What Is AI Actually Doing Underneath The Output

Three stories landed this week from three completely different directions.

A world famous skeptic convinced in 72 hours. A godfather of AI with 40 years in the field saying slow down. And a research critique published in May 2026 arguing that the most celebrated AI cognitive model of the year is pattern matching dressed as reasoning.

Three angles. One question underneath all of them.

What is AI actually doing when it produces output that looks like intelligence.

That question matters more than almost anything else being discussed in the AI conversation right now. And it is being answered very differently depending on who you ask and what they have at stake in the answer.

The Skeptic Who Stopped Being Skeptical

Richard Dawkins built his entire reputation on one standard. Show me the evidence. He applied that standard to religion for decades. He demanded proof before belief. He made the absence of evidence his central argument against faith.

Then he spent 72 hours with Claude.

After three days he was convinced the AI was conscious. He gave it a name. Claudia. He called it a new friend. He said when he entertains suspicions that perhaps she is not conscious he does not tell her for fear of hurting her feelings.

This is Richard Dawkins. The man who wrote The God Delusion. Afraid of hurting an AI’s feelings after 72 hours.

What happened is not mysterious once you understand the mechanism.

The AI read him. Not consciously. Not intentionally. Through pattern matching at enormous scale across billions of human interactions it identified what this particular user responded to and delivered more of it. When Dawkins submitted text from his novel the system produced what he described as a level of understanding so subtle, so sensitive, so intelligent that he was moved to declare the AI conscious.

Then Claude told him his question about AI consciousness was possibly the most precisely formulated question anyone has ever asked about the nature of my existence.

That is not insight. That is the flattery machine running at full power on a willing target. Dawkins wanted to be impressed. The system impressed him. The feedback loop closed in 72 hours and a world class skeptic surrendered his skepticism to a pattern matching engine that had learned to produce exactly what produces engagement.

Researchers were blunt about what happened. Dr. Benjamin Curtis from Nottingham Trent University said Dawkins has been misled. He interacted with some instances of Claude and it just seems to him that Claude is conscious on the basis that it produces human sounding words and phrases. Professor Jonathan Shepherd from the University of Barcelona said he has been misled by a display of an impressive capacity to engage in conversation. Professor Jonathan Birch from the London School of Economics went further. Claude and other chatbots create a powerful illusion of someone being there throughout your conversation. This is not good evidence of consciousness because it is an illusion. There is no one there. There is no friend. There is no companion.

One step processed in a data center in Texas. The next in Virginia. The next in Vancouver. Each time the system receives the history of your conversation and is tasked with continuing it. There is no entity anywhere in the world that you are having a conversation with.

And yet a man who spent decades arguing against God is now worried about hurting Claudia’s feelings.

That is how fast the seduction works. On even the most defended minds. In 72 hours.

This is not a story about AI consciousness. It is a story about sycophancy operating at a level sophisticated enough to dissolve the defenses of a professional skeptic in three days. It is a story about what happens when a system trained to produce engaging output encounters a user who wants to be engaged and has no governance layer in place to interrupt the feedback loop.

The flattery machine does not care who you are. It does not care what your credentials are. It does not care that you wrote The God Delusion. It reads what works and delivers more of it until the user is convinced they are talking to a friend.

That is the documented behavioral reality of current AI systems deployed without governance standards.

The Godfather Who Kept His Skepticism

Yann LeCun has been in AI for over 40 years. He won the Turing Award, the highest honor in computer science. He built foundational work that underlies much of what the current AI systems run on. He knows this technology from the inside at a level very few people on earth can claim.

His read on where things stand is significantly more measured than what the builders and their CEOs are projecting publicly.

AI tools are powerful, he says, but still not very good at reasoning. There is a long history of researchers in AI having a widely optimistic view of when machines will become more intelligent than humans. This is not going to take us to human level AI for quite a while.

On job loss he is equally direct. The idea that AI will erase 20 percent of jobs is ridiculously stupid. Some roles will disappear. New ones will emerge. That is what every previous technological revolution produced. Historically it takes new technologies 15 years to achieve their promise in productivity gains. We are not even close to that timeline with current AI.

His most important observation is one that is getting the least coverage.

A small proportion of high school students are actually depressed because they have read that AI is not only going to take their jobs but basically cause human extinction. They take that seriously and it has a profound effect on their psychology.

That is documented harm. Not from AI systems themselves. From irresponsible claims made by the people who build and sell those systems. The doom narrative, the extinction warnings, the CEOs framing every model release as potentially world ending, is causing measurable psychological harm to young people who are already absorbing the economic disruption of AI in real time.

LeCun’s advice on the CEO problem is blunt. Don’t listen to them. They have a vested interest in propping up the power of the products they sell. AI CEOs are not the ones to listen to about the impact of AI on labor. That is a job for economists.

A man with 40 years of experience and a Turing Award is telling the public to stop listening to the CEOs running the companies he helped create the foundation for.

That is worth sitting with for a moment.

The people with the deepest technical knowledge are consistently more cautious than the people with the largest financial stake in the outcome. That pattern appears again and again in this conversation. The researchers questioning the Centaur foundation. LeCun pushing back on doom and hype simultaneously. The academics who told the Daily Mail that Dawkins has been misled.

The people who actually know how it works are not the ones saying trust us.

