Four AIs Agreed. We Still Hadn't Verified Anything.

THE FIRST QUESTION

We started with a private company.

A real one.

It had a website. A history. Documented projects. Public profiles. Articles. Services.

Enough information scattered across the internet to form an opinion.

It did not publicly disclose its revenue. It did not publish EBITDA. There was no public valuation.

We asked an AI system a simple question.

What can you tell me about this company?

The answer was good. Surprisingly good.

It understood what the company did. It identified patterns across its work. It connected the founder’s background to the way the company described its services.

Some conclusions were simplified. A few details were wrong. But the overall picture was coherent.

So we kept asking. Is this a serious company? What is unusual about it? Who are its competitors? Does it appear to have real clients? Where is the business going?

Again, the answers were interesting.

Then we asked a more dangerous question.

How much revenue might this company generate?

The system gave us a number. We continued. Eventually, it gave us a valuation range.

The later answers became surprisingly expensive.

THE FIRST NUMBER

We went back through the reasoning.

The company described its organization using language similar to: more than two dozen specialists, collaborators and partners.

The phrase described an operating network.

The AI read it differently.

Employees.

That was the first number. Except it was not really a number. It was an interpretation.

From there, the system found an industry benchmark for revenue per employee. The arithmetic was simple. Estimated headcount multiplied by estimated revenue per employee produced estimated annual revenue.

Then came an assumed margin. Revenue became estimated EBITDA. Then a market multiple. EBITDA became valuation.

The final answer looked sophisticated. There were ranges. Industry terminology. Caveats. The reasoning was presented step by step.

And most of the arithmetic was perfectly reasonable.

There was only one problem.

The first number was never calculated. It was interpreted.

Everything after it was calculated.

That distinction is easy to miss.

THE ARITHMETIC MADE THE ASSUMPTION LOOK AUDITED

Imagine the public signal suggests a network of approximately twenty-five specialists and collaborators.

The model interprets this as twenty-five employees. It applies an estimated $180,000 in annual revenue per employee. Now the company appears to generate $4.5 million in revenue.

Assume a 22% EBITDA margin. We have $990,000 in EBITDA. Apply a 5x multiple. The company is suddenly worth approximately $5 million.

The figures in this example have been changed. The sequence of reasoning has not.

Look at the chain again: public language → headcount interpretation → revenue benchmark → revenue estimate → margin assumption → EBITDA estimate → market multiple → valuation.

By the time we reach the bottom, the answer looks financial. The uncertainty looks quantified. The valuation looks calculated.

But the entire structure may still be standing on the interpretation of one sentence.

The arithmetic did not remove the uncertainty. It buried it under increasingly professional mathematics.

SO WE CHECKED THE AI

We suspected the estimate was weak. So we did what people increasingly do. We asked another AI. Then another. Then another.

We tested the same company across Google AI Mode, ChatGPT, Claude and Perplexity.

The systems disagreed. One was aggressive. Another was cautious. One reduced the implied scale of the company. Another questioned parts of the financial estimate.

Good. This felt like verification.

Four systems. Four companies behind them. Different interfaces. Different answers.

And yet the general direction remained surprisingly similar.

The company appeared larger than the underlying facts justified. Its financial scale was repeatedly inferred upward. Its organizational sophistication became a signal of economic scale.

The numbers moved. The story survived.

For a moment, that made the story feel more credible.

Then we realized what we had actually done.

FOUR MODELS ARE NOT FOUR SOURCES

We had not found four independent pieces of evidence.

We had asked four systems to interpret largely the same public information.

That is not the same thing.

We confused model plurality with evidence plurality.

Researchers are already looking at a related problem. A July 2026 paper asked a remarkably direct question: When LLMs Agree, Are They Right?

The uncomfortable answer is: not necessarily.

Agreement can be useful. But agreement itself is not accuracy. Models may converge because of shared biases, memorized heuristics or other common tendencies rather than truth.

Another large-scale study examined more than 350 language models. On one leaderboard dataset, models agreed roughly 60% of the time when both were wrong.

