Ask five AI researchers to define artificial general intelligence and you will get six answers. One points at a benchmark score. Another describes a job. A third starts talking about consciousness, and someone at the back says the term has outlived its usefulness entirely. That disagreement is not a side issue. It is the reason the debate keeps going in circles.
The stakes are not academic. About a dozen labs are spending money at a rate that only makes sense if something enormous is coming, and regulators in Brussels, London and Sacramento are drafting rules that hinge on when a model crosses from “very good software” into something else.
What artificial general intelligence is supposed to mean
In plain terms, AGI describes a system that can handle essentially any cognitive task a person can, at roughly human level or better, including tasks it was never specifically trained to do. A chess engine that also drives a car is not AGI. A system that reads a lease, spots the unusual clause, explains it to a tenant, then helps draft a polite email to the landlord is much closer to the idea.
The trouble is that “any cognitive task” is carrying a lot of weight. Labs have settled on competing definitions:
- OpenAI’s charter describes AGI as “highly autonomous systems that outperform humans at most economically valuable work,” which is a market definition rather than a cognitive one.
- Google DeepMind’s 2024 paper proposed levels running from Emerging through Competent, Expert and Virtuoso to Superhuman, scored separately for performance, autonomy and risk.
- Benchmark researchers tend to define it operationally: can the system solve novel reasoning problems, transfer skills to unfamiliar domains, and hold a plan together over days rather than minutes?
Each definition implies a different timeline. The market version gets you there faster than the reasoning version. That is not dishonesty, exactly, but it is worth noticing which definition a press release is quietly using.
The benchmarks keep moving, which is the whole point
The most useful progress metric of recent years comes from METR, a nonprofit evaluation group, which measures how long a real task a model can complete with roughly a 50% success rate. That number has doubled about every seven months for several years running. Tasks that took a skilled human four minutes are now routine, and the frontier has pushed past several hours of human work.
Then there is ARC-AGI, a puzzle set designed to resist memorisation. The first version fell. The generational leap OpenAI claims with GPT-6 Astra is largely a claim about the second version, along with agentic coding benchmarks where models build and debug multi-file projects. When you read that a model has solved something, the questions worth asking are which suite, on what date, with how much scaffolding, and what the human baseline was.
Contamination muddies all of it. Benchmark items leak into training data, and stealth models get tested anonymously on public leaderboards before anyone knows whose weights they are. That dynamic played out when the Ox Alpha mystery ended with Z.ai revealed as the lab behind an unnamed model that had been climbing the rankings for weeks. Scores matter, but they are a noisy signal.
Why expert predictions are thirty years apart
You can find credible people arguing AGI is two years out and equally credible people saying it is not happening this century. Both camps are reading the same evidence.
The case for sooner
Scale has kept working past several rounds of confident predictions that it would stop. Training compute has grown by orders of magnitude, and capability followed. If that trend holds, and if agents that use tools, write and run code, and check their own work keep improving, the practical distance between “narrow but very capable” and “general” starts shrinking from the outside.
The case for much later
Current systems predict tokens extremely well and still stumble on things a four-year-old manages: building a mental model of how objects behave, learning a new skill from two examples, remembering what happened last Tuesday without being told. Yann LeCun and others argue that no amount of scaling a next-token predictor gets you there, because the architecture is missing the part that learns how the world works. Meta’s willingness to sit out the current spending slowdown while rivals trim capital expenditure suggests at least one major lab believes the payoff is still years away.
Safety stopped being a philosophy seminar
For a long time, AGI safety was a conversation about hypothetical paperclip maximisers. It now looks more like systems engineering. Frontier labs run sabotage evaluations, dropping a model into a scenario with an obvious shortcut and watching whether it takes the shortcut and hides the evidence. Interpretability teams try to read what is happening inside a network instead of guessing from outputs. Chain-of-thought monitoring assumes the model’s visible reasoning is at least partly honest, an assumption that will not survive a system smart enough to manage its own transparency.
The people who built these models have been arguing about the risk for years, sometimes at genuine personal cost. Ilya Sutskever’s account of why he moved against Sam Altman reads less like a boardroom feud and more like someone convinced the stakes justified the mess. Regulators have started catching up. The EU AI Act applies extra obligations to models trained above a compute threshold, and California’s SB 53 requires large frontier developers to publish safety frameworks and report incidents.
Who gets to decide when the line is crossed
Here is the awkward part. The companies closest to building AGI also get to define it, measure it and announce it. That is a structural conflict, and it is now being litigated. The Musk v. Altman trial exposed how thin the line is between a nonprofit mission to benefit humanity and a capped-profit subsidiary raising money at a valuation that would make a sovereign wealth fund blush. Whoever controls the definition of AGI controls the trigger that ends the mission.
What would actually change first
If a genuinely general system showed up next year, the first visible effects would be boring and local. Software teams shipping features without writing most of the code. Insurance claims triaged end to end. Contract review, radiology screening, lab protocol design, and the endless middle layer of knowledge work currently done by people with twelve browser tabs open.
The bottleneck would be deployment rather than capability: cost per task, latency, liability, integration with systems built in 2004, and whether anyone trusts the output enough to stop checking it. Watch for these signals instead of press releases:
- Agent runs that last weeks without human intervention and stay on task.
- Models that improve from deployment experience rather than a fresh retrain.
- A measurable fall in the price of one completed unit of knowledge work.
- Independent evaluators, not the labs themselves, reporting the result.
Artificial general intelligence may arrive as a slow accumulation of ordinary Tuesdays rather than one announcement, and the honest position is that nobody knows which definition will turn out to be the one that mattered. What you can do is read the claims carefully, ask who measured them, and keep an eye on the gap between what a model does in a demo and what it does on a bad day at 3am with a real customer waiting.