AI tests mostly check knowledge, reasoning and style — but do they measure the skills you'd need to trust an AI to work alone?

This explores which separate kinds of AI ability today's tests actually capture, and which important ones they leave unmeasured. The corpus suggests assessments cover four territories reasonably well: knowledge and reasoning, surface style, social interaction, and step-by-step agent behavior. The abilities needed for autonomous work and safe deployment are barely measured.

"AI competence" isn't one thing, and it doesn't grow as one thing. FLASK breaks evaluation into 12 skills and finds they scale very differently. Metacognition saturates around 7B parameters and logical efficiency plateaus near 30B, while reasoning and knowledge keep improving with size Do all AI skills improve equally as models scale?. Style is the sneaky one. Open-source models copy the surface style of stronger models but not the reasoning underneath, so one overall score can flatter a model that only got the form right. At the frontier, Humanity's Last Exam uses 3,000 expert-written questions to separate models where MMLU has saturated. It measures closed-ended expert knowledge, and it discriminates only temporarily Can frontier exams really measure cutting-edge AI capability?.

The second domain is social. SOTOPIA scores social intelligence on seven dimensions at once: Goal, Believability, Knowledge, Secret, Relationship, Social Rules, and Financial. An agent can therefore hit its goal while leaking a secret or breaking a social norm. It also makes efficiency a measurable skill: humans average 16.8 words per turn against GPT-4's 45.5 Can social intelligence be measured across seven dimensions?. Agent benchmarks make a parallel move. They shift from scoring the final answer to scoring the whole interaction, including process quality, recoverability, coordination, and robustness How should we evaluate agent behavior beyond final answers?.

What's missing is telling. Autonomous science needs hypothesis generation, experimental design, data analysis, and iterative self-correction. Standard benchmarks don't reliably test any of them, and self-correction, the hardest, is where reasoning accuracy is documented to degrade What capabilities do AI systems need for autonomous science?. Whether a system's errors stay visible, contained, and recoverable is only partly instrumented. Chain-of-thought disclosure covers visibility, incident counts cover containment, and rollback timing covers recoverability. No single measure spans all four conditions or the human-institution factors around them How can we measure whether AI errors stay visible and recoverable?. Endpoint scores also mislead for a deeper reason. AI separates the finished form of intellectual work from the reasoning that produced it Does AI separate intellectual form from the thinking behind it?, so a polished output can't certify the competence behind it.

The measuring tools need checking too. An agent-based judge that gathers evidence showed 0.27% judge shift against 31% for a standard LLM judge on complex tasks, though its memory module cascaded errors Can agents evaluate AI outputs more reliably than language models?. Self-report isn't a shortcut either. Pooled across three studies, self-rated and objectively measured AI competence correlated at just .055, with confidence intervals including zero Can self-ratings replace objective performance scores for AI competence?. Tests are good at measuring what a model knows and how well it reasons, and they are getting better at social behavior and process. "Can it do original research, and would we catch its mistakes?" is still mostly unmeasured.