In 2023, a model scored around the 90th percentile on the Uniform Bar Exam. A year later, the same class of systems still couldn’t reliably book a plumber, notice that a PDF had two columns, or remember what you told it on Tuesday. Both things are true, and together they explain why the question of whether AGI is possible produces so much shouting and so little agreement.
What the question actually asks
Artificial general intelligence usually means a system that can pick up and perform any intellectual task a human can, at roughly human level, across domains it wasn’t specifically trained for. That’s the short version. The long version is a minefield, which is why even the people building these systems disagree about what “artificial general intelligence” actually means.
There are two separate questions hiding inside “is AGI possible”. The first is physical: does anything about the universe forbid a machine from doing what a brain does? Almost nobody serious argues that it does. The second is practical: can we get there with silicon, electricity, data and money on a sensible timeline? That one is wide open.
The case for yes
Start with the obvious. Human brains are physical objects. They run on roughly 20 watts, about the draw of a dim lightbulb, and they manage general intelligence just fine. Nothing in physics says cognition requires meat. If a lump of tissue can do it, some arrangement of transistors probably can too, even if we don’t yet know the arrangement.
Scaling has embarrassed the sceptics before
Every few years someone declares a wall, and every few years a model walks through it. Translation, chess, Go, essay writing, code completion: each was supposed to be the thing machines couldn’t do. Each fell. That pattern has made a generation of researchers suspicious of confident never claims, and reasonable people now treat “it’s just autocomplete” as a claim that needs updating every six months.
Narrow miracles hint at something wider
AlphaFold predicted structures for roughly 200 million proteins, work that would have kept structural biologists busy for decades. Demis Hassabis has argued that this is the shape of what comes next: AI as an engine for compressing scientific discovery rather than a single clever tool. If a system can absorb an entire field and then generate novel insight inside it, the gap between narrow and general starts to look like a matter of scope rather than kind.
Then there’s transfer. Models trained mostly on text can write working code, explain a lease clause, and walk a teenager through algebra. Nobody explicitly programmed those skills in. They emerged from a training objective that only asked for the next token.
The case for no
The pessimists aren’t Luddites. Most of them work in the field, and their objections are specific.
- Data scarcity. Human-written text is finite, and the best models have already consumed a large fraction of it. More compute pointed at less fresh data is a bad trade.
- Sample efficiency. A four-year-old learns “dog” from three examples. Frontier models need billions of tokens for the same trick, and they still get confused by a dog in a Halloween costume.
- Continual learning. Current systems are frozen after training. They don’t consolidate experience, don’t build expertise over years the way a surgeon does, and forget everything the moment a session ends.
- Grounding. Language models learn the statistical shape of text, not the world. Nobody has shown convincingly that this is enough for genuine causal reasoning about a kitchen, a factory floor, or a patient.
- Energy and cost. A frontier training run can cost hundreds of millions of dollars and draw power like a small town. The brain manages on a sandwich and a night’s sleep.
- Reliability. An 85% success rate is impressive in a demo and useless in a payroll system. General intelligence probably needs something closer to 99.9%, consistently, on tasks nobody anticipated.
Half the debate is about words, not machines
Chess was once treated as a hallmark of human intellect. A machine won, and the target moved to Go. Go fell. Then it was creativity, then humour, then “real understanding”, and now it’s holding down a job for a month without supervision.
This is the goalpost problem in AGI, and it makes the original question slippery. If AGI means whatever we haven’t built yet, the concept is unfalsifiable by construction, and “possible” becomes a claim about language rather than technology. The more useful move is to decide in advance what evidence would settle it.
What would actually count as proof
Testing whether AGI is achievable means agreeing on the tests before the results arrive. Reasonable candidates:
- A system that learns a genuinely new professional domain, say marine insurance or crystallography, from modest data in weeks rather than years.
- Multi-day autonomy: an agent that runs a project for a week, recovers from its own mistakes, and reports accurately on what it did and where it guessed.
- Transfer without retraining, such as reading a service manual and then operating the machine it describes.
- Benchmarks that resist memorisation, like ARC-AGI, where scores have climbed steadily but remain far below human levels.
Notice that none of these require consciousness, and none are settled by a chatbot producing a decent sonnet on request.
AGI is not the same as superintelligence
People slide between the two constantly. AGI means human-level across the board. Superintelligence means decisively better than the best humans at everything, and it’s a separate leap with its own arguments and its own skeptics. Even if human-level generality arrives, the line between AGI and superintelligence is where most of the risk conversation actually lives, and the two should never be used interchangeably in a serious discussion.
What to watch instead of timelines
Timeline predictions have a dreadful track record. Five-year forecasts from 2015 look quaint now. Ten-year forecasts being published today will probably look the same. Better signals to track:
- Cost per completed real-world task, rather than cost per token or per query.
- Whether models measurably improve from experience inside a deployment.
- How much human correction a long task needs before it ships.
- Whether new capability comes from scale or from architecture changes. The second is the stronger signal by far.
A design tool can produce a passable poster in four seconds and still has no idea why the poster works. That’s as good a small illustration of the gap as any; you can read a full breakdown of what Canva Magic Design nails and what it misses to see narrow competence dressed up as general help.
So is AGI possible? Almost certainly yes in the physical sense, since nothing forbids it. In the practical sense, nobody has a clean argument either way, and anyone who claims certainty is selling something. The useful version of the question isn’t binary at all. It’s a cluster of measurable claims about learning speed, reliability and autonomy, and those you can actually check, one deployment at a time.