Why the AI Industry is Quietly Redefining Artificial General Intelligence

When the first chess computers defeated human grandmasters in the 1990s, the definition of machine intelligence simply moved down the hall to the Go board. But as frontier artificial intelligence models begin executing multi-hour software engineering tasks autonomously, the definition of Artificial General Intelligence (AGI) is not just moving—it is fracturing entirely. The difference this time is that the goalpost is no longer a board game with fixed rules; it is the entirety of the global knowledge economy.[7]

The catalyst for this fracture arrived in early September 2026, when OpenAI launched its GPT-6 Astra model. During the press briefing, OpenAI President Greg Brockman declared, "Welcome to the AGI era," pitching the system as a generational leap capable of operating software and filling out forms autonomously.[1]

Yet, just days before that launch, OpenAI CEO Sam Altman reportedly dismissed the very acronym his company was founded to achieve, calling AGI "a very poorly defined term" and "an irrelevant marketing term." That contradiction captures the current state of the artificial intelligence industry, as reported by Bloomberg on September 7: the leaders building the technology no longer agree on what the finish line actually looks like.[1]

To understand the shift, one must look at what is actually shipping versus what was historically promised. For decades, computer scientists defined AGI as a theoretical machine matching or surpassing human cognition across every conceivable domain—a concept now often labeled "Full AGI."[7]

What OpenAI and others are currently deploying is increasingly referred to as "Functional AGI." This narrower definition is anchored in economics rather than philosophy. OpenAI's own charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work."[1]

Under that economic framework, the capability threshold looks different. Astra, for instance, is designed to operate more like an agent than a chatbot, carrying out sequences of actions across different environments, such as executing database migrations or formatting documents without step-by-step human prompts.[1]

Anthropic, a primary rival to OpenAI, has similarly distanced itself from the traditional AGI framing. In a January 2026 interview, Anthropic President Daniela Amodei called AGI an increasingly outdated concept. "Many years ago, it was kind of a useful concept to say, 'When will artificial intelligence be as capable as a human?'" Amodei said.[2]

She noted that the framing is breaking down because models like Anthropic's Claude already surpass average human capabilities in specific domains, such as software development, while simultaneously failing at tasks humans handle effortlessly. "By some definitions of that, we've already surpassed that," she added, highlighting the jagged capability profile of modern large language models.[2]

To impose order on this definitional chaos, Google DeepMind researchers published a 57-page paper on June 10, 2026, titled "From AGI to ASI." The researchers, including DeepMind co-founder Shane Legg, proposed a rigorous framework that abandons the binary idea of AGI in favor of a continuum toward Artificial Superintelligence (ASI).[3]

Under the DeepMind taxonomy, a "competent" AGI outperforms 50 percent of skilled adults across a wide range of non-physical tasks, while a "superhuman" system hits a 100 percent threshold. To put this theory into practice, Google launched a $200,000 Kaggle hackathon in March 2026 to crowdsource evaluations for cognitive abilities like metacognition and social cognition.[4][6]

The urgency to define these metrics stems from rapidly compressing timelines. A March 24, 2026, report from the RAND Corporation synthesized expert forecasts and found that estimates for achieving AGI have shifted substantially from mid-century toward the near term, with some projections landing in the late 2020s.[5]

"Decisionmakers are making decisions based on methodologies that are in nascent stages of development," the RAND authors noted, warning that the field lacks benchmarks resistant to saturation and gaming. The report emphasized that while definitional ambiguity drives some disagreement, fundamental debates remain over whether current architectures can truly generalize.[5]

That architectural debate is the core of the skepticism surrounding "Functional AGI." Critics argue that systems built purely on language models lack the flexible, autonomous reasoning required to invent genuinely new things. While OpenAI Chief Research Officer Mark Chen claimed in August 2026 that the company is "80% of the way" to AGI, skeptics view such percentages as arbitrary marketing for systems that still rely on pattern matching.[1][7]

Ultimately, the industry is trading the philosophical pursuit of a digital human for the pragmatic deployment of a digital workforce. Whether a system possesses true self-understanding is becoming less relevant to developers than whether it can successfully execute a 40-hour software engineering ticket. The definition of AGI has not been solved; it has simply been redefined to match what the industry knows how to sell.[7]