Three campaign slogans. Two groups of readers. One difference: half were told a marketing professional wrote them, half were told AI marketing software did. The slogans were word-for-word identical. The group that believed AI produced them was 3.4 percentage points less willing to contribute an idea of their own โ against a 26% baseline, a 13% relative collapse in voluntary effort (Brookings, 2026).
Nothing about the work changed. Only the byline.
That is the cleanest causal result yet on AI attribution inside a workflow, and it lands somewhere most operating plans have no line item for. Every discussion you have had about AI and your team has been about capability: what the model can do, how good the output is, where the error rate sits. This finding says the label on the output moves human behavior on its own, holding quality fixed. Attribution is not a communications detail downstream of the work. It is an input to how much work you get.
What the Experiment Held Constant, and What It Moved
The study is a preregistered survey experiment by Nikolova, Milanova and Wang, run identically on nationally representative samples in the United States (N=1,511) and the Netherlands (N=2,117), pooled to N=3,628 (GLO Discussion Paper 1788, 2026). The design was registered in advance in the AEA registry (AEA RCT Registry #15363, 2026) โ meaning the outcomes were specified before the data came in, which is the difference between a finding and a story.
Participants were given a small public-health brief โ people do not drink enough water โ and asked to help with a campaign slogan. Before seeing anything, they rated how meaningful the task felt, averaging about 5 on a 7-point scale. Then they read three slogans. Half were told a marketing professional wrote them; half were told AI software did. Same slogans, both arms. Afterwards they re-rated the task's meaningfulness, and were offered the chance to submit a slogan of their own. That voluntary submission was the effort measure.
Two results came out of it. Task meaning fell about 0.07 standard deviations in the AI-label condition. Willingness to volunteer an idea fell 3.4 points from a 26% base โ the 13% figure. The effects replicated across both countries.
Note what is being measured. Not satisfaction. Not sentiment. The thing people did when nobody required them to do anything: discretionary contribution. That is the exact input your process improvements, your escalations, and your quality catches actually run on.
The Penalty Lands on the Work, Too
The second finding gets less attention and deserves more.
Participants who were told the slogans were AI-made rated those same slogans as less creative and less persuasive than participants who were told a human wrote them (Brookings, 2026). Identical text, lower marks. They also reported less trust in AI's ability to do that kind of work.
For an operations lead, this is the sharper edge of the result. If your team is now reviewing AI-drafted material โ policy language, first-pass analysis, customer replies, internal briefs โ the review is not happening on a level playing field. The reviewer's assessment of quality is partly a function of the label, not only the text. Some of that discount is warranted skepticism, and you want it. But it is not calibrated to the actual defect rate of the draft in front of them, which means it produces two failure modes at once: over-editing of good AI output, and a general erosion of confidence in the pipeline that produced it.
You cannot separate those two effects by asking people. They will report, honestly, that the draft was weak.
Why AI Attribution Reprices the Standard Workflow
The default pattern that mid-market operations adopted over the past eighteen months is "AI drafts, human reviews." It was chosen because it looked motivationally neutral โ the person still owns the outcome, the machine just saves them the blank page.
This evidence says that neutrality was an assumption, not a property.
In the standard pattern, the human contribution is framed as editing. The AI contribution is framed as authorship. That framing is a choice, usually made by whoever configured the tool, and it is now measurably load-bearing. The same work reframed โ human sets the brief and owns the argument, AI supplies raw material โ is the same activity with a different byline, and the byline is the variable that moved effort by 13% in a controlled setting.
This connects to a gap already visible in the field data. BCG's 2026 survey of 11,749 workers found 47% now spend more time managing and directing AI than doing the work themselves, while 66% report no meaningful guidance on how to use the time AI frees (BCG, 2026). Roles are being restructured around review and direction โ precisely the position the experiment shows is motivationally exposed โ with no corresponding design decision about how the human contribution gets named.
