Ask OpenAI, Google DeepMind, and Anthropic to define AGI, and you’ll get three fundamentally different answers — and that disagreement isn’t a technicality. It’s the reason “AGI is coming” headlines mean almost nothing without asking: coming by whose definition?
OpenAI’s bar is economic, not philosophical. Its charter defines AGI as a system that outperforms humans at most economically valuable workOpenAI’s charter defines AGI as systems that outperform humans at most economically valuable work, an economic test rather than a test of understanding, which the company itself grades. That’s a moving target by design — “economically valuable work” expands every year as new tasks get automated, which means OpenAI’s own finish line quietly shifts forward with it.
DeepMind treats AGI as a spectrum, not a switch. Its widely-cited 2023 taxonomy breaks capability into five levels — emerging, competent, expert, virtuoso, and superhuman — each measured on both performance and generality rather than a single pass/fail thresholdDeepMind published a widely-cited taxonomy in 2023 defining levels of AGI from emerging through competent, expert, virtuoso, to superhuman, each qualified by both performance and generality. Under this framework, today’s best models already sit somewhere in the middle — which makes “have we reached AGI yet” the wrong question entirely.
Anthropic won’t even use the term. CEO Dario Amodei has publicly called AGI a marketing term and prefers to describe the goal in more concrete language — something closer to a “country of geniuses” running inside a data center, with systems smarter than a Nobel laureate across most subjects, and a target as near as 2026Amodei calls AGI a marketing term and prefers to describe a country of geniuses in the data center, discussing systems smarter than a Nobel Prize winner in most subjects with a target year of 2026. Anthropic frames its mission instead as building advanced AI responsibly for humanity’s long-term benefit — deliberately avoiding a single capability threshold to declare victory at.
Why the disagreement isn’t academic:
These aren’t just branding choices — they shape real decisions. An economic-output definition justifies rapid deployment because the finish line is “useful at scale.” A capability-tier framework justifies incremental rollout because each level demands new safety evaluation. A refusal to name a finish line at all justifies slower, more cautious scaling because there’s no target to race toward.
Regulators drafting AI policy, investors pricing AI companies, and the public forming opinions about existential risk are all, often unknowingly, adopting one lab’s definition over another’s — usually whichever one made the most recent headline. That’s not a neutral choice. It determines whether you think AGI arrived in a benchmark last month or is still decades away.
The real risk isn’t that we don’t know when AGI will arrive. It’s that three different labs are already building toward three different finish lines, using the same word to describe none of them.