Google DeepMind Plans to Track AGI Progress With These 10 Traits of General Intelligence

The company that defines the AGI benchmark owns the AGI conversation

Google DeepMind has a proposal: stop arguing about when AGI will arrive and start measuring it.

The company published a paper this week titled "Measuring Progress Toward AGI: A Cognitive Taxonomy," which attempts to do something the AI industry has largely avoided — build a rigorous, empirical framework for tracking capability progress. Rather than declare a finish line, DeepMind has built a measuring system.

The framework deconstructs general intelligence into 10 cognitive faculties. Eight are basic building blocks: perception, generation (producing text, speech, and actions), attention, learning, memory, reasoning, metacognition (monitoring one's own thinking), and executive functions (planning, inhibition, cognitive flexibility). Two are composite — problem solving and social cognition — which require combinations of the basic faculties working together.

To evaluate performance, DeepMind proposes a three-stage protocol: test AI systems on held-out cognitive tasks designed to isolate each faculty, establish human baselines by having demographically representative adults complete the same tasks under identical conditions, then map each model's performance against the human distribution. The result is a cognitive profile — a structured view of where a system is strong, weak, or above average relative to people.

The practical implementation is being crowdsourced. DeepMind launched a Kaggle hackathon in partnership with the research community, with $200,000 in prize money. The focus is on five cognitive abilities where reliable evaluation benchmarks do not yet exist: learning, metacognition, attention, executive functions, and social cognition. Submissions close April 16; winners are announced June 1.

Why this is a strategic move, not just a science project

The paper is genuine cognitive science. But it is also a power play. "Whoever defines the metrics wins the narrative" — and in the AGI debate, that is worth more than any single capability claim.

The AI industry has spent years debating whether systems have achieved AGI, with each claim contested, vague, or self-serving. OpenAI says AGI is near. Anthropic says it requires more safety work first. Academics say the term is meaningless. Policymakers trying to write rules do not know what they are regulating. Enterprise buyers cannot evaluate competing claims without a common language.

DeepMind is offering that language. And as the company that defines it, DeepMind gets to shape how progress is measured, reported, and governed.

This is not the first time DeepMind has tried to own the AGI definition. In December 2023, the company published a "Levels of AGI" paper that categorized AI systems by depth (performance) and breadth (generalizability across tasks) — a framework that drew on how self-driving levels are categorized. But that paper described the problem; it did not solve it. The new paper goes further by proposing the actual measuring stick.

The framework also has a governance function the paper does not fully explore. If there is a credible, agreed-upon cognitive profile for "average adult human general intelligence," regulators and enterprise buyers gain a tool for assessing when AI systems cross meaningful thresholds. That has implications for everything from AI liability to procurement decisions.

The reasonable objection

The framework is only as good as the cognitive theory it rests on. The paper explicitly acknowledges that there is no guarantee the 10 identified faculties actually capture the essence of human general intelligence. The taxonomy draws on decades of psychology and neuroscience — which is a strength — but those fields themselves debate what general intelligence means. Psychometricians have argued for decades about whether IQ-style measures capture anything meaningful about human cognition. Transplanting that debate into AI evaluation does not resolve it.

There is also the problem of test contamination. The paper notes that many existing benchmarks are public, meaning training data may have included their contents. The DeepMind team is working with academics to build held-out evaluations — a real acknowledgment of the problem — but building robust, non-leaked benchmarks for metacognition and social cognition is genuinely hard. Those are exactly the faculties the hackathon is trying to measure.

The Register's coverage of the hackathon put it plainly: the framework is only useful if acing it actually predicts real-world performance better than narrower specialist systems. That connection has not been established.

What this means for the industry conversation

The practical immediate effect is on enterprise buyers and policymakers, who now have a structured way to ask the question "how does this AI compare to a capable human at these specific cognitive tasks?" That is more useful than asking whether a model has "achieved AGI."

The harder long-term effect is on competitive positioning. If DeepMind's 10-trait framework becomes the standard reference point for AGI discussion — in academic papers, in regulatory filings, in enterprise RFPs — then DeepMind has anchored the conversation on terrain it mapped. Rivals OpenAI and Anthropic will be evaluated against a framework designed by a competitor.

Whether the framework earns that authority depends on whether it actually predicts useful performance — a question that will take years of data to answer. In the meantime, DeepMind has made the first substantive move to define the measuring stick. In a debate defined by assertion and counter-assertion, that is itself a significant act.