Think about that for a second. They agreed while being wrong. Not one strange model. Not one bad answer. Correlated errors across hundreds of systems.

Our observation was much smaller. One company. Four systems. Evolving conversations rather than controlled prompts. This was not a scientific study. And we are not pretending it was.

But it made us notice something one step further down the chain.

What does a human do when AI agreement becomes a personal verification workflow?

ASK ANOTHER AI

A person receives an AI answer. They are uncertain.

Ten years ago, they might have searched for the original source. Opened a financial filing. Found the company registry. Read an industry report. Called someone.

Today there is an easier option. Open another AI.

Ask: Does this estimate make sense?

The second system examines the idea. It finds similar market benchmarks. It recognizes the same company signals. It may even disagree with the first estimate.

Perhaps the first AI says approximately $5 million. The second thinks $3 million to $4 million is more realistic. A third places the company somewhere in the low single-digit millions.

Different numbers. Different systems. The human sees a range. The range feels like consensus.

But what if all three systems began with the same ambiguous signal?

We did not independently verify the answer. We repeated the same epistemic act several times.

THE RANGE MADE IT FEEL SAFER

One system gave a higher estimate. Another lowered it. A third was more cautious.

Strangely, the disagreement increased our confidence.

A range feels more honest than a point estimate. It acknowledges uncertainty. It looks like someone considered the edges.

But a wider range does not repair a bad variable.

If the first assumption is wrong, $3 million to $6 million may not be more responsible than $5 million. It may simply be a more sophisticated expression of the same mistake.

Uncertainty around an estimate is not the same as uncertainty about the premise behind the estimate.

The models disagreed enough to look independent. But not enough to force us back to the original signal.

That may be the dangerous range. Not perfect agreement. Plausible disagreement around a shared premise.

WE CARRY THE STORY WITH US

There is another problem. Humans rarely move information between AI systems neutrally.

We carry the previous answer with us. Sometimes literally.

Copy. Paste. What do you think?

Now the second system is no longer evaluating only the company. It is evaluating the company and a coherent narrative about the company.

That narrative already contains selected facts. Connections. Interpretations. Benchmarks. Perhaps even a valuation methodology.

The second AI may challenge it. But it is now reasoning inside a frame created by the first.

We do this constantly. I do it. One model writes something. I read it. I think. I move the answer to another model. I ask it to critique the logic. Then I take the critique back.

The work often gets better. Much better. Weak arguments disappear. Obvious errors are corrected. The language becomes cleaner. The structure becomes stronger.

The final work may be genuinely good.

But there is a question I had not really asked before. At what point did we stop verifying the original assumption and start optimizing the explanation built around it?

Those are not the same task.

A BETTER STORY CAN STILL BE WRONG

This is where the problem becomes more interesting than hallucination.

A hallucination is easy to understand. The system invents a company. A person never held that position. A study was never published. Wrong. Find the source. Done.

But that is not what happened here.

The company was real. The projects were real. The people were real. The public signals were real. The industry benchmarks could be real. The EBITDA multiple could be reasonable.

The problem lived between the facts. It lived in the transitions.

collaborators became employees which became organizational scale which became revenue capacity which became financial performance which became valuation.

No absurd fact appeared. Nothing looked obviously broken.

In fact, the final answer became more convincing because so many reasonable steps separated it from the original assumption.

Researchers have documented a related problem with what they call self-consistent errors — cases where models repeatedly produce semantically similar wrong answers. These errors can be particularly difficult for common detection methods to identify.

For a human, the mechanism can be even simpler.

A bad assumption does not always look worse after more reasoning. Sometimes it looks better.

A bad assumption can become more persuasive as the quality of the reasoning around it improves.

THE WORK TRUCK

Imagine two people arriving at the same job site.

One drives a new heavy-duty truck worth more than $100,000. The other arrives in an old SUV worth perhaps a tenth of that.

The truck is objectively more capable. It can tow more. Carry more. It was built for work.

Now suppose both people carry the same tools to the same site every morning. The tools fit in the SUV with the rear seats folded.

