A language model clears the bar exam at roughly the 90th percentile. The same generation of models still fumbles the question of whether a glass of water will spill if you tip it over. That mismatch is the whole story of human-level artificial intelligence in a single paragraph: brilliant spikes of competence sitting right next to gaps a seven-year-old would never have.
So where does the field actually stand, and what would it even mean to get there?
The definition is squishier than it sounds
Human-level AI, often written as HLAI, means a system that can do anything a competent adult can do, across domains, without being rebuilt for each new task. Not a specialist that crushes grandmasters at chess but can’t hold a conversation. A generalist.
That already puts it in a different category from almost everything in commercial use. Most deployed systems are narrow AI: excellent at one bounded job, helpless the moment you step outside it. If you want the full ladder from spam filters up to superintelligence, there’s a clear breakdown of the different types of artificial intelligence that maps where each rung sits.
The goalposts keep moving
Every time a milestone falls, the definition shifts. Passing the Turing test was the bar for decades. Chatbots cleared it and the response was a shrug and a new test. Beating humans at Go, writing passable poetry, scoring well on medical licensing exams: all of it got reclassified as not real intelligence within about eighteen months of each achievement.
Some of that scepticism is healthy. Some of it is just people quietly moving the finish line because the old test turned out to measure the wrong thing. The honest version of the question isn’t whether a machine can pass a particular exam. It’s whether that exam measured what we thought it did.
General isn’t the same as expert
Here’s a distinction that trips up a lot of discussion. A human-level system would be about as capable as an ordinary person, not as capable as a whole room of specialists. There’s a solid working definition of what artificial general intelligence actually means that keeps those two ideas apart, and it matters. A system that reasons like a bright generalist is still a very different thing from one that out-thinks every expert in every field at once.
Four things that make this genuinely hard
1. Common sense
A five-year-old knows that a dropped cup breaks, that a wet floor is slippery, and that saying a friend looks tired might land badly. None of that is written down in the text a language model trains on, because it’s so obvious nobody bothers to explain it. Models pick up fragments from sheer volume, but the gaps are oddly shaped. Ask one to plan a dinner for eight with one vegetarian and it’s fine. Ask it to reason about why a stool tips over when you lean on one leg and you find the edge of the map.
2. Transfer across domains
Learn to drive a van and you can drive a truck with an hour of adjustment. The skills move with you. Neural networks mostly don’t work that way. Fine-tune a model on medical text and watch its coding ability wobble. Something transfers, but it’s fragile and hard to predict in advance.
3. Sample efficiency
Frontier models train on trillions of tokens, which is roughly a hundred thousand times the language input a child gets before age ten. AlphaGo’s predecessor learned from tens of millions of games. Humans need a few dozen hands of cards to learn a new table game. Data-hungry learning works, but it’s a very different mechanism from the one humans use, which is why some researchers doubt it converges on the same destination.
4. Memory that persists
Every conversation with a large model starts fresh. It doesn’t build up expertise over months of work with you, doesn’t remember that you hate being interrupted, doesn’t carry a grudge. Human-level AI implies continuous learning from experience, and today that’s mostly bolted on with retrieved notes rather than built into the system.
Where today’s systems actually sit
The honest picture isn’t a smooth progress bar. It’s what Wharton’s Ethan Mollick calls a jagged frontier: models that are superhuman at summarising a hundred documents and mediocre at counting the letters in a sentence. AlphaFold has predicted structures for over 200 million proteins, a feat no research team could match in a lifetime. The same era of AI still gets confused about which object is on the left in a photograph.
That unevenness is why concrete examples are so useful. Looking at real artificial general intelligence examples and, crucially, at the ones that get excluded, shows how much depends on where you draw the line between impressive and general.
How would we know we’d arrived?
Any serious test would need to go beyond a score on a fixed exam. Reasonable criteria look something like this:
- Pick up a genuinely new field from a few examples, the way a new hire learns a niche software tool in a week.
- Carry a project across weeks or months without losing the thread of what was decided and why.
- Notice when it’s wrong, rather than confidently inventing a source that doesn’t exist.
- Reason about the physical and social world well enough to plan a lab experiment or mediate an argument.
- Chain those abilities together on a task nobody designed a benchmark for.
Each of those is partly solved and partly not. That partial state is exactly why forecasts diverge so wildly.
Why the timelines are all over the map
A 2022 survey of several hundred AI researchers put the median estimate for human-level machine intelligence around 2060, with a wide spread from a decade to never. Geoffrey Hinton left Google in 2023 warning that the thing is closer than people think. Yann LeCun has argued, in the same period, that current language models are a dead end for real general intelligence and something else is needed.
Both men know the field deeply and they’re describing different gaps. One camp treats scaling as the main missing ingredient. The other thinks scaling hits a wall without new ideas about world models, planning, and persistent memory. There’s a fuller walk through the real arguments for and against human-level AI if you want to weigh them properly rather than take a headline’s word for it.
What would actually change
Less than the movies suggest, and more than most companies are ready for. The change wouldn’t arrive as a robot that does your job. It would arrive as the boring tasks inside every job becoming cheap overnight: reading the contract, drafting the first version, triaging the inbox, summarising the meeting, chasing the invoice. Radiologists would still run scans. They’d just spend far less time on the parts a machine handles.
The step from there to something that improves itself is a separate question entirely, and it’s worth being precise about it. There’s a good explanation of where the line between AGI and superintelligence actually falls, because almost every scary scenario people describe is the second one, not the first.
Three questions to ask about the next big claim
Announcements come weekly, so a filter helps. First: was the system trained on the test? Benchmarks leak into training data constantly, and a model that has seen the answers is not evidence of anything. Second: what was the base rate? If the task can be done by a competent intern in an afternoon, beating a specialist on it is less impressive than the press release implies. Third: does the capability survive a change of setting? A demo that works in one narrow configuration and falls apart when you alter the input by ten percent tells you something about robustness, which is what actually matters in practice.
Human-level artificial intelligence isn’t a switch that flips on a Tuesday. It’s a frontier that keeps getting pushed into weird shapes, advancing fast in some directions and stalling in others for reasons nobody fully predicted. The useful thing to track isn’t the headline number attached to a benchmark. It’s whether the gaps between those spikes are shrinking, and whether the systems that produce them can carry what they learned from one problem into the next.