Is AGI real — or just good business?
When test scores, hype cycles, and shifting definitions blur the line between progress and profit.
Artificial General Intelligence — or AGI — sits at the heart of AI’s grandest promises. But scratch the surface, and the foundations begin to crack. Definitions shift, benchmarks mislead, and emerging abilities often vanish under scrutiny. And perhaps most telling of all: AGI is increasingly defined in terms that benefit business models, not science.
I have just made a preprint of an article titled “How Smart is Smart? A Critical Interdisciplinary Perspective on Artificial General Intelligence” that I am submitting to a scientific journal. Please download it from here. I’d love your thoughts, comments, and suggestions.
The paper looks at several limitations in our definitions and work with AGI. Based on this, here’s what we need to look at more closely.
We still don’t know what intelligence is
For over a century, psychology and neuroscience have debated the nature of intelligence. Is it one thing? Many things? Modular? Emergent? The