Both people complete the work.

For that specific task, purchase price tells us very little about the economic value each vehicle contributed to the output.

This does not make the SUV a better heavy-duty truck. It means we asked the wrong question.

Visible capacity is not utilized capacity. And the price of an asset is not the same as its contribution to output.

We understand this fairly easily with vehicles. We are much worse at seeing it in companies.

WE VALUE THE TRUCK

A company has fifty employees. Another has ten.

The first raised $20 million. The second is bootstrapped.

The first has an executive team. The second has a founder and a laptop.

The first publishes research every week. The second suddenly begins producing work at a similar intellectual density.

Historically, these signals told us something. A large, coherent public information footprint was expensive to produce. Research required researchers. Publishing required writers and editors. Brand consistency required a team. Strategic analysis required consultants. Technical prototypes required developers.

The output gave us clues about the machine behind it. We learned to estimate the machine by looking at the exhaust.

Then the engine changed.

THE COST OF EXPOSING THOUGHT COLLAPSED

AI is usually discussed as a productivity tool. One person can do more. A small team can move faster. Companies may need fewer people for certain tasks. All true.

But there is another consequence.

For years, a person can accumulate experience without documenting much of it. Hundreds of decisions. Failed ideas. Client conversations. Patterns. Things they notice but never write down.

The value exists. It is simply poorly indexed.

Then AI enters the process.

The person starts talking. The model organizes. The person reads it.

No. That is not what I meant. Again. Better. Still wrong. Again. Now this part is right. Move it. Cut that. You made me sound smarter than the thought actually is. Put the uncertainty back. Again. Again. Again.

Eventually, something strange happens.

Ten or twenty years of poorly indexed experience begin appearing as structured public information. Articles. Cases. Frameworks. Connections between previously isolated parts of the person’s work.

Did the person suddenly acquire twenty years of experience? No.

Did the company hire a research department? No.

Did the underlying knowledge appear yesterday? No.

The cost of exposing it collapsed. That is different from the cost of creating it.

And I am not sure our systems — human or artificial — are good at separating the two yet.

THE COMPANY DID NOT GET FIFTY TIMES BIGGER

There is already evidence that generative AI can materially improve productivity in certain kinds of work. That matters.

But productivity is only one side of what happened. The other side is perception.

A research footprint that once suggested a team may now come from one analyst. A publishing operation may be one founder. A coherent strategic narrative may emerge from hundreds of hours of human experience reorganized with a general-purpose AI model on an ordinary laptop.

The output may be real. The thinking underneath it may be real. The experience may have taken decades to accumulate.

But our historical assumptions about what it costs to expose that experience to the world may no longer be real.

This creates a strange problem.

AI systems observe output. Humans observe output. Investors observe output. Customers observe output.

We have spent decades learning to infer organizational capacity from visible signals.

But what happens when the production cost of those signals changes faster than our heuristics?

IT SEES THE OUTPUT AND IMAGINES THE TRUCK

Imagine a small company with ten years of accumulated experience.

Its knowledge is fragmented. Some of it lives in emails. Some in old projects. Some in the founder’s head. Some in conversations that happened at midnight and were never documented.

Then the company begins using AI well. Not to invent expertise. To extract it. To challenge it. To organize it. To publish it.

The public information footprint expands dramatically.

Now an AI system looks at the company. What historical examples does it have? What kind of organization used to produce this volume of coherent output? How many people did that usually require? How much revenue did companies of that apparent scale usually generate?

The model may be looking at a small system producing signals historically associated with a much more capital-intensive one.

It sees the output. And imagines the truck.

HUMANS DO IT TOO

We see an expensive office and infer stability. We see headcount and infer scale. We see funding and infer success. We see publishing volume and infer a content department. We see a polished strategy and infer expensive consultants. We see three AI systems produce similar conclusions and infer independent confirmation.

We use proxies because we have to. Nobody can fully investigate everything.

The problem begins when we forget that a proxy is a proxy.

Revenue is real. It is not profit. Headcount is real. It is not output. Funding is real. It is not value creation.

