A new paper from Google DeepMind, co-authored by company co-founder Shane Legg, is getting attention across the AI world. It proposes a cognitive framework for measuring progress toward artificial general intelligence, and the core insight is one that every business leader deploying AI should understand: today’s AI systems have a “jagged cognitive profile.”

That phrase captures something genuinely important. AI models can outperform most humans in mathematics, factual recall, and pattern recognition. But they still trail the average person in learning from new experience, maintaining context across long interactions, and reading social situations. The gap is not random. It follows predictable patterns that the paper maps in detail.

What the Framework Says

The paper, titled “Measuring Progress Toward AGI: A Cognitive Taxonomy,” identifies ten core cognitive abilities that matter for general intelligence: perception, generation, attention, learning, memory, reasoning, metacognition, executive functions, problem-solving, and social cognition.

Eight of these are classed as foundational faculties. Two, problem-solving and social cognition, are described as composite abilities that require all the others working together.

The framework’s practical threshold is this: for an AI system to qualify as AGI, it would need to at least match median human performance across all ten areas. By that standard, no current system qualifies. Today’s frontier models are extraordinary in some columns and average or below average in others.

This is a deliberate departure from how AI has traditionally been benchmarked. Existing evaluations like MMLU and BIG-bench test knowledge and task performance. This framework asks a different question: what kind of cognitive machinery does the AI actually have? Are those results coming from genuine reasoning, or from pattern-matching on training data?

DeepMind’s answer matters because it’s grounded in decades of psychology and neuroscience rather than engineering convenience. The researchers introduce a three-stage evaluation protocol that checks whether AI systems exhibit human-like problem-solving patterns, match average human capabilities, and eventually surpass top human experts in specific domains.

The Kaggle Hackathon

Alongside the paper, Google DeepMind launched a competition through Kaggle to crowdsource better evaluation tools. The hackathon runs through April 16, 2026 and offers a total prize pool of $200,000.

The focus is on five areas where existing evaluation methods are most limited: learning, metacognition, attention, executive functions, and social cognition. These happen to be the same areas where today’s AI systems are weakest.

Top two submissions in each of the five tracks win $10,000 each. Four grand prizes of $25,000 go to the overall best submissions. Results are announced June 1.

The decision to open benchmark design to the broader community is notable. DeepMind is essentially saying: we cannot objectively measure our own progress in isolation. External scrutiny produces more credible standards. For anyone who has watched AI labs self-report capability milestones using benchmarks they designed, this is a meaningful shift in approach.

What This Means for Business

The “jagged profile” concept is directly useful for anyone making decisions about where to deploy AI.

If your AI use case involves tasks that fall into the areas where models already exceed human performance, deployment risk is low and productivity gains are real. That covers a lot of ground: data analysis, report generation, code review, document summarization, customer query routing, and anything involving retrieval and pattern recognition at scale.

If your use case requires the AI to learn on the job, navigate ambiguous social dynamics, exercise judgment in novel situations, or manage competing priorities over extended time horizons, the picture is more complicated. These are the areas the framework identifies as genuinely underdeveloped, and they are precisely where businesses run into trouble when AI deployments fail to meet expectations.

The practical implication is not to avoid AI in these areas. It is to understand what you are getting and design accordingly. Build human checkpoints into workflows that require learning and adaptation. Pair voice AI agents with clear escalation paths when social or emotional complexity enters the conversation. Set expectations with your team that AI is strong in certain cognitive lanes and still developing in others.

This framework also matters for the vendors you evaluate. As the Kaggle hackathon produces new benchmarks built on this taxonomy, you will start to see more granular capability data for commercial AI products. Right now it is nearly impossible to compare models across labs because every lab uses different metrics. DeepMind’s cognitive taxonomy, if it gains adoption, could create a common language for those comparisons.

The Bigger Picture

It is worth noting the timing. This paper arrived the same week that Nvidia’s Jensen Huang told an interviewer that “we’ve achieved AGI,” and the same month that Anthropic announced its model is “industry-leading by a wide margin” across coding and tool use. Everyone in AI has a definition of AGI that makes their current work look like success.

DeepMind’s contribution is a definition that is harder to game. By grounding the standard in cognitive science rather than benchmark scores, and by opening the evaluation process to independent development, the framework creates accountability that internal metrics cannot provide.

For businesses, the practical question is not whether any given system has achieved AGI. It is whether the system handles your specific task reliably. The cognitive taxonomy gives you a useful lens for asking that question more precisely.

If you’re deciding where to start with agents, start here. The free Working With Claude field guide walks through the ecosystem, Claude Code, and a real rollout plan. Get your copy.