The productivity implication is the one the authors themselves raise: suppressed discretionary effort may be part of why large experimental AI gains have not shown up in economy-wide productivity numbers. You can get more output per task and less contribution around it, and the second effect will not appear in any tooling dashboard.
Where This Shows Up in a 50 to 500 FTE Operation
Four places, in rough order of how quickly they bite.
Internal communication. Policy updates, process changes, and all-hands material drafted with AI and recognizably so. The message is identical to the human-written version; the reception is not. If you need people to act on it voluntarily rather than comply with it, the label is working against you.
Peer review and quality gates. Adaptavist's 2026 study of 2,500 knowledge workers found 52% regularly correct AI-generated work produced by colleagues (Adaptavist, 2026). That correction load is already real. What this experiment adds is that the reviewer's judgment of severity is partly driven by the label, so your rework volume is not a clean signal of your defect rate.
Recognition and performance conversations. When a strong piece of work is known to be AI-assisted, both the contributor's sense of ownership and the manager's read of the contribution shift. Nothing in a standard performance process accounts for this.
The compounding problem
None of these four is severe in isolation. The mechanism is exposure frequency. A single AI-labeled artifact costs you a fraction of one person's discretionary effort on one task. A pipeline where most first drafts, most summaries, and most routine analyses arrive labeled means the same small discount applies dozens of times a week to everyone downstream. The experiment measured one encounter. Your operation runs thousands.
That is also why the effect is hard to see from the top. No individual instance produces a visible failure. What it produces is a slow flattening of the voluntary layer โ the suggestions that never get made, the second look nobody takes โ and that layer has never appeared on an operations dashboard because it was always free.
Customer and candidate-facing copy. The same discount applies outside the building, and the study's respondents were general population samples, not employees. Disclosure practices designed for trust and transparency carry a measured effect on perceived quality. That is not an argument against disclosure โ it is an argument for knowing the price of the form you choose.
What This Evidence Does Not Say
Three limits, stated plainly, because a plan built on overreading this will fail.
It is a short creative micro-task, not a work process. Participants judged output labeled as AI-made; they did not collaborate with an AI system over weeks on work they were accountable for. External validity to sustained professional work is untested and the authors say so.
The meaning effect is small. 0.07 standard deviations is a modest standardized effect size. The headline 13% is a relative change off a 26% base rate, which amplifies a 3.4-point absolute move. Both numbers are accurate; only one is dramatic.
It is a working paper. Preregistered and replicated across two countries, which is more than most evidence in this space carries โ but not yet peer-reviewed.
What survives all three caveats is the direction and the mechanism: attribution moves effort independent of quality. If the effect is even half this size in a real workflow, and workers encounter AI-labeled output constantly rather than once, it accumulates.
Making Attribution a Process Parameter This Quarter
The response is design, not messaging. Four moves, none of which require new spend.
Name the human contribution, not the tool. Where work carries an AI provenance label, make sure it also carries what the person did โ the brief, the judgment call, the decision to ship. The authors' own recommendation is to design workflows that keep human agency visible and individual contributions attributable. Costless to implement; currently absent from most templates.
Separate provenance from evaluation. In any review gate that matters, have the reviewer assess the artifact before they see how it was produced. You will learn quickly how much of your rework is defect and how much is discount.
Stop framing people as editors. "Review this draft" and "own this output, here is raw material to start from" describe the same hour of work and different jobs. Change the language in your process docs and your tickets. It is the highest-leverage item here and it costs a wording change.
Instrument discretionary contribution, not adoption. Seat counts and prompt volumes cannot detect this. Track the voluntary stuff: unsolicited improvement suggestions, defects caught before the gate, people adding scope that nobody asked for. If that curve is bending down while output volume climbs, you have found the effect in your own data.
The decision in front of you this quarter is not whether to disclose AI involvement in work. It is narrower than that, and it is a design question: whether the AI attribution baked into your current processes is something you chose deliberately, or something a tool default chose for you. In this experiment, the label was worth 13% of voluntary effort. It is the cheapest variable on your board and, so far, the only one nobody owns.
