AI Does Not Just Produce Slop. It Produces Feedback Slop.

A team ships faster, generates more reviews, runs more design critiques, and logs more retrospective notes than ever. The activity dashboard is green. The decisions are getting worse. Both things are true at once, and the second one is invisible until something expensive breaks.

AI did not cause the dysfunction. It accelerated dysfunction that was already there. A team with a healthy feedback culture gets a faster healthy culture. A team that punishes dissent gets a faster, better-documented machine for suppressing it. Most teams are the second kind and do not know it.

Pavel Samsonov made this argument in Product Picnic in June 2026, focused on design and product work. The mechanism he describes is general. It applies to any function now routing its feedback through models.

Debunked rules come back as automated best practice

Samsonov points to a specific failure that should worry anyone using AI to evaluate work. Design “laws” that were disproven decades ago are now baked into LLM training data. Miller’s “7 plus or minus 2” as a limit on interface elements. “Fewer clicks is always better.” These were never sound, and the field moved past them. The training corpus kept them.

So when a model reviews a design, it resurfaces them with confidence. The bad rule comes back wearing the costume of a best practice. A junior designer who reads that critique has no way to know the rule was retired before they were born. The model does not flag uncertainty. It asserts.

This is feedback that looks like quality control and functions as misinformation. The reviewer is fast, available, tireless, and wrong in a way that is hard to catch because it sounds like the consensus of the entire industry.

Abundant feedback is not the same as useful feedback

The promise was that AI would give everyone a reviewer. The result is that everyone now drowns in review.

A model will generate fifty comments on a document where a senior colleague would have left three. Forty-seven of them are noise: style preferences, restatements of obvious points, generic suggestions that apply to any artifact. The three that matter are buried in the pile, indistinguishable by volume from the rest. The reader’s attention is a fixed budget. Flood the channel and you bury the signal instead of enriching it.

This is where the dysfunction compounds. Samsonov references organizations claiming roughly 300% efficiency gains while still missing their deadlines. The activity went up. The throughput of feedback, drafts, and revisions went up. The thing that was supposed to follow, shipping on time, did not. The metric measured motion. The deadline measured outcome. They came apart.

When feedback becomes abundant and cheap, it stops being a signal and becomes a texture. People learn to skim it, then to ignore it, then to generate it for the same reasons they ignore it: because the system rewards volume.

The critique that mattered gets buried on purpose

Here is the part that turns an annoyance into a governance problem. Abundant feedback gives a leader a tool they did not have before: plausible selection.

When ten reviewers leave one critical comment each, a manager has to engage with ten humans. When a model leaves fifty comments and a team leaves another thirty, a leader can scroll, find the forty that praise the direction they already chose, and quote those. The single comment that said “this will not work for the actual user” is still in the document. It was just outvoted by noise that the leader curated.

Samsonov invokes a line attributed to Erika Hall: seek feedback, not approval. The AI-flooded channel inverts it. Leaders get a mechanism for harvesting approval that wears the clothing of feedback-seeking. They asked. The system responded at volume. They picked. Every step looks open. The outcome was decided before the first comment landed.

Employees read this fast. They watch which feedback gets acted on and which gets buried under a hundred model-generated suggestions. Then they stop offering the critique that matters. Samsonov’s phrase is exact: they become conditioned to keep their mouths shut. Not because anyone forbade dissent. Because dissent now disappears into the same bucket as autocomplete, and arguing with a bucket is not worth the political cost.

That is learned helplessness, installed quietly, measured nowhere. The org chart still shows feedback flowing. The real signal went underground.

Why this is worse than output slop

We have written before about fabricated marketing data and productivity theater, in the marketing governance reckoning and the two percent productivity problem. Output slop is bad. A fabricated prospect list is bad. But output slop is at least inspectable. You can check a number against a source.

Feedback slop attacks the mechanism you would use to catch every other failure. The feedback loop is how a team learns it is wrong. Corrupt the loop and you lose the ability to detect that anything is corrupt. The team keeps shipping confident decisions on suppressed signal, and the first hard evidence of the problem arrives as customer churn, a failed launch, or attrition of the people who stopped speaking up.

By then the cause is months upstream and untraceable. The decision that sank the launch looked well-reviewed at the time. It had forty supportive comments.

Govern the feedback, not just the output

The fix is to tie feedback to decisions and refuse to let volume stand in for signal. Concrete moves:

Separate generated feedback from human feedback in the channel. A model comment and a colleague comment should never sit in the same undifferentiated list. Tag the source. A leader who wants to cite “the feedback” should have to say which kind.

Require decision logs that name the dissent. For any decision above a threshold, the record must state the strongest objection raised and why it was overruled. If the strongest objection on file is a style nitpick, that is itself a signal that the real critique was buried.

Cap the model’s reviewer role. Use AI to surface specific, checkable categories of issue, accessibility violations, broken links, contract mismatches. Do not use it as a general “what do you think” reviewer that generates open-ended commentary by the dozen. Open-ended volume is exactly the failure mode.

Audit retired rules. If your team uses AI to critique work, keep a list of debunked “best practices” the models still parrot, and review against it. The model will not tell you a rule is dead. Someone has to.

Do this now: pick your last three significant decisions and find the strongest objection that was raised against each. If you cannot find one, your feedback loop is not healthy, it is quiet. A quiet loop with rising activity is the exact signature of suppression, and AI makes it cheaper to produce than ever.

---

This analysis synthesizes Feedback Loops Require Real Feedback. AI Drives It Underground. (Product Picnic, June 2026).

Victorino Group helps teams keep feedback tied to decisions as AI floods the channel. Let’s talk.