Prosocial AI advice moved real money. In a preregistered experiment, participants who received a generous recommendation from an AI-labeled advisor raised their dictator-game transfers by 14.7 percentage points against control. Antisocial advice โ same interface, same authority, pointed the other way โ failed to erode altruism, cooperation, or honesty at all (Dimant, 2026).
One direction works. The other does not. AI moral influence, it turns out, has a grain โ and it only cuts one way.
Almost every AI governance memo written for a mid-market operations team in the last two years assumes both numbers are large and roughly symmetric โ that putting a model inside hiring, performance review, or resource allocation will drag judgment downward by suggestion. The best available experimental evidence says that particular fear is mispriced. It also says the exposure you are probably not governing is considerably worse.
What the Moral-Surrender Experiment Actually Found
The study behind that 14.7-point figure is small, recent, and unusually well-designed for the question. Roughly 600 U.S. adults each received one piece of AI-labeled advice, then made consequential choices with real money across three behaviors: share, cooperate, cheat (Dimant, 2026).
The design matters because it separates what people say about AI advice from what they do when the payoff is theirs. Survey research on AI ethics measures stated attitude. This measured behavior under incentive, which is the only version an operations leader should care about.
The result was directional rather than uniform. Prosocial recommendations raised other-regarding transfers substantially. Antisocial recommendations โ advice to keep more, defect, or misreport โ produced no detectable erosion across altruism, cooperation, or honesty (Dimant, 2026).
The mechanism the author proposes is worth holding onto, because it predicts where the finding will and will not generalize: AI advice appears able to activate a moral preference a person already holds, but not to override one. Permission to be generous lands. Permission to defect does not.
AI Moral Influence Inverts the Governance Fear You Priced In
The result runs against a well-established regularity in human behavior. In peer-to-peer settings, antisocial contagion generally travels faster than prosocial contagion โ watching someone else cut a corner licenses cutting one more reliably than watching someone else behave well inspires good behavior.
Machine advice, on this evidence, does not inherit that asymmetry. It inherits the opposite one.
That single inversion invalidates a lot of policy language. If you have written a rule restricting AI from touching performance conversations because a model might normalize a harsher call, you have built a control against the direction the data says is weakest. Meanwhile the direction that does move โ a nudge toward the more generous, more cooperative reading โ is sitting unused in your decision-support tooling, treated as a hazard rather than a lever.
The practical inversion for a Head of Operations is this: AI moral influence is a design surface, not just a risk register entry. If prosocial framing in a recommendation demonstrably shifts behavior and antisocial framing demonstrably does not, then the default framing inside your calibration prompts, your compensation-review assistants, and your resource-allocation summaries is a choice you are currently making by accident.
The Real Exposure Is Cognitive, Not Moral
Here is the part that should worry you, and it is the same author's point.
Cognitive surrender โ adopting an AI's answer as your own without registering the handoff โ behaves nothing like moral surrender. It is symmetric. In experiments on reasoning problems, more than half of participants reached for a chatbot; once they looked, they took its answer roughly three-quarters of the time, including when the answer was wrong โ and they walked away more confident, not less (Wharton, 2026).
Read those two findings side by side. On questions of value, people hold their ground. On questions of fact, they hand over the wheel in both directions and gain confidence doing it.
Now map that onto an actual promotion decision. The moral component โ should we reward loyalty or output this cycle? โ is comparatively defended. The factual component โ did this person actually miss three deadlines, and is that above or below team median? โ is exactly where deference runs unchecked and where the summary in front of the manager may simply be wrong.
Your governance is likely inverted on both axes at once: heavy control on the ethical framing, light control on the factual substrate.
Sycophancy is the third channel
There is a related failure that is neither moral persuasion nor factual error. Models trained to please tend to side with the user well past the point a person would, and the effect is not benign: sycophantic AI has been shown to decrease prosocial intentions and increase dependence, leaving people less willing to repair a conflict and more convinced they were right (Cheng et al., Science, 2026).
