People Want You to Disclose AI Use, Then Trust You Less When You Do

A 2026 study in Computers in Human Behavior: Artificial Humans tested how people judge workplace messages depending on who or what wrote them. Human-written, AI-assisted, fully AI-generated. Then it measured trust, authenticity, and whether recipients actually used the information.

Two findings, and the second one is the awkward one.

First, an AI penalty: messages involving AI were rated less trustworthy, less authentic, and less useful for knowledge uptake. Second, a disclosure paradox: participants said disclosing AI use was important, and then penalised the message when the disclosure was there. The authors note the obvious consequence, which is that this creates an incentive not to disclose.

You can read the paper on ScienceDirect. It is worth going to the source, because this is precisely the sort of result that gets summarised into whichever conclusion the summariser already held.

What the Study Does Not Say

It does not say AI-assisted writing is worse. Perceived trustworthiness and actual quality are different measurements, and this study measured perception.

It also does not settle the question. Other 2026 work points the other way: a survey of 887 working adults found professionals rating AI-generated messages as professional, effective, and direct across conditions. Both can be true. Rating a message in isolation and rating a message labelled as AI-written are not the same experiment.

So the honest state of the evidence as of August 2026: labelling a message as AI-produced costs you something in how it is received, and how much depends heavily on context and on who is receiving it.

The Incentive Problem Is Real and It Is Not Yours to Solve

A norm that asks for disclosure while punishing it will produce less disclosure. That is not a moral failing of the people surveyed. It is what happens when you attach a cost to honesty and no cost to silence.

There is a version of this post that ends with a stern paragraph about disclosing anyway, because integrity. You already know that, and it is not the useful part, because most workplace disclosure questions are boundary questions rather than integrity questions: is this the kind of AI involvement anyone would want to know about.

We drew that line in should you disclose that you used AI to write an email: disclose when AI produced the judgment, not when it produced the grammar. The new study does not move the line. It explains why the line matters, because a disclosure attached to a trivial edit spends your credibility on nothing. For the cases where disclosure is required rather than optional, when you actually have to disclose that you used AI covers the obligations.

The Better Response Is to Reduce the AI, Not Hide It

Here is the reading of the study that actually helps.

If recipients detect something and penalise it, the something is worth investigating. There is a well-documented candidate: AI-written workplace messages have a recognisable shape. Warm opener, context nobody asked for, the actual point in paragraph three, closing offer of further assistance. Fluent and generic, and generic reads as low-effort regardless of what produced it. That is a draft problem rather than a disclosure problem, and it is the argument in why AI-written emails get ignored.

So use AI for the part with no judgment in it and keep the judgment yours. Tone, register, grammar, length: fine. What to say, what to commit to, what to leave out: yours. A message where you made every call and the machine fixed the phrasing does not read as AI-written, because structurally it is not.

Before:

Hi Sam, I hope this message finds you well. I wanted to reach out regarding the Q3 reporting timeline, as I know this is an important priority for the team. Unfortunately, we have encountered some challenges which may impact our ability to deliver on the original schedule. I would welcome the opportunity to discuss this further at your convenience. Please let me know if you have any questions.

Best regards

After:

Hi Sam, the Q3 report will be two days late. The finance export broke on Tuesday and we rebuilt it manually. New date is Thursday 27th. Nothing else in the timeline moves. Happy to talk if that causes a problem downstream.

The first one contains no information. The second one contains four facts and a date, and nobody reading it wonders what wrote it, because there is nothing generic left to notice.

A Context That Does Not Sand Off the Specifics

The failure mode when you hand a blunt draft to a model is that it politely removes exactly what made the message useful. Constrain it.

A Wrivio Context for internal updates could say:

Rewrite this as a direct internal message. Put the outcome or the ask in the first sentence. Same length or shorter. Keep every name, date, figure, and commitment exactly as written. Do not add a greeting, an opening pleasantry, background, or a closing offer to help. Do not soften a stated deadline or a stated problem.

Press Ctrl+Shift+Space, paste the draft, and read the word-level diff rather than the finished text. The diff is where you catch a hedge being inserted into a commitment, which is the edit most likely to cause an argument later. More on that in how to review AI rewritten text.

If you want output that sounds like you rather than like a model, the Styleprint examples attached to a Context are the mechanism, and how to make AI-written emails sound like you covers choosing examples that carry your voice.

What to Take From This

Do not read one study as licence to stop disclosing, or as proof that AI writing is bad.

Read it as evidence that recipients are sensitive to something in AI-mediated messages, that a disclosure label carries a cost you should spend deliberately, and that the most reliable way to avoid the penalty is to send messages with your judgment in them. Which was already the advice, and now has a citation behind it.

Common Questions

What is the AI penalty in workplace communication?

It is the finding that messages identified as AI-assisted or AI-generated are rated lower on trustworthiness, authenticity, and usefulness than human-written equivalents, reported in a 2026 study in Computers in Human Behavior: Artificial Humans.

What is the disclosure paradox?

Participants in the same study said disclosing AI use was important, but rated the communication lower when the disclosure was present. The authors flag that this creates an incentive against disclosing.

Should I stop telling colleagues I used AI?

Not as a blanket rule. Disclose when AI shaped the judgment or the substance, and skip it for grammar and tone edits, which is the boundary most people already apply to spellcheck.

Does the research agree that AI writing is perceived worse?

Not uniformly. A separate 2026 survey of 887 working adults found professionals rating AI-generated messages favourably on professionalism and effectiveness, so context and framing clearly matter.

How do I avoid my messages reading as AI-written?

Lead with the point, keep specifics and dates intact, cut opening pleasantries and closing offers of help, and make the substantive decisions yourself rather than asking a model to decide what to say.

Download Wrivio for Windows to fix the phrasing of a message you wrote while the facts, the ask, and the judgment stay yours.