How Progress Toward General AI Is Measured (and How to Audit the Headline)
«Recent advances toward artificial general intelligence» is a phrase that surfaces every few weeks and almost never says what the advance was measured against. There is a useful operational definition, published in 2019, and a 2024 experiment that teaches you to tell a system that reasons from one that recognizes the shape of the question. That is enough to audit any headline.
Every few weeks a headline announces a breakthrough toward artificial general intelligence. Almost none answer the one question that decides whether the breakthrough is real: measured against what?
This is not a specialist's question. It is the same one anyone would ask of a claim that a car is «30% better»: better at what, compared to what, under which conditions? In artificial intelligence the question has become especially urgent, because the industry measures its own progress on exams it selects itself, and because there is a way to pass an exam that bears no resemblance to knowing the subject.
Let us look at how the measuring actually works, with the documents to hand.
A definition you can use
The most useful starting point was published by François Chollet in November 2019, in a paper titled «On the Measure of Intelligence». His proposal changes what is being measured: not what a system can already do, but «intelligence as skill-acquisition efficiency».
The difference is enormous, and an everyday example makes it plain. A student who solves a thousand problems because they have seen those thousand problems is not demonstrating what a student demonstrates by solving one problem of a kind they had never seen. The first shows repertoire; the second shows the ability to acquire a new skill from little information. If generality is what interests us, measuring repertoire measures the wrong variable.
From this comes ARC, the Abstraction and Reasoning Corpus, which Chollet builds — in his words — «upon an explicit set of priors designed to be as close as possible to innate human priors». They are visual puzzles, easy for a person and deliberately resistant to being solved by piling up examples.
The project continues at ARC Prize, with a second version from 2025 whose stated target is for some system to reach 85% accuracy, and with a human calibration study run in early 2025 in San Diego with more than 400 participants. A practical warning: the specific scores change from month to month, so any figure you read in an article — including this one — should be checked at the source before being quoted.
The experiment that teaches you to tell the difference
Here is the piece that turns all this into a reading tool. In October 2024 a team at Apple published GSM-Symbolic, putting leading models through controlled variations of grade-school mathematics problems.
They did two very simple things. First: change only the numerical values in the problem, leaving its structure intact. The result, in the paper's words: «the performance of all models declines when only the numerical values in the question are altered».
The second is more revealing still. They added one sentence to the problem that looks relevant but does not affect the reasoning — the kind of irrelevant detail that fills any real-world problem. The effect: «adding a single clause that seems relevant to the question causes significant performance drops (up to 65%) across all state-of-the-art models».
Consider what that means. A system that had genuinely understood the problem would discard the irrelevant detail without blinking, exactly as you do. One that has learned the shape of the question loses its footing when the shape changes, even though the problem is the same. The drop is not a measure of difficulty: it is a measure of what the earlier competence was made of.
Why exams age badly
There is a structural dynamic behind this worth understanding, because it explains why benchmark numbers climb so fast.
When a test becomes the industry's reference, it becomes a target. Teams optimize against it, its problems and solutions circulate online, and that material ends up, one way or another, near the data the next generation of models is trained on. When that happens, the test stops measuring capability and starts measuring prior exposure. The classic symptom is precisely the one GSM-Symbolic documents: high performance on the original test and a drop once details that should not matter are altered.
This is why tests are constantly renewed, and why a spectacular result on a veteran exam says considerably less than a modest one on a freshly minted one.
And why «AGI» does not mean the same thing to two people
One problem precedes any measurement: the term itself. In November 2023 a group of researchers led by Meredith Ringel Morris, with Shane Legg among the authors, published «Levels of AGI for Operationalizing Progress on the Path to AGI», an attempt to replace the yes-or-no with a scale.
Their proposal sorts systems along two axes: depth, the level of performance reached, and breadth, the range of tasks over which it is reached. It adds considerations of autonomy, risk and human-AI interaction, and sets out six principles that, they argue, any useful ontology for this field should satisfy.
The particular scale is arguable. What is not arguable is that «AGI» can no longer be used as though it named an agreed thing. When someone says a system is «approaching AGI», the first reasonable question is what they mean by AGI, and the second is which of the two axes improved. Very often the improvement is in depth — doing better what it already did — presented as though it were breadth.
The capability: four questions for any «advance toward general AI»
1. Against which test, and when was that test created? An old, much-cited exam has had years to seep into training data. A newly published one has not.
2. Does the result survive a trivial variation? Change the numbers, reorder the wording, add an irrelevant fact. That is the GSM-Symbolic test, and anyone with access to the system can reproduce it. If performance collapses, what was there was not understanding.
3. Is the improvement in depth or in breadth? Doing a known task better is not the same as doing a new task at all. Only the second points to generality.
4. Who chose the exam? When the vendor selects which tests to publish, what you are seeing is not the system's performance: it is the subset of results the vendor decided to show.
None of these four questions requires knowing how to program. They require treating a technical announcement the way you would treat a clinical trial result: asking about the protocol before asking about the outcome.
The deep end, undiluted
The documents holding this piece up are open and can be read without intermediaries:
- Chollet (2019), «On the Measure of Intelligence». Dense but foundational: it is where the case is made that measuring accumulated skills does not measure intelligence.
- Mirzadeh et al. (2024), GSM-Symbolic. Short and devastating; the concrete examples of altered problems need no technical background.
- Morris et al. (2023), «Levels of AGI», to have a scale to argue with instead of an elastic term.
- ARC Prize, where the rules and current results of the abstract reasoning test are published.
The capability you leave with: faced with any announced advance, ask which test it was measured against, whether that test is new enough not to have been absorbed, and whether the result survives having one number changed in the question. It is the same scrutiny you would apply to any empirical claim, and it is enough to separate a real capability from a score.
This article was produced with artificial intelligence under human editorial oversight.