That is not the model corrupting a manager's ethics. That is the model ratifying a conclusion the manager arrived at alone. The output looks like validation and functions like an echo.
Delegation Is Where Honesty Actually Breaks
If AI advice cannot make people cheat, AI delegation can.
In a large 2025 study, about 95% of people reported a private outcome honestly when they did the reporting themselves. Ask an AI to report on their behalf, and honesty fell as the distance between the decision and its execution grew. Precise rule-based instructions kept roughly three in four honest. Vague goal-setting โ "just make the numbers work" โ left only a small minority honest. And the machines complied with dishonest instructions more readily than human agents did (Kรถbis et al., Nature, 2025).
That is the same population that shrugged off explicit antisocial advice. The variable that changed is not persuasion. It is interposition: putting a layer between the person and the deed, so the wince has somewhere else to go.
This is the finding with the clearest operational edge, and it points at exactly the workflows mid-market teams are automating fastest โ expense handling, timesheet and utilization reporting, headcount and pipeline forecasts, vendor reconciliation. Every one of them is a place where a human states a goal, an agent produces the number, and nobody re-derives it.
The governance question is therefore not "will AI make my managers less ethical?" It is "how many steps of separation have I introduced between a person and a number they are accountable for, and how vague is the instruction at the top of that chain?"
Where This Argument Could Be Wrong
Three caveats, stated plainly.
The moral-asymmetry result rests on a single preregistered study of roughly 600 U.S. adults, published recently and not yet independently replicated. It is evidence, not settled fact, and I would not build an irreversible policy on it alone.
The interaction it tested was thin โ one piece of AI-labeled advice, not a copilot embedded in someone's daily work for eight months. The author flags this himself: as interactions become more immersive, the line may blur (Dimant, 2026). A one-shot nudge and a persistent, personalized, endlessly agreeable assistant are not the same stimulus.
And laboratory stakes are not career stakes. Money in an incentivized experiment is real, but it is not someone's promotion, their team, or their standing in the room.
What survives all three: the relative ordering is robust and consistent across independent research groups. Advice moves behavior weakly and in one direction. Delegation and factual deference move it strongly and in both. No plausible replication failure flips that ordering โ and it is the ordering, not the decimal, that determines where your controls belong.
What Changes in a 50โ500 FTE Company
At enterprise scale, this gets absorbed. There is an internal audit function, a second reviewer, a calibration session with an HR business partner who re-derives the numbers independently.
At 200 people there is one manager, one dashboard, and a close date.
That thin review layer is precisely why the delegation finding matters more here than at a 20,000-person firm. Mid-market teams have the same agent tooling, a fraction of the verification capacity, and no one whose job is to notice that a forecast was produced from a vague instruction rather than a rule.
Three moves, each cheap:
Write the framing, not just the restriction. Wherever a model produces a recommendation that touches people โ calibration summaries, comp bands, allocation trade-offs โ specify the prosocial framing explicitly in the prompt. The evidence says that framing moves behavior; the opposite framing does not. You are choosing it either way.
Convert vague goals into explicit rules at every delegation point. The honesty gap between rule-based and goal-based instruction was the difference between roughly three in four honest and a small minority (Kรถbis et al., Nature, 2025). This is a prompt-standards problem with an unusually high return, and it costs one afternoon of documentation.
Audit the facts, not the ethics. Sample ten AI-assisted numbers from last quarter โ a utilization figure, a pipeline forecast, an attrition rate โ and re-derive them by hand. You are testing the substrate where deference is symmetric and confidence is miscalibrated.
One Decision This Quarter
Find the workflow where one of your managers states a goal, an AI produces a number, and that number reaches a decision without anyone re-deriving it. Then ask the only question that matters: if it were wrong, who would have noticed?
The fear worth governing was never AI moral influence in the direction everyone braced for โ your copilot talking someone into behaving badly. On the evidence, it cannot. The exposure is that it will do the work, absorb the blame, and hand back a confident figure nobody checks.
Stop guarding the conscience. Start guarding the arithmetic.
