Disclosure and records

Disclosure is a trust decision before it is a compliance one, and customers are more forgiving of AI assistance than of discovering it was hidden. The workable principle: disclose when a reasonable person would want to know. A drafted email you reviewed and sent is your message and needs no label. An AI-generated image presented as a photograph of your actual work, an AI voice on a call the customer believes is human, or an AI-written review or testimonial all cross the line — and the last of those is straightforwardly deceptive in most jurisdictions. Some contexts carry explicit duties rather than judgment calls. Advertising claims must be substantiated whoever wrote them, so "the AI wrote it" is not a defense for an unsupported claim. Chat interfaces that answer as a person invite regulation in a growing number of places. Regulated advice — legal, medical, financial — has professional standards that AI assistance does not relax. When your industry has a rule, the rule governs, and the tool is irrelevant to whether you followed it. Records are the other half. Keep enough of a trail that you could reconstruct how a customer-facing or consequential output came to exist: which workflow produced it, who reviewed it, when, and what changed. This does not require a system — a column in a spreadsheet or a note in the ticket is usually enough. Set a retention period, keep the log free of sensitive data, and make sure someone owns it. The trail costs almost nothing to maintain and is enormously valuable the one time a customer, an auditor, or your own team asks what happened. Disclosure decisions are easier when made in advance as categories rather than case by case under deadline pressure. Write the rules once and apply them without renegotiating each time. Records should be minimal by design. Storing full prompts and responses creates a retention liability of its own, while four fields — workflow, reviewer, date, changes — provide accountability at almost no cost. A design studio labels all generated imagery and logs the reviewer on every client deliverable. When a client asked whether a concept image depicted a real installation, the answer was already documented and took seconds to give.

Key takeaways

• Some contexts carry explicit duties rather than judgment calls.
• Records are the other half.
• Presenting generated imagery as documentation of real work misrepresents reality.

Sign in to complete activities, save progress, and use the AI Coach.