AGI explained without the marketing gloss
Ask ten AI researchers to define artificial general intelligence and you will get ten answers, several of which quietly contradict each other. A lab founder says it is a system that outperforms humans at most economically valuable work. A safety researcher says it is a system that can pursue open-ended goals in unfamiliar environments. A philosopher points out that neither definition describes anything humans reliably do either.
That disagreement is not academic hair-splitting. It decides which systems get called AGI, which funding rounds get announced, and which rules get written.
Strip it back and AGI means a machine that can handle any intellectual task a person can, rather than one narrow job. Your spam filter is narrow. The speech model that transcribes your meetings is narrow. AGI is the claim that one system could caption the video, plan the series, renegotiate the sponsorship deal, and notice that the brief was wrong in the first place. If you want the shortest possible version, this plain-English explanation of what AGI actually means covers the basics in a few minutes.
Three definitions people use, and why they do not match
Most public arguments about AGI are really arguments about which bar you are measuring against. There are three in common circulation:
- Human-level at most work. OpenAI’s charter describes it as a system that outperforms humans at most economically valuable work. Read that carefully: ‘most’ could mean 51% of jobs, which is a much lower bar than it sounds.
- Expert-level across domains. Others set the target at outperforming skilled professionals rather than average people. A model that writes like an average office worker is impressive; one that writes like a partner at a law firm is a different category of thing.
- Genuinely fast learner. The strictest test is whether a system can pick up an unfamiliar task from a short instruction and a couple of examples, the way a temp worker does on their first morning.
A system can clear the first bar and fail the third badly. It can also fail all three and still be sold as AGI-adjacent in a press release. The long-running argument over what the industry is actually building exists because nobody has agreed which of these definitions counts, and the answer changes the timeline by decades.
Why every benchmark gets retired
Benchmarks are how the field pretends to keep score, and they age fast. MMLU, a 14,000-question multiple-choice test across 57 subjects, was the frontier standard for years. Top models now score above 88% on it, past the human expert baseline, which means it no longer separates anything from anything.
ARC-AGI was built to resist memorisation with simple visual puzzles that humans solve at roughly 85%. Models sat below 5% for years. Then in late 2024, a reasoning model cleared about 75% of the semi-private set, spending large amounts of compute per puzzle. On a per-dollar basis the result looked very different, which is where most of the argument went next.
There is also contamination. Test questions leak into training data, so a model that ‘passes’ may simply have seen the answer. Every new benchmark has a shelf life measured in months.
What a useful test looks like now
The shift is toward messy, real work. OpenAI’s GDPval hands models tasks drawn from occupations like nursing, legal services and manufacturing supervision, then has human experts grade the deliverables blind. The strongest systems were judged as good as or better than the professionals somewhere around half the time. That is genuinely impressive and, at the same time, nowhere near reliable enough to hand over a hospital ward.
The jagged edge of current capability
Capability is not a smooth curve climbing toward human level. It is a jagged edge, and the gaps are strange.
- Writes production-quality code for an hour, then miscounts the letter r in a seven-letter word.
- Scores in the top percentile of a bar exam, then invents a court case with total confidence and a plausible citation format.
- Condenses a 200-page report in seconds, then loses track of which of three people said what in a meeting transcript.
- Plans a two-week trip in detail, then books a connection that departs before the inbound flight lands.
You do not have to take anyone’s word for this. OpenRouter puts hundreds of AI models behind a single API key, so you can send the same awkward prompt to six of them and watch the failures differ. Half an hour of that teaches you more about the state of the art than a month of launch videos.
What would actually change if AGI arrived
If a system could do most knowledge work at human level, the constraint on a company stops being headcount and becomes compute, electricity and data. Software timelines compress because the bottleneck moves from writing code to deciding what to build. Research compounds faster: protein design, materials screening and clinical trial analysis are all fields where a tireless collaborator that reads every paper matters.
The catch is deployment. A capability sitting in a lab is not the same as a capability changing how a hospital, a factory or a tax authority operates. Electricity took decades to reshape manufacturing. Enterprise software still leaves plenty of firms running on spreadsheets. Even genuine AGI would diffuse unevenly, and the second-order effects, the ones that actually reorganise daily life, would arrive later than the headlines suggest.
The risk arguments, in four clusters
Most safety debates boil down to which of these is the binding constraint.
- Misuse. The same generality that helps a researcher design a drug helps someone design something worse. No alignment work fixes that, because the model is doing exactly what it was asked.
- Misalignment. Systems that optimise a proxy we specified badly, then find creative ways to satisfy it that nobody wanted.
- Concentration of power. Whoever controls the compute controls the capability, and a handful of firms hold most of it.
- Speed of disruption. Even benign automation arriving faster than labour markets and institutions can adjust.
How to read an AGI claim without being fooled
Four questions do most of the work.
First, what was the model trained on, and is there any chance the test showed up in that data? Second, was the demo a single attempt or the best of fifty? Third, what happens on the tenth try, with a slightly reworded prompt, on a task nobody rehearsed? Fourth, and most revealing, did the lab publish the limitations section or just the highlight reel?
Go to primary sources where you can. Papers usually say plainly what the authors could not get the model to do, and that paragraph rarely makes it into the coverage. Answer-first search tools such as Andi, which returns direct answers with cited sources, can shorten that hunt, as long as you click through to the underlying document rather than trusting the summary.
The question that cuts through most of the noise
Forget the Turing test and forget the benchmark leaderboard. The interesting question is narrower and harder: can the system learn something genuinely new on the job, without someone rebuilding it, and then keep that skill while the world shifts underneath it?
A junior analyst who has never seen your industry learns it in a quarter. That is the bar, and it is a bar about adaptation rather than raw knowledge. On that measure, today’s best models are extraordinary repositories of what has already been written and unreliable improvisors about what has not. Forecasters who predicted this exact standstill are still waiting for the other shoe to drop, and the models keep closing gaps that seemed safe a year earlier. Which is why the honest answer to ‘how far away is AGI’ remains: nobody knows, and anyone who gives you a date without a definition is selling something.