Nobody adds a footnote saying they used spellcheck. Almost everyone agrees that handing over an entirely generated report as your own analysis is a problem. Between those two points sits nearly all real work, and most organizations have no useful guidance for it — which leaves individuals guessing and produces both over-disclosure that means nothing and under-disclosure that damages trust when discovered.
A workable line exists. It is about judgment, not tools.
What changed in 2026
- Use became near-universal and disclosure did not. With AI assistance embedded in most professional software, the gap between how much it was used and how much it was mentioned widened considerably.
- Regulated professions got specific. Legal, medical, financial, and academic contexts developed concrete disclosure requirements, replacing general ethical hand-waving.
- Blanket labels lost credibility. Organizations that required a disclaimer on anything AI-touched found the labels became invisible within weeks.
- Shadow use grew where policy was restrictive. Bans correlated with undisclosed use rather than with less use, which made review impossible.
A test that works
| Situation |
Disclose? |
| Grammar, spelling, formatting help |
No |
| Rephrasing your own sentences |
No |
| Research assistance you then verified independently |
Generally no |
| Analysis you reviewed line by line and stand behind |
No, but be ready to explain your process |
| A draft you edited lightly and did not fully verify |
Yes |
| Generated content passed on largely as produced |
Yes, clearly |
| Anything to a client, regulator, or court |
Follow their rules, always |
| Creative work presented as your own craft |
Yes, norms are strict here |
The organizing principle is whose judgment stands behind the output. If you reviewed every claim, checked the numbers, and would defend each sentence under questioning, the work is yours and the tool is irrelevant. If you did not, the reader is relying on something you have not verified, and they deserve to know.
That test has a useful property: it makes disclosure a statement about your review process rather than an admission. "I drafted this with AI assistance and verified the figures against source" is more informative and more professional than either a bare label or silence.
Writing a policy people follow
Three properties separate policies that work from ones that get ignored.
Be specific about tasks, not tools. Naming approved and prohibited uses survives the arrival of new products; naming approved products does not.
Address data separately from disclosure. What may be pasted into which tool is a different question from what must be disclosed, and conflating them confuses both. Client data, personal information, and unreleased material each need clear handling rules, which connects to the inventory work in AI tool sprawl.
Make the default permissive with defined exceptions. Restrictive policies produce undisclosed use, which is the outcome you were trying to prevent. Permitted-with-review, plus a short list of contexts requiring disclosure and a shorter list of prohibited uses, is far more likely to be followed.
For hiring specifically, the same reasoning applies to what candidates should be told — AI interview cheating covers where employers are drawing that line.
Common mistakes
- Requiring disclosure on everything. Universal labels stop carrying information.
- Banning use outright. It moves usage out of sight rather than eliminating it.
- Naming specific products in policy. The list is stale within a quarter.
- Ignoring client requirements. External obligations override internal comfort, and violating them has consequences beyond embarrassment.
- Leaving individuals to guess. Absence of policy is itself a policy, and it produces inconsistency people resent.
FAQ
Should I disclose AI use in an email?
For routine correspondence you reviewed, no. For substantive analysis or advice the recipient will act on, describing your process is worthwhile.
What about code?
Team norms vary considerably. The consistent expectation is that you understand, tested, and can maintain what you commit, regardless of how it was written.
Does using AI make work less valuable?
The value was always in judgment, framing, and accountability. Those remain yours. Being open about process tends to increase trust rather than reduce it.
What if my employer has no policy?
Apply the judgment test, follow any client or regulatory requirements, and ask if you are unsure. Asking is never the wrong move.
Where to go next
For auditing what your team is actually using, read AI tool sprawl. For the hiring-side version of the same question, AI interview cheating.