Ask what a system can do, across which tasks, with how much help, and how reliably. A label or deadline cannot answer those questions.
Artificial general intelligence, usually shortened to AGI, sounds like a single destination. In practice, a headline using the term may be talking about very different things: broad competence across intellectual tasks, an economically useful worker, or a system that can operate independently for long periods. Those are related possibilities, but they are not interchangeable tests. Before asking when AGI will arrive, ask what the speaker means by it and what evidence could change their mind. Otherwise, an apparent disagreement about timelines may actually be a disagreement about the finish line.
Start with a working definition
For this guide, AGI means AI with broad, adaptable capability across many intellectual tasks, rather than exceptional performance in just one narrow activity. This is a working description, not a universal scientific definition. A system’s ability to answer many types of questions is relevant evidence, but the range of its topics alone does not establish dependable general competence.
The research paper Levels of AGI, in its May 2024 version, proposes separating the depth of performance from the breadth of tasks and considering autonomy separately. Its value for a reader is the distinction: a highly capable assistant and a highly autonomous agent need not present the same benefits or risks.
Four questions behind any impressive demonstration
AI Affect suggests reading capability claims through four questions. This is an editorial checklist, not an AGI benchmark.
- What was measured? Writing a convincing explanation, solving a new problem and completing a real workflow are different achievements. Ask for the actual task and the scoring rule.
- How broad was the success? A system might excel in one domain and struggle elsewhere. Look for a profile of strengths and weaknesses, rather than a single headline score.
- How much support was needed? Detailed prompts, repeated attempts, specialist tools and human corrections may all be part of the result. They should be described.
- How reliably did it work? A best example demonstrates possibility. Repeated trials, reported failures and independent evaluation tell you more about dependability.
Why a benchmark is evidence, not the whole answer
A benchmark measures performance under a particular design. Before extending its result to everyday life, inspect its scope. An evaluation might use short tasks, a restricted environment or a limited set of languages. These choices are not inherently wrong. They define what the result supports.
A May 2026 paper by Burnell and colleagues proposes a cognitive taxonomy spanning ten faculties, alongside targeted, held-out tasks to produce a capability profile. It is another attempt to make the discussion more measurable. It should be read as a research proposal, not an official certificate or proof that one agreed threshold has been reached.
Capability is not the same as permission to act
Imagine two systems that produce the same quality of travel plan. One offers suggestions. The other can buy tickets and change bookings using your accounts. Even if their underlying reasoning were identical, the second system would require additional decisions about authority, oversight and recovery from mistakes. That example illustrates why autonomy deserves separate attention.
Likewise, a system can create consequential harms without meeting someone’s definition of AGI. A label is therefore a poor substitute for evaluating a specific use. Ask which actions are possible, who approves them, and what happens when the result is wrong.
Treat future timelines as forecasts
Claims that AGI will arrive in a certain year depend on assumptions about technical progress and the threshold being used. A prediction can be worth discussing without becoming an established fact. Useful forecasts state their definition, evidence and uncertainty.
For humanity, the immediate task is to improve how capabilities are evaluated and how powerful systems are governed. That work remains useful across competing definitions. You can take AI’s potential seriously while keeping a clear boundary between a demonstrated result, an extrapolation and a speculative scenario.