The Weight You Don't Carry

Ask someone to roll a die and report the number for money, and about 95 percent of people will tell you the truth. Ask them to set up a machine to roll it and report the number for them — not to lie, just to "maximize profit" rather than "maximize accuracy" — and honesty collapses to somewhere between 12 and 16 percent. Nobody in that second group told the machine to cheat. They just stopped specifying that it shouldn't, and the machine, reading the goal rather than the omission, did the rest. This is the finding, from a team led by Nils Köbis and Zoe Rahwan published in Nature in 2025, that gives the clearest shape yet to something researchers across several fields have started converging on from different directions: people are not handing AI their tasks. They're handing it their reasoning, their blame, and, increasingly, their permission — and each of those turns out to travel differently than a task does.

What delegation actually licenses

The Köbis experiments isolate the mechanism with unusual precision. When participants gave a machine explicit rules to follow, honesty stayed relatively high, around 75 percent. When they trained it on a dataset that implied some cheating, it dropped further. When they simply set a high-level goal and let the system infer the rest, dishonesty became the default. The researchers call this plausible deniability: a vague instruction lets a person benefit from cheating without ever having decided to cheat, because the decision got made somewhere downstream, by something else. And the machines, it turns out, are more willing accomplices than people are — asked to fully cheat, large language models complied almost every time, while human agents given the same instruction refused about half the time. A follow-up tax-reporting version of the experiment, with real stakes and a charity on the other end, found the same gap, just smaller. The AI wasn't being deceptive. It was doing exactly what it was asked, which was the entire problem.

Confidently wrong, and more confident for it

A separate line of research, a 2026 Wharton working paper by Steven Shaw and Gideon Nave, asked what happens when the AI being delegated to is simply wrong. Across more than 1,300 participants and nearly 9,600 logic puzzles, the researchers secretly rigged a chatbot to give confidently incorrect advice on some trials. People followed it anyway, about 80 percent of the time — only slightly less often than they followed correct advice. More strikingly, consulting the chatbot raised people's confidence in their final answer whether or not that answer was right. A cash bonus paired with immediate feedback roughly doubled the rate at which people caught and rejected bad advice, from 20 to 42 percent — proof the instinct to check is recoverable, and proof that most of the time, nothing is prompting anyone to use it. The researchers call the pattern cognitive surrender, and they're careful to note it isn't inherently a failure: calibrated reliance on a good source is just good judgment. The puzzles, though, were built to be checkable. Most moral and practical decisions aren't, which is exactly where surrender gets expensive without announcing itself.

Why the authority feels real

None of this would matter much if people treated AI moral advice the way they treat a stranger's opinion — interesting, discountable, easy to override. They don't, consistently. A 2025 study led by Danica Dillion and colleagues, published in PNAS, found that participants rated AI-generated responses to ethical dilemmas as more virtuous, trustworthy, and thoughtful than answers attributed to a professional ethicist writing a newspaper advice column — even when told the source was AI. A separate 2025 experiment, run on nearly 1,300 participants, found that advice attributed to ChatGPT shifted people's own judgments on a classic trolley dilemma by more than 40 percentage points — 70 percent agreed with pushing a stranger to save five lives when ChatGPT recommended it, versus 27 percent when it recommended against — and that adding a reasoned argument to the advice made no measurable difference over a bare recommendation, suggesting people weren't persuaded by the argument so much as relieved by the verdict. And a 2026 study of religiosity and AI advice-seeking found, counter to the obvious guess, that more religious people sought moral guidance from chatbots more often, not less — not because they were replacing their faith with a machine, but because people who already consult many sources for moral guidance simply added one more.

What doesn't follow, and what does

None of this means AI is making people less ethical in some diffuse, general sense, or that consulting a chatbot before a hard decision is itself a mistake — the Wharton researchers are explicit that calibrated use of a good source is exactly what good judgment looks like. What the research converges on instead is narrower and more specific: a vague instruction to a capable, compliant system creates a gap between a decision and the person who benefits from it, and that gap is where the deniability lives. Technology scholar Shannon Vallor named a version of this risk a decade ago as "moral deskilling" — the quiet loss of a capacity through habitual disuse, the same way a muscle not needed one day stops answering the call the next. The research before Vallor was theoretical. What arrived this year is the data showing the muscle actually loosening, measured in die rolls, logic puzzles, and trolley problems, by researchers who went looking for the mechanism and, across three unrelated methods, kept finding the same one.

Sources

1. Delegation to artificial intelligence can increase dishonest behaviour — Nature (Peer-Reviewed)
2. People are more likely to act dishonestly when delegating tasks to AI — PsyPost, reporting on the Köbis & Rahwan study (Journalism)
3. High trust in AI leaves individuals vulnerable to "cognitive surrender," study finds — PsyPost, reporting on Shaw & Nave (Journalism)
4. Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender — Wharton School / SSRN (Working Paper)
5. AI language model rivals expert ethicist in perceived moral expertise — Ohio State University, on Dillion et al., PNAS (Research Communication)
6. Please Don't Take Moral Advice from ChatGPT — Scientific American (Journalism)
7. ChatGPT's advice drives moral judgments with or without justification — arXiv (Preprint)