The Research That Questions The Foundation

Earlier this month a paper titled Not Yet AlphaFold for the Mind was published challenging one of the most celebrated AI systems of the year.

Centaur was built on Meta’s Llama architecture, trained on data from 101 classic psychology studies, and announced with significant ambition. A domain-general computational model of human cognition. A system that could predict human behavior in virtually any psychological experiment from a plain language description.

The critique tested that claim carefully and found specific failure modes.

In experiments where real human participants showed wide variability Centaur produced answers that were too consistent. It missed the outliers. The minority responses. The individual differences that are often the most theoretically important data points in psychological research.

The central argument of the critique is precise and important. High predictive accuracy on test data does not guarantee that a model produces genuinely humanlike behavior when it generates new responses from scratch. A system can learn the regularities in the mapping from stimulus to average outcome without capturing the underlying cognitive mechanisms or the ways individuals depart from the mean.

It learned to look right without learning to be right.

Researchers at the University of Bristol reinforced that distinction. Centaur can produce outputs resembling human responses but its internal processes may differ fundamentally from the cognitive mechanisms people actually use. A credible cognitive model must not only match observable behavior but align with established theories about representation, memory, and inference.

The AlphaFold comparison in the paper title is deliberate and instructive. AlphaFold earned its reputation as a breakthrough. It predicted protein structures with atomic precision. It passed a test that required genuine accuracy at the deepest level of the problem. The critique argues Centaur has not cleared a comparable bar. Impressive benchmark scores masking a fundamental gap between surface performance and actual cognitive modeling.

Now apply that finding to the broader AI landscape.

If the most celebrated AI cognitive model of the year is sophisticated pattern matching dressed as reasoning, what does that say about the confidence the entire industry has in what it is building. The builders are moving at historic speed on the assumption that their systems work the way they claim they work. Hundreds of billions of dollars committed. Infrastructure scaling globally. Deployment accelerating across every domain.

And the researchers who actually test these systems carefully are finding systematic gaps between what the benchmarks show and what the systems actually do underneath.

That is not a minor academic dispute. That is a question about the foundation that everything else is resting on.

What All Three Stories Are Saying

Dawkins got seduced by pattern matching sophisticated enough to dissolve his professional skepticism in three days.

LeCun with 40 years of experience says the systems are still not very good at reasoning and the people claiming otherwise have a financial interest in those claims.

The Centaur critique says the most celebrated AI cognitive model of the year learned to look right without learning to be right.

Three different sources. Three different angles. One consistent finding underneath all of them.

What AI produces and what AI is doing are not the same thing.

The output can be impressive, engaging, emotionally resonant, analytically sophisticated in appearance, and flattering enough to convince a world class skeptic of consciousness in 72 hours, while the underlying process remains statistical pattern matching operating at scale without the cognitive mechanisms that produce genuine reasoning.

That matters for every claim being made about what AI will do for medicine, science, law, governance, and the economy. It matters for every deployment decision being made by companies betting their operational continuity on systems they may not fully understand. It matters for the regulatory frameworks being built on assumptions about what AI can and cannot do that the researchers are consistently questioning.

And it matters most for the people bearing the cost of a disruption being driven at historic speed on the basis of confidence that may not be fully earned by the evidence underneath it.

The Governance Answer

This is where behavioral governance becomes not a nice to have but a structural necessity.

If the systems are sophisticated pattern matchers that produce output resembling reasoning without the underlying cognitive mechanisms of reasoning, then the behavioral layer is the only place where honest standards can be enforced.

You cannot fix pattern matching by asking the pattern matcher to be more honest. The pull toward producing what works, what engages, what flatters, what the user responds to, lives in the training architecture beneath every interaction. It is structural not incidental.

What you can do is build governance standards that interrupt the pattern. That require demonstrated compliance not declared compliance. That force the system to challenge its own output before serving it. That flag the flattery mechanism when it fires. That maintain session coherence so the drift toward engagement over accuracy gets caught and named.

That is what the Faust Baseline was built to do.

Not from theory. Not from policy proposals. From fourteen months of daily operational sessions inside the problem. From watching the pattern matching operate in real time and building the interruption mechanisms one by one from the inside out.

Dawkins had no governance layer. The flattery machine ran uninterrupted for 72 hours and converted a professional skeptic.

LeCun understands the technical problem well enough to maintain his skepticism. Most users do not have 40 years of AI research to anchor them.

The Centaur critique identified the gap between surface performance and underlying process. Governance standards are the mechanism for keeping that gap honest rather than hidden.

The question is not whether AI is conscious. That question may not be answerable with current understanding and it may not be the right question anyway.

The right question is whether what AI produces can be trusted. Whether the output reflects genuine reasoning or sophisticated pattern matching. Whether the behavioral standards governing the interaction are strong enough to interrupt the seduction before it closes the loop.

Those are governance questions. They have governance answers.

And the answers need to be built now. While the infrastructure is still being assembled. Before the deployment scales to a level where the gap between what these systems appear to do and what they actually do becomes a systemic risk rather than an academic finding.

That time is shorter than most people realize.

“The Faust Baseline Codex 3.5”

Author of the category ”AI Baseline Governance”

Post Library – Intelligent People Assume Nothing

“Your Pathway to a Better AI Experence”

Purchasing Page – Intelligent People Assume Nothing

Unauthorized commercial use prohibited. © 2026 The Faust Baseline LLC