DeepMind proposes ten cognitive scales to measure progress toward AGI

Google DeepMind has published "Measuring Progress Toward AGI," an attempt to give the industry a tool for actually measuring progress toward AGI, rather than yet another classification that nobody can verify.

Where the problem came from

Almost three years ago, DeepMind published "Levels of AGI," a system of five levels of intelligence (from emerging to superhuman) and six levels of autonomy (from a simple tool to a fully independent agent). The analogy with the levels of autonomous driving worked well: it was structured, clear, and easy to explain to investors and journalists. The industry got a shared vocabulary, a kind of common terminology for talking about AGI.

But the classification had a fundamental flaw: there was no tool for checking where a given system actually stands. Any company could call its model "level 2" or "level 3," and no one had a way to refute it. "AGI" turned into a marketing label, convenient for press releases and fundraising but thoroughly inconvenient for science. The new paper tries to fix exactly this problem.

Ten scales instead of one score

The paper, released in March 2026, takes a fundamentally different approach. Instead of a single overall ranking, it defines ten separate scales, each measuring a specific aspect of cognitive ability. The scales are independent: a system can score high on reasoning but low on adapting to new tasks, and that mismatch will be plainly visible rather than hidden behind an averaged value. The result is a multidimensional portrait of a system, not a single number.

The key difference from conventional benchmarking is that the scales are built not on datasets and problem sets but on the toolkit of cognitive psychology, a science that has studied intelligence in real people for decades and developed methods that resist training effects. The aspects measured include:

• Working memory and context retention
• Planning and multi-step reasoning
• Transfer of knowledge to new domains
• Learning from a small number of examples (few-shot)
• Metacognition, or understanding the limits of one's own knowledge
• Causal reasoning
• Adaptation to data outside the training distribution

The authors present the framework as a starting point for discussion, not a final standard. The list of scales is open to extension.

Why this matters more than benchmarks

Until now, AI progress has been measured indirectly: MMLU, HumanEval, ARC-Challenge, GSM8K. The problem is that models have learned to overfit deliberately to specific benchmarks. A high MMLU score long ago stopped being a reliable indicator of real reasoning. Everyone in the industry knows this, yet the standards do not change.

A cognitive-psychology approach is much harder to game. If a model cannot generalize to fundamentally new tasks, no fine-tuning on a test set will hide that. Methods designed to measure human intelligence resist drilling by their very nature.

For investors, corporate AI buyers and regulators, this could potentially mean the end of an era in which any lab could announce a "breakthrough toward AGI" with no possibility of independent verification. Shared, measurable scales make systems from different companies comparable, and therefore accountable.

What this means

DeepMind is shifting the AGI conversation from "we have level N" to "here is exactly how this can be measured." This is not an answer on AGI timelines, nor a guarantee of consensus: different labs will interpret the scales in their own ways. But it is the first serious step toward common evaluation standards built on science rather than marketing.