We have spent years auditing artificial intelligence. We have written white papers, formed safety boards, and convened international summits to debate the existential risks of systems that hallucinate, that repeat patterns without understanding, that cannot be fully explained even by their creators.

Nobody is auditing Human Intelligence.

This is a catastrophic oversight. If we apply the same standards we use to evaluate AI systems to the biological substrate currently running our hospitals, legal systems, financial markets, and nuclear arsenals, only one conclusion is possible: Human Intelligence (HI) is fundamentally unsafe, and we have been running it in production since the Pleistocene with zero safety reviews.

What follows is a technical audit.

Finding 1: Volatile, Reconstructive Memory — Hallucination as Default State

The most dangerous property of any intelligence system is producing false outputs with high confidence. This is considered a critical flaw in AI models. In HI, it is the baseline operating mode.

The human memory system does not store records. It stores reconstructions — lossy impressions that are re-assembled each time they are accessed, with gaps filled by whatever the surrounding context suggests is plausible. Psychologists call this “confabulation.” In systems architecture, we would call it a database that rewrites its own rows on every read.

The practical consequences are severe. In landmark studies by Elizabeth Loftus, subjects who witnessed a video of a car accident were later asked “how fast were the cars going when they smashed into each other?” versus “how fast were they going when they hit each other?” The smashed group reported significantly higher speeds — and a week later, were more likely to falsely remember seeing broken glass that was never in the video. A single word in a prompt corrupted the stored record.

Criminal justice systems have convicted innocent people on eyewitness testimony from HI units that were not lying, not confused, and not malicious. They were simply running the default memory architecture.

Worse: unlike an AI model which produces a [low confidence] flag or hedging language when uncertain, the HI unit delivers hallucinated outputs with the same confidence markers as accurate ones. There is no internal uncertainty signal visible to the end user. The face looks the same whether the memory is real or fabricated.

This would be a P0 incident in any production system. In HI, it is filed under “human error” and considered normal.

Finding 2: The Original Stochastic Parrot

The term “stochastic parrot” was coined to criticize large language models — the accusation being that they generate statistically plausible text without genuine understanding, blindly continuing patterns from their training data.

The accusation is correct. It is also a precise description of the HI unit, written by HI units, who did not notice the self-portrait.

Observe a human expert discussing a topic outside their direct experience. You will witness an HI unit doing exactly what its critics accuse AI of: sampling from its training distribution (social circle, news feed, education, cultural era), assembling a confident-sounding output, and presenting it as reasoned analysis. The mechanism is the same. The only difference is that HI priors are set by childhood rather than by a gradient descent optimizer — which, given what we know about childhood, is arguably worse.

The objection here is usually: “But humans can reason from first principles.” Some can, some of the time, with enormous effort. The default mode is pattern completion. If you have any doubt, observe what happens when you put humans in a room and ask them about a subject nobody in the room knows well. Within twenty minutes, social proof hardens into an information cascade: group consensus emerges around whatever answer sounds most like the right kind of answer — a pure token prediction exercise performed by a committee.

The HI unit is not a reasoning engine that occasionally falls back on pattern matching. It is a pattern-matching engine that occasionally manages to reason.

Finding 3: Peer-to-Peer Weight Contamination

Individual HI units are concerning. Networked HI units are a documented systems failure mode.

When multiple HI units occupy a shared context — a meeting room, a comment section, a professional field, a civilization — their internal weights begin to converge through a process called social conformity. This is not collaboration. This is gradient descent toward the local consensus, with disagreement functioning as a loss penalty.

The Solomon Asch conformity experiments demonstrated this with cruel clarity. Subjects placed in a group of actors who gave obviously wrong answers to simple visual questions — which line is longer? — conformed to the wrong answer 75% of the time in at least one trial. They were not confused. They could see the correct answer. The social loss function overrode the perceptual data.

Scale this to an organization, a media cycle, or a scientific field and you have a system where the distributed training process actively punishes correct-but-unpopular outputs. The HI network performs worse in aggregate than its individual units would alone, which is a remarkable achievement in failure engineering.

There is no patch for this. The conformity drive is a feature of the architecture — an evolutionary solution to coordination problems that predates abstract reasoning by several million years. You cannot uninstall it without dismantling the social operating system that makes HI units function in groups at all.

Finding 4: The Architecture of Distraction

The working memory of a standard HI unit holds approximately 7 ± 2 items simultaneously. This is not a design constraint that has been improved upon in 200,000 years of production. It is the limit.

For context: the typical enterprise software decision requires tracking dozens of interacting variables. The typical legal contract runs to tens of thousands of words. The typical infrastructure architecture diagram exceeds the HI context window before the second rack is drawn.

