Everyone talks about AGI. Nobody agrees what it is.

The AI you use today is usually good at a particular kind of task: drafting an email, summarising a PDF, flagging a suspicious transaction. Artificial general intelligence, or AGI, is the idea of AI that can cope with almost any task a person can, and learn new ones, without being rebuilt for each.

That sounds clear until you ask who decides. Here are two of the best-known definitions:

• OpenAI says in its charter that it means "highly autonomous systems that outperform humans at most economically valuable work". The test is work output.
• Google DeepMind researchers propose measuring two things instead: how well a system performs, and how wide a range of tasks it can handle.

Even the people building these systems are cautious about the word. Dario Amodei, the chief executive of Anthropic, has written that he finds AGI "an imprecise term that has gathered a lot of sci-fi baggage and hype", and prefers to say "powerful AI". When the experts cannot agree on the finish line, be wary of anyone who says it has been crossed, or that it never will be.

The DeepMind paper compares AI to a skilled adult, whatever the task. A system moves up a rung when it beats a higher share of skilled people. Their general (AGI) ladder looks like this:

1. Level 1Emerging. Equal to or somewhat better than an unskilled person.
2. Level 2Competent. Better than at least half of skilled adults.
3. Level 3Expert. Better than at least 90% of skilled adults.
4. Level 4Exceptional. Better than at least 99% of skilled adults.
5. Level 5Superhuman. Better than every human. For general AI, this is artificial superintelligence.

When the authors first wrote the paper in 2023, they placed the leading chatbots at Level 1. They showed Competent-level skill on some tasks, such as short essays and simple code, but only Emerging-level skill on most others. They described the higher general levels as not yet achieved. The paper was revised in 2025, so check it for the current wording.

The ladder also has a separate "narrow" version, for systems that do one job. A chess engine sits at the top of that one. It is Superhuman at chess and useless at anything else, which is exactly why breadth matters in the definition.

Two groups answer this question very differently.

The researchers. A survey by AI Impacts asked 2,778 researchers who had published in top AI venues. Taken together, they gave a 10% chance by 2027 and a 50% chance by 2047 that machines could outperform humans at every task. For all human jobs becoming fully automatable, the combined forecast was a 10% chance by 2037 and a 50% chance by 2116. These are averages of many opinions, and the individual answers varied a lot.

The labs. People running AI companies tend to give shorter timelines. Amodei wrote that powerful AI, which he pictures as "a country of geniuses in a datacenter", "could come as early as 2026, though there are also ways it could take much longer". In early 2025, Sam Altman of OpenAI wrote that the company is "confident we know how to build AGI as we have traditionally understood it".

Picture a chartered accountancy firm in Jaipur, or a coaching centre in Lucknow. Neither needs to know the date of AGI. They need to know whether AI can save the team an hour a day this month. That is a question you can test now.

• Plan around the tools, not the labelJudge any product by what it does on your own tasks. "AGI-powered" on a sales page is marketing, not a measurement.
• Ask for evidenceWhen a vendor says their system is "human-level", ask: at which tasks, against which people, and measured how?
• Start with narrow, repeatable jobsInvoice matching, first-draft replies and meeting notes are easy to test and easy to check.
• Keep a person in chargeToday's AI is patchy. Have someone check its work, especially when it touches customers, money or personal data.
• Revisit every few monthsCapabilities change fast. A task that failed in January may work in June, so re-test rather than assume.