Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

Every AI gives you the same answer—that should worry you

We need to start building systems that verify what is actually true.

[Photo: Getty Images]

Profile photo of Michael Quoc

Founder & CEO · Product.ai

If you searched for a product review in 2010, you found a messy sprawl of forums, blogs, and personal websites. It was inefficient, but the disagreement was the point. Five different people with five different opinions gave you something to work with. You could triangulate.

Today, that disagreement is disappearing. Ask AI for a recommendation and you get a single, polished, authoritative answer, with no indication of how much conflict it smoothed over to get there. I call this the beige singularity: the moment when the full spectrum of human knowledge collapses into one confident, often incorrect answer.

We are caught in a feedback loop. As large language models flood the web with synthetic content, they are poisoning their own future training data. When an AI learns from content generated by another AI, the output becomes a Xerox of a Xerox. The edges get smoothed. The nuance disappears. What remains is information optimized for statistical probability rather than truth.

THE POLLUTION LAYERS

The web was already in trouble before AI arrived. Three structural forces had been degrading information quality for years.

The first is the SEO economy. Google’s algorithm created a market that rewards ranking, not accuracy. A content farm can publish thousands of articles about a product category without testing a single product. Same five recommendations. Same recycled copy. New headline.

The second is what I think of as marketing hallucination. Every product page claims “precision engineered” and “professional-grade.” This is not dishonesty. It is a structural feature of commercial speech. The problem is that AI cannot distinguish a marketing claim from an empirical one. An LLM reads “commercial-grade” on a spec sheet and treats it as fact.

The third is affiliate corruption. Every “best of” list is a monetization vehicle. The author gets paid when you click, not when you keep the product. High-commission products rise. Low-commission ones vanish. The AI reads these lists without knowing the rankings are bought. It treats “best” as an empirical claim. It is not. It is a business model.

These forces created a web where noise is the majority and signal is the exception. Then we pointed AI at it and asked it to summarize.

THE CONFIDENCE PROBLEM

We have moved from the search engine, which offered a list of possible sources, to the AI answer, which offers a definitive statement. That shift feels like progress. It is faster to read one paragraph than to click through five websites.

But speed is not accuracy. An LLM does not operate on “truth.” It operates on plausibility. It constructs a response that sounds correct based on its training data. If that training data is polluted with SEO content, marketing copy, affiliate rankings, and other AI-generated errors, the model will reproduce those errors with total confidence. That is the protocol now.

That confidence changes behavior. When a machine gives you a single, assured response, the friction required to verify it feels disproportionate. You accept the answer. Or you spend hours doing your own research, opening tabs, reading forums, building a mental spreadsheet. The internet promised to make decisions easier. For a lot of people, it has made them a part-time job.

The AI is confident. The user is anxious. This is exactly backwards.

THE COMPOUNDING ERROR

AI-generated summaries are now going back onto the web. Someone posts a ChatGPT answer on a forum. The next generation of models trains on a web that includes the previous generation’s errors.

The errors do not cancel out. They compound. A hallucination from 2024 becomes a training datapoint in 2025. By 2026, it has been repeated so many times it passes for consensus.

Variance is signal. When sources disagree, the disagreement tells you something real about edge cases, failure modes, and use patterns. Summarization kills that variance. It converges on the median. The distinctive voice of the frustrated user who documented a real problem gets compressed into “some users report issues.” Run this loop enough times and the specific warning inverts into a generic endorsement.

This is the beige singularity in its final form. A closed system of recycled ideas. An internet that has finished eating its own tail.

HOW WE BUILD OUR WAY OUT

We cannot turn off the AI. The efficiency gains are real and the adoption is irreversible. But we can change the architecture.

The core issue is that current AI systems are built to summarize. When the sources are polluted, the summary is polluted. Better models will not fix this. They will make the pollution more fluent.

The alternative is verification. Instead of systems that average claims, we need systems that test them. Collide the marketing page against the user review. Collide the spec sheet against the real-world performance data. The output should not be a summary. It should be a verdict, with the evidence visible and the reasoning auditable.

First, we need to value primary sources over recycled content. The user who bought a product, lived with it for a year, and documented a specific failure has produced more signal than a thousand SEO articles that rewrote the spec sheet. AI systems should weight sources by epistemic quality, meaning how close the source is to direct experience, not by volume or search ranking.

Second, we need transparency you can verify. Every recommendation system claims to be unbiased. Almost none can demonstrate it. We should demand systems where the sources are traceable, the conflicts between sources are visible, and the reasoning behind a recommendation is clear enough to audit. That conflict is the information. Hiding it is the problem.

The current trajectory, where every AI converges on the same answer because every AI trained on the same polluted data, is a future where trust erodes and decisions get worse.

The internet does not have to be beige. But it will be, unless we stop building systems that average the noise and start building systems that verify what is actually true.

Michael Quoc is founder and CEO of Product.ai.