the interface actively encourages that wrong model. It looks like a chat window. It feels like typing to yourself. What should not go in Specific and non-negotiable for any tool not explicitly cleared for it: Customer personal data. Names with contact details, account numbers, health information, anything identifying a real person who did not agree to this. Employee records. Performance notes, disciplinary matters, salary information, medical details, anything from an HR file. Credentials. API keys, passwords, tokens, connection strings. These turn up in pasted logs and config files more often than anywhere else. Unreleased financials. Results before announcement, forecasts, deal terms, anything with a disclosure obligation attached. Third-party confidential material. A client's documents, licensed research, anything under an NDA. Your obligation to them does not pause because it was convenient. Anonymise as a habit The practice that makes most of this manageable. Replace real names with placeholders. Swap actual figures for representative ones of similar magnitude. Strip identifying details that are not needed for the task. "Customer A, on the Enterprise plan, has raised eleven tickets this quarter about integration latency" preserves everything the model needs to draft a useful response. It contains nothing you would mind seeing outside the company. Then put the real details back in yourself, afterwards. Thirty seconds of work, and you have removed the entire category of problem. The habit part is what matters. Deciding case by case means eventually deciding wrong on a day when you are rushed. Anonymising by default means you do not have to get the judgment call right. The question to ask When unsure, one test cuts through: Would I be comfortable if this appeared outside the company? Not "will it." Whether you would be comfortable if it did. If the honest answer is no, either anonymise it or use a tool your organisation has cleared for that level of sensitivity. This test works because it bypasses the argument you would otherwise have with yourself about how likely a leak is. Probability is not the point. You are making a decision about exposure, and you should make it deliberately. The organisational side For anyone setting policy rather than following it: Approve specific tools for specific sensitivity levels, and say so plainly. "Use AI responsibly" is not a policy. "These two tools are approved for internal documents, this one is approved for customer data, nothing else is approved for anything" is. Make the approved path easy. People route around friction. If the cleared tool is slow to access, they will use the one in their browser, and you will not find out. Ask before, not after. Someone checking whether a document can go into a tool is a good day. Someone disclosing that it already did is a worse one, and the gap between those two outcomes is almost entirely about whether asking feels safe. The short version Assume it is not private. Anonymise by default. Use cleared tools for sensitive work. Ask before you paste, not after. Four habits, and they cost almost nothing once they are automatic.