The Threshold Has Already Been Crossed

When researchers at various institutions began benchmarking large language models against standardized human performance measures, the results were uneven at first. AI systems outperformed average humans on some narrow tasks, underperformed on others, and failed in obvious ways on things any child could handle. That picture has shifted. Across a growing range of cognitive benchmarks — reading comprehension, legal reasoning, medical diagnosis, mathematical problem solving, coding — current AI systems now exceed the average adult human, often substantially. This is not a prediction about the future. It happened. The question worth sitting with is not whether it will occur but what follows now that it has.

What Average Means

The framing of "average human capability" requires some care. AI systems do not perform like a person. They have no persistent memory across conversations by default. They cannot initiate action in the world without tools attached to them. They lack embodied experience. They can produce confident-sounding errors in domains where their training data was sparse or low quality. But average human performance on knowledge and reasoning tasks is not especially high, and this is not an insult — it is an honest look at what ordinary cognitive output looks like at scale. Most people, on most days, are not performing at peak analytical capacity. They are tired, distracted, operating under time pressure, and drawing on knowledge that was last updated years ago. On these terms, AI systems running in optimal conditions consistently outperform average human output on structured tasks. A research team at Microsoft and Carnegie Mellon University evaluated GPT-4 across 57 subject areas using professional-level standardized tests and found performance exceeding the 80th percentile of human test-takers across most domains. That benchmark has since been moved by newer systems.

The Credential Gap

One underappreciated consequence of AI exceeding average human capability is what it does to credentialing. Professional credentials exist partly to signal that a person has crossed a competence threshold. Lawyers, doctors, engineers, and accountants earn credentials that are meant to assure the public they know more than a non-specialist. If AI systems can now match or exceed credentialed professionals on the knowledge and reasoning components of those fields — which they can, on structured tasks — the signals that credentials send become murkier. This does not mean credentials are useless. It means the value proposition is shifting toward what humans provide that AI does not: accountability, relationship, judgment under genuine uncertainty, the capacity to be held responsible for outcomes.

A Tangent on What Averages Hide

Averages in human performance obscure enormous variance. The average human driving performance is modest; the best human drivers, surgeons, or chess players operate at levels that no AI system currently touches in embodied, unstructured real-world conditions. Narrow benchmarks favor AI. Complex, embodied, relational, and novel tasks still favor exceptional humans. The error is treating the average as if it represents the ceiling rather than the midpoint.

What Institutions Are Actually Doing

Most organizations are not publicly discussing what it means to operate in a world where their average employee produces output that an AI system could match or exceed. That conversation is happening quietly, in workforce planning meetings, in strategic restructuring, in the gradual de-prioritization of roles centered on information retrieval and routine analysis. Research from Stanford's Digital Economy Lab has tracked the adoption of AI tools in professional services and found that the productivity gains are concentrated at the lower and middle performance bands — workers who were already high performers gained less in relative terms, while workers who were average or below-average gained substantially. This suggests AI is raising the floor of human professional output more than it is raising the ceiling.

The Psychological Dimension

What no benchmark captures is the psychological adjustment required of humans who built their identity around cognitive superiority. Being the one who knows things, analyzes things, solves problems — these are not just jobs. They are sources of meaning and self-worth for large portions of the population. When a system can do those things faster and more consistently, the adjustment is not only economic. It is existential. The people navigating this most successfully appear to be those who reframe their role from knowledge producer to judgment provider — from answering questions to knowing which questions matter and taking responsibility for what happens next.

The Question That Remains

Exceeding average human capability on benchmarks is not the same as wisdom, and it is not the same as consciousness. These remain genuinely open questions. But the practical reality is that organizations, individuals, and governments are making consequential decisions in a world where AI cognitive output is already more reliable than most human cognitive output on well-defined tasks. Sitting with that honestly is the precondition for responding to it well.