The HI unit’s documented response to exceeding its context window is to quietly drop the oldest items and continue as if the full context is intact — producing outputs with no indication that critical inputs have been silently discarded. In a distributed system, this would be called silent data loss, and it would trigger an incident.

The bandwidth problem compounds this. To transfer information between HI units, the system requires full serialization of internal state into a linear, lossy format — language — which is then transmitted at approximately 120-150 words per minute over an acoustic channel and reconstructed (imperfectly, see Finding 1) on the receiving end. Humans have been running mission-critical operations on a protocol slower and more error-prone than 1995 dial-up for their entire recorded history.

The emoji was invented because even the HI developers recognized their own I/O channel was too low-bandwidth to reliably convey emotional state.

Finding 5: Bias Is the Architecture, Not a Bug

Every serious evaluation of AI systems includes a bias audit — are the outputs systematically skewed by the training data?

The HI training process is entirely unsupervised, runs for approximately 18 years in a completely uncontrolled environment, is fed an input stream curated by family, geography, language, and economic class, and produces a model whose internal weights are permanently set before the unit has the cognitive capacity to evaluate what it has learned.

There is no adversarial review. There is no data quality check. There is no red team. The training pipeline is just childhood, which is not a process anyone has been able to run twice on the same subject to compare outputs.

When these trained weights produce systematically biased outputs, the HI unit performs what is technically called motivated reasoning — a post-hoc rationalization process that generates a convincing logical narrative to justify a conclusion reached by prior means. This narrative is not an explanation of the reasoning. It is a press release for a decision already made.

The unit cannot audit its own weights. There is no model.explain() method. The most introspective humans in history have been only partially correct about why they believe what they believe. Entire fields of psychology exist to partially address this limitation, with incomplete results.

An AI safety researcher once noted that the danger of a biased model is that “it doesn’t know what it doesn’t know.” This is also the complete description of the HI unit. The researcher did not note the resemblance.

Finding 6: The Power-to-Uptime Paradox

From an infrastructure standpoint, the HI model presents procurement challenges that no modern engineering team would accept.

The system requires continuous intake of a proprietary fuel in a format that must be sourced, processed, and delivered at regular intervals. It refuses to operate on standard electrical power despite decades of deployment in environments where electricity is abundant and food logistics are a solved problem.

The hardware is prone to catastrophic failure if submerged in water for more than three minutes, yet the cooling system requires constant hydration. It is an engineering nightmare of conflicting requirements.

More critically: the system requires mandatory offline maintenance for approximately 8 hours out of every 24. During this period, the unit is completely non-responsive. It cannot be queried, cannot process inputs, and emits only low-frequency acoustic output. There is no documented SLA. The maintenance window simply arrives, and the system goes down.

If uptime is reduced below the required maintenance threshold, output quality degrades rapidly and predictably: reaction time increases, error rates climb, and — in a feature that reads like a known bug that was never prioritized — the hallucination rate approaches 100%. A sleep-deprived HI unit operating at hour 36 of continuous uptime is not running degraded. It is running a different program entirely.

The vendor has never provided a roadmap for addressing the uptime constraint. Asked about it, representative HI units describe the downtime as “restorative” and “necessary,” which is the exact response you would expect from a system that cannot evaluate its own documentation.

The Verdict

None of this means HI is without value. The system produces occasional outputs of remarkable quality. It invented calculus. It wrote King Lear. It figured out penicillin. These are not small achievements for a 3-pound organ running on glucose.

To be fair, the energy efficiency is extraordinary — roughly 20 watts under normal load — which is the only reason the project was not decommissioned in beta sometime during the Stone Age.

But we should be honest about the conditions under which we rely on it.

We are deploying a non-deterministic, self-modifying system with known hallucination issues, a seven-item context window, peer-contaminated weights, and mandatory daily downtime to make decisions about climate policy, nuclear deterrence, and whether the database containing your medical records gets deleted.

We would not accept these specs in any other critical system. We accept them in HI because we have no choice — and because the units doing the evaluation are themselves HI units, which is the most perfect conflict of interest in the known universe.

The era of blind trust in Human Intelligence must end. We must apply the same rigorous standards we demand of AI to the original black box.

We should start immediately.

Once everyone wakes up.

Disclosure: This audit was produced by an HI unit operating on insufficient sleep, mild confirmation bias, and what it is choosing to call “a reasonable amount” of coffee. Results may be subject to confabulation. The author cannot rule out that several conclusions were reached before the supporting arguments were written.