A real Monet was posted online in May with a lie attached to it. An X account shared a painting from Monet’s Water Lilies series, claimed to have generated it with AI, and asked people to describe, in as much detail as possible, what made it inferior to a real Monet. More than 1,300 people obliged. They found the composition lacking (“no coherent composition… there’s nothing really to focus on”), the colours wrong (“an incoherent muddle of inconsistently saturated greens”), the whole thing empty of feeling — “it is not Monet”, one self-described art professional concluded, more like “an undergrad art student’s study from a museum visit”. Somebody called it garbage. Then came the reveal. The painting was real, the post had reached 6.8 million people, and some of the most confident critics deleted their replies.

Phill Agnew wrote the episode up for HubSpot as a story about how we judge AI content, and the psychology he points to is real enough: we devalue work we believe took no effort, and the anti-AI reflex runs deep. But the more useful lesson sits somewhere else entirely, and it has nothing to do with AI. Nobody in that thread stumbled across those flaws while admiring the painting. They were asked to find them, so they went looking, and people who go looking for flaws usually find them. I’ve had a saying for years: if you ask for someone’s opinion, they’ll give it.

The opinion may not have existed

I wrote an essay last year arguing that opinions are not data — that what people say they’d do is a poor predictor of what they actually do, and that organisations keep treating stated intentions as behavioural evidence. That argument was about the answer. This one is about the question — about what asking for feedback does to the person being asked, because a question can create the very dissatisfaction it appears to discover. Opinions are not data, and sometimes they weren’t even opinions until you asked for them.

The mistake starts with how we picture feedback. We imagine it as retrieval: the concern already exists in someone’s head, and asking simply collects it. Some feedback does work that way, which is what keeps the picture alive. Much of what comes back, though, is made to order. The question sets an evaluative task; the person on the receiving end works out what’s expected of them, inspects the work through that frame, constructs a judgement on the spot, and reports it as though it had been sitting there all along. The judgement is usually sincere, which is what makes this hard to spot in the room. The critics in the Monet thread weren’t pretending — the muddled greens looked muddled to them once they’d been told to look for muddle.

We’ve all been in the meeting where this happens. The CEO goes around the table asking what everyone would change, the first four people offer something, and by the time the question reaches us, saying “nothing comes to mind” feels like failing to take part. So we find something — and the something is real to us the moment we say it. The behavioural literature has names for every piece of this — Norbert Schwarz on how a question’s wording shapes the self-report it collects, Bettman, Luce and Payne on preferences constructed at the moment of the task, Martin Orne’s “demand characteristics” — but you don’t need any of it to recognise the meeting.

What the question smuggles in

Look again at the wording of the Monet post, because it’s a small masterpiece of loaded questioning. It did three things at once. It labelled the work: you’re looking at AI. It presupposed the conclusion: the image is inferior, and the only open question is how. And it set an assignment: explain that inferiority in as much detail as possible. The people who replied were briefed as prosecutors rather than invited to serve as jurors. The verdict came supplied, and their job was to produce the evidence for it.

A caveat, before building anything on it: this was a clever social-media stunt, not a controlled experiment. The respondents were self-selecting, they were performing for an audience, and their motives were mixed. There’s no way to isolate how much of the pile-on was anti-AI prejudice, how much was conformity to the replies already visible, how much was the wording, and how much was people enjoying the chance to perform expertise. It illustrates the mechanism vividly; it can’t measure it.

The reason it deserves a manager’s attention anyway is that its structure is the structure of most workplace feedback requests. “What would you change about this proposal?” presupposes change. “Does anyone have concerns?” signals that concerns are expected, and that voicing one is the way to be useful. “Can everyone give me one suggestion for improvement?” makes criticism the entry fee for participation. None of these is as crude as the Monet post, but each supplies a frame and sets an assignment, and people complete assignments.

When answers become evidence

Now watch what an organisation does with the results. A founder shows a new proposition to ten colleagues and asks each of them what they’d change. Each produces a suggestion, because that’s the task, and each suggestion is individually plausible. The founder walks away with a list of ten apparent defects, and the list feels like evidence — it’s written down, it came from ten different people, it has the texture of due diligence. None of this makes the founder foolish, either: asking felt responsible, even generous, because inviting input is what good leadership is supposed to look like.

Consider what that list actually establishes, though. It doesn’t show that any of the ten was dissatisfied before being asked, that a customer would ever notice any of the issues raised, that the suggestions are compatible with one another, or that implementing them would leave the work better than it started. The volume of feedback has been mistaken for the severity of the problem, and they’re different measurements — one records how many people you asked, the other would require evidence nobody collected.

