You know whether AI touched the work you sell. Your customer doesn’t. An AI disclosure is the short, plain statement that closes that gap: what the model did, what you did, and who is accountable when the output is wrong. Done well, it ends an awkward conversation before it starts. Done badly, it reads as either a legal shield or an apology, and both make buyers uneasy.
What an AI Disclosure Is For — and What It Isn’t
It is not a confession, and it is not a liability waiver. Its job is narrow: let a buyer predict what they are getting, and know where to go when something is wrong.
Run the materiality test before you write a word. Ask yourself: if this customer learned the full story six months from now, would they feel misled? Would it have changed whether they bought, what they paid, or how they used the deliverable? If yes, it belongs in the statement.
If no — your spell checker, the machine learning buried in your invoicing app, the autocomplete in your email — leave it out. A statement that lists every tool you have ever opened is noise, and noise hides the parts that matter.
This is operational guidance, not legal advice. If you work in a regulated field, or your client contracts already contain AI clauses, have a lawyer read your wording before you publish it.
Say Four Things, in This Order
A workable AI disclosure answers four questions and then stops:
- What the AI did. Name the task, not the vendor. Drafted first-pass outlines and alternate headlines beats leverages advanced AI technology.
- What a person did before it shipped. The review step, written in verbs.
- What AI was not used for. Boundaries tell buyers more than capabilities do.
- Where to raise a problem. An address or a form, plus a rough response time.
Here is a worked example for a freelance content writer:
I use AI models to generate outlines, alternate headlines, and rough first drafts. Every sentence that ships has been rewritten or approved by me. Facts, figures, quotes, and names are checked against original sources by hand — I don’t ship a number I haven’t traced. I do not use AI to write testimonials, case study results, or anything presented as my personal experience. Found an error? Email hello@example.com and I’ll correct it within two business days. Last updated March 2026.
Five sentences. No may, no from time to time, no where appropriate. And note the date line: it turns a vague promise into a claim about a specific process at a specific moment.
Where the Statement Belongs
Keep one canonical version at a stable URL and point everything at it. Rewriting the wording in four places guarantees that three of them go stale.
- A one-line summary plus a link in your site footer.
- On sales pages for anything content, research, design, or analysis related, where the question is genuinely live.
- In your proposal or SOW template, as a clause the client reads before signing rather than after a dispute.
- In the read-me file of a delivered asset, when that asset itself was drafted with AI help.
Where it does not belong: buried in paragraph 14 of your terms of service, or blown up into a site-wide banner that treats normal tool use as a warning label. If a marketplace, ad network, or social platform requires a per-item AI label, follow that platform’s rule there and let your own statement carry the fuller explanation.
Describe the Human Review Step So Someone Could Check It
Reviewed by a human is the weakest sentence in most statements. It is unfalsifiable, which is exactly why readers discount it. Replace it with verbs and objects:
- Numbers are traced back to the source document or the original export.
- Quotes are compared against the recording or transcript.
- Names, titles, prices, and dates are confirmed against a system of record.
- Structure and argument are rewritten by the person whose name is on the work.
Then set a threshold and publish it. A defensible default: 100% line-level review for anything containing a number, a claim, a name, a price, or a promise; lighter spot checks for internal drafts and throwaway variants. If you sample rather than review everything, say so. Buyers forgive a stated limit far more easily than a discovered one.
Back the wording with a record: date, item, reviewer, what changed. One spreadsheet row per item is enough. When a customer asks whether a person really read it, a log is the only answer that isn’t just a louder yes. The Responsible AI Use & Human Review SOP Kit packages that structure — a 16-page guide, a 9-sheet workbook, and 16 policy and SOP templates — if you would rather adapt something than build the logging from scratch.
Avoid Overclaiming and False Modesty
Overclaiming is the obvious failure. 100% human written stops being true the moment you accept an AI-generated outline. Our proprietary AI engine is a stretch if you are calling a commercial API with a good prompt. Fully AI-powered, no humans needed is a promise you have to keep on a bad day, and you won’t.
False modesty is the quieter failure. Writing I just used ChatGPT a bit makes your hours of editing invisible and hands the buyer a reason to negotiate your price down. So does the catch-all hedge we may use AI tools in our workflow — it covers everything, commits to nothing, and tells the reader you would rather not be pinned down.
The calibration rule: your statement should be specific enough that someone could catch you if it were false. If no realistic audit could ever contradict your wording, the wording is too vague to earn trust.
When Customers Ask Hard Questions
Three questions come up repeatedly. Write the answers before you need them, and make sure everyone who talks to customers uses the same ones.
Is my data being fed into a model?
Answer with specifics: which tool, what you send, what setting you have chosen, and whether the vendor’s terms allow training on your inputs. If you don’t know, say you will check and then actually check the vendor’s data-use terms. Guessing here is how a small question becomes a contract problem.
Did a real person read this?
Name the person and the pass they made. Offer the review log if the client wants it. This is the payoff for keeping records.
Can I get a version made without AI?
Decide your answer in advance: yes at a different rate and timeline, or no with a straight explanation. Both are respectable. Improvising this one under pressure is how people promise workflows they cannot actually deliver. A short internal policy makes these answers consistent, and our walkthrough on building a practical AI policy for a small team covers the decisions worth writing down first.
When something slips through anyway — a fabricated citation, a wrong figure — do not quietly delete your statement. Fix the deliverable, tell the customer which review step failed, and say what you changed so it fails differently next time. A disclosure that survives one public mistake is worth more than one that was never tested.
Keep Your AI Disclosure True as Your Process Changes
Your statement is a claim about a process, and processes drift. Treat these as triggers for a same-week rewrite:
- You add a tool, or switch to a model with different behavior.
- You remove, shorten, or automate a review step.
- A new contractor or employee starts producing client-facing work.
- You launch a product type your current wording never anticipated.
- A client’s own AI policy requires terms you have to match.
Beyond those triggers, re-read the whole thing once a quarter with your actual workflow open beside it, line by line. Keep a last updated date in the published text and archive prior versions with their dates. If a question ever arises about work delivered last spring, you want to show what you promised last spring, not what you promise now.
If you want a running start, the SOP kit above gives you the templates and the review log to sit behind your wording, and the rest of the Cursiqa workbook library covers the adjacent operational pieces. Either way, write the statement this week — the first draft will be wrong, and you can only fix wording that exists.