TLDR: “Human-competitive” is a concept from the early days of AI meaning an algorithm performs as well as a human. Now that AI is superhuman at most digital tasks, the concept is obsolete. The relevant question is the reverse: for which tasks is a human AI-competitive?
Human-competitive
The term “human-competitive” was coined by John Koza in the context of genetic programming: a result is human-competitive if it is as good as or better than what a human expert could produce. Since 2004, the annual Humies awards at GECCO have formalized this into a competition rewarding evolutionary computation results that are human-competitive.
The concept was useful. The human was the gold standard, and the question was whether a machine could reach it. A notable milestone: Repairnator produced patches indistinguishable from human developer patches, accepted and merged into open-source projects without reviewers realizing they came from a bot.
This framing is now mostly obsolete. LLMs write code, papers, legal briefs, and scientific reviews. AI dominates at chess, Go, protein folding, and image classification. For most digital tasks, “human-competitive” no longer marks a frontier, it marks a floor.
AI-competitive
The interesting question has flipped. AI-competitive describes tasks or agents where human performance is on par with, or better than, the best available AI. It is the exact symmetric of human-competitive, with the reference point switched.
The digital world is where the concept bites. Here, AI has raw throughput advantages: millions of tokens per second, perfect recall, no fatigue. Yet humans retain pockets of genuine advantage.
Token efficiency. The key human advantage in the digital world is token efficiency. A human reads at roughly 200–300 words per minute but with aggressive semantic compression: skimming headers, reading the abstract and conclusion, jumping to the one paragraph that matters. That selective attention produces a high-value summary at a cost of perhaps a few hundred tokens of equivalent content consumed. An AI agent given the same task will often read the full document linearly, process it exhaustively, and emit a long chain-of-thought — consuming orders of magnitude more tokens to reach a similar conclusion. Vitalik Buterin frames it as “AI as the engine, humans as the steering wheel”: humans provide a small amount of high-quality signal. The value is in where you point the engine, not in the engine itself.
$$$ efficiency. Token efficiency has a direct cost consequence. As of 2026, AI can cost more than human workers for a growing set of tasks — Microsoft’s own data shows AI agents exceeding the cost of equivalent human labor in some workflows. A human who skims a 40-page report and extracts the key insight in ten minutes is not just token-efficient — they are cost-efficient. For such tasks the human is strictly AI-competitive: better output per dollar.
Humans are AI-competitive in the digital world when they skim well, steer efficiently, and intervene at the right level of abstraction. The productive human role in a world of superhuman AI is precisely the set of tasks that remain AI-competitive.