Once that confusion settles in, the costs turn concrete. Picture the same proposition six weeks on. The ten suggestions have become a tracked backlog with an owner; the launch has slipped a month while three of them are implemented; and the proposition itself has grown longer, safer and less distinctive as each reviewer’s contribution is accommodated — all without anyone having established that a single customer had a problem with the original. Even the suggestions nobody acts on have to be logged, triaged and diplomatically declined. And every one of them carries a production cost the suggester never pays. Feedback is cheap to give and expensive to absorb: the giver spends thirty seconds, the receiver spends days, and the accountability for the outcome stays with the receiver too. I’ve argued before that “I was just following advice” has never excused a decision, and the same holds for feedback you didn’t need but implemented anyway.

The subtler costs are political. Ten people naming ten different minor issues gets summarised upward as “everyone agreed it needed work” — a consensus that never existed, assembled from unrelated gripes. And founders learn to mistake repeated review for rigour, sanding away the conviction that made the idea worth reviewing in the first place — changed, certainly, but not improved. What emerges from five rounds of solicited feedback often carries the fingerprints of everyone who was asked and the conviction of no one.

Generative AI has given all of us a daily rehearsal of this mechanism, incidentally. Ask a model to improve a piece of writing and it will oblige — every time, forever, regardless of the writing’s quality. Its inexhaustible willingness to produce another revision tells you nothing about whether revision was needed. People, handed the same assignment, behave much the same way.

The failure on the other side

None of this licenses the conclusion some managers would love to reach: that feedback is unreliable, so the listening can stop. That’s the opposite failure, and it’s at least as expensive. Real problems go unreported because people fear contradicting authority. Customers leave without ever explaining why. Leaders end up surrounded by agreement at precisely the moments they most need challenge.

So there are two distinct failures worth keeping apart: suppressed feedback, where a real problem exists and nobody surfaces it, and commissioned dissatisfaction, where no problem has been demonstrated but everyone has been instructed to produce an improvement anyway. Most feedback processes are designed to fight the first and end up manufacturing the second. The remedy is to design better questions and to hold a higher evidential bar before acting: create conditions in which real problems can surface, without building a process in which everyone is required to invent one.

Give people permission to say it’s fine

A good feedback process must make “leave it alone” an acceptable outcome. That single change — letting “nothing material” count as a real contribution rather than a failure to participate — is the foundation the rest of the process builds on. In practice, it looks like a gate that input has to pass before it becomes work:

  1. Establish whether a problem exists: instead of “what would you improve?”, ask “is there anything here that would prevent this achieving its purpose?” or “would you be comfortable proceeding with this as it stands?”. Both leave room for the answer to be no.
  2. Ask for observation before interpretation: “where did you hesitate, get confused or lose interest?” tells you more than “what don’t you like?”, and “what have customers done in comparable situations?” tells you more than “would customers want this?”.
  3. Classify what comes back: a demonstrated defect, an observed friction, a personal preference, a speculative idea and a risk worth investigating are all useful, but they don’t deserve equal weight.
  4. Apply a value threshold: would the change create enough value to justify its cost, complexity or delay? A surprising share of suggestions fail this test the moment it’s asked.
  5. Protect the original objective: a clearer sentence isn’t better if it loses the personality, and an added feature isn’t better if it weakens the simplicity people were buying.
  6. Test where you can: prototype it, put it in front of people, watch what they do. Watching behaviour will often settle an argument that another round of opinions would only prolong.

One refinement for the cases where improvement genuinely is the goal: a team led by Hayley Blunden found, across a field experiment of more than 27,000 comments, that asking for “advice” rather than “feedback” draws out input that’s more specific and easier to act on. Notice, though, which problem it solves — how to improve, once improving has earned its place as the objective. Establish that a material problem exists, decide the fix is worth the cost, and only then commission advice on how. Most organisations run that sequence backwards, starting with the suggestions and working back towards a justification.

The Monet was not improved

The thread produced well over a thousand detailed criticisms, and not one of them established that the painting needed correcting. Had Monet been alive and dutiful enough to action every comment, the result wouldn’t have been a better painting — merely a more reviewed one. So hold to a simple rule: don’t ask people to improve something until you’ve given them permission to conclude that it doesn’t need improving. They’ll take the assignment seriously either way. Ask for an opinion, and you’ll get one.