AGI Definition Drift: What OpenAI's Claims Actually Mean

When OpenAI released its Astra model in September 2026 and suggested artificial general intelligence had arrived or become imminent, the announcement marked more than a technical milestone. It highlighted how dramatically the meaning of AGI has shifted since the term emerged in academic circles.

The concept of AGI once belonged to a small community of researchers working through the AI winter of the 1990s. For these pioneers, the term described machines capable of general reasoning across domains—not necessarily superhuman performance, but broad adaptability. Today, that definition has blurred considerably.

Why it matters

The stakes extend beyond semantics. AGI claims steer trillions in investment, shape regulatory frameworks, and fuel public anxiety about AI displacement. When companies declare AGI achieved without shared definitional standards, they risk policy decisions based on marketing rather than technical reality. Understanding what different stakeholders mean by AGI becomes essential for separating genuine capability advances from rebranding exercises.

Competing Definitions Among Researchers

Blaise Agüera y Arcas, vice president of Technology & Society at Google, argues we reached AGI when large language models demonstrated general capability. In a 2023 essay with Stanford's Peter Norvig, he suggested that mastering language—an "AGI-complete" task—unlocks broader capacity because it requires acquiring substantial world knowledge beyond grammar.

Agüera y Arcas notes that unlike humans, who can lose language ability through brain damage while retaining other skills, LLMs don't exhibit clean separation between linguistic and reasoning capabilities. The ability to call on other computational modules makes language a gateway to open-ended functionality.

Other researchers disagree. Joscha Bach of the California Institute for Machine Consciousness contends truly intelligent systems shouldn't require consuming the entire Internet and vast computing resources for simple queries. A genuine AGI would learn and reason more efficiently than humans, not just faster.

Ben Goertzel of the SingularityNET Foundation, an AGI pioneer, reports that OpenAI's Astra model doesn't accomplish anything he couldn't do independently—it simply accelerates his work. He finds current systems lack creative spark and the ability to learn on-the-go from experience. As Agüera y Arcas acknowledges, today's models can learn temporarily during a session, but that learning doesn't persist.

The Fusion Problem

The original AGI concept explicitly separated generality from superhuman performance. Early researchers expected machines would eventually surpass humans—transitioning from AGI to artificial superintelligence—but viewed human-level capability as an arbitrary benchmark, not the definition itself.

Recently, these dimensions have merged. Many now use AGI to mean either superintelligence or something involving consciousness—concepts the pioneers considered distinct. Intelligence level and generality represent independent axes: pocket calculators are superhuman but narrow, while early LLMs were general but barely articulate.

Beyond Binary Thinking

The definitional confusion suggests AI development may not produce a single AGI moment. Intelligence itself isn't monolithic, and different AI systems excel at different tasks. This fragmentation has practical implications: society can continue developing systems that help cure disease while restricting those enabling cyberattacks. The stumbling block for major challenges like climate change isn't knowing what to do but executing solutions—a problem AGI won't solve regardless of definition.

These details were first reported by George Musser in Scientific American.