A $100,000 truck is real. Its purchase price does not tell us how much economic value it added to the work performed today.

An AI-generated valuation is real in the sense that the model genuinely produced it. That does not tell us whether the first variable in its reasoning was ever verified.

NIST has warned about automation bias and excessive deference to automated systems.

But we may now be creating a strange variation of that problem. We distrust the first automated answer. So we ask another automated system. Our skepticism remains. The verification layer has simply become automated too.

THE BUBBLE MAY BEGIN BEFORE THE SPREADSHEET

Eventually, serious transactions reach real numbers. P&L. Contracts. Cash flow. Customer concentration. Debt. Adjusted EBITDA. Technical due diligence.

Nobody should buy a company because four AI systems liked its website.

But that misses the earlier part of the process.

Before diligence, there is attention. Before the spreadsheet, there is interest. Before a buyer requests documents, someone decides the company may be worth looking at. Before an advisor builds a model, a narrative begins to form.

AI is increasingly present in that layer. It researches. Summarizes. Compares. Finds competitors. Explains positioning. Builds scenarios. Challenges assumptions.

And then another AI checks the work.

The danger is not necessarily that AI will fabricate the final purchase price.

The danger may be quieter. AI can influence which stories arrive at the spreadsheet already feeling plausible.

A weak company can be made to look coherent. A strong but poorly documented company can remain invisible.

A soft public signal can become a headcount estimate. The estimate can become revenue. Revenue can become EBITDA. EBITDA can become a valuation range.

And a valuation range, repeated often enough, can become an anchor.

Not a fact. An anchor. Those are different things.

EVERYBODY ALREADY KNEW PEOPLE LIE

Companies have always managed perception. Founders exaggerate. Pitch decks select convenient numbers. Advisors build favorable narratives. Sellers know which quarter to show. Buyers know which weakness to emphasize.

None of this began with AI. We already knew people lie. We built entire professions around checking them. Auditors. Analysts. Lawyers. Investigators. Due diligence teams.

The interesting change is not that AI can be wrong. Of course it can.

The interesting change is that AI can take real information, make a reasonable interpretation, build mathematically coherent consequences from that interpretation and then have another AI partially validate the same structure.

No conspiracy. No malicious system. No invented company.

Just a chain of increasingly defensible reasoning around a premise nobody returned to verify.

That may be harder to catch than a lie. Because nobody had to lie.

WE ASKED FOUR SYSTEMS

Our experiment proves almost nothing on its own. One company is not a dataset. Four systems are not the AI industry. The prompts evolved during the conversations. The models may behave differently next month.

A proper study would need controlled prompts. A larger sample of companies. Known financial data. Repeated trials. A way to measure public information density.

It would need to ask a much more precise question: Do general-purpose AI systems systematically overestimate organizational scale when small companies produce public information footprints historically associated with larger organizations?

I don’t know. But I think that question is now worth asking.

Researchers are already showing that model agreement is not a simple proxy for truth and that errors can be correlated.

Our question begins one step later. What happens when ordinary people turn cross-model comparison into a verification habit? Especially when we move not only a question, but a narrative, from one system to another?

I don’t know that either.

But I know what happened to us. Different systems challenged different details. Some were cautious. Some were not. The exact numbers moved. The interpretation remained surprisingly stable.

And every time another system produced something in roughly the same direction, our confidence increased faster than the amount of independent evidence.

That is the part I cannot stop thinking about.

HOW MANY TIMES?

We started by asking AI what a private company might be worth.

The answer was probably wrong. So we checked. And checked again.

By the fourth system, the exact number mattered less. The story had become familiar. Reasonable. Defensible.

We almost felt that we had verified something.

We had not.

We had asked several extraordinarily capable systems to reason over similar information. There is enormous value in that. There is also a trap.

A repeated interpretation is still an interpretation. Even when the arithmetic is correct. Even when the language is professional. Even when another AI agrees.

We started with a question about valuation. We ended with a different question.

How many times does an estimate need to be repeated before a human stops asking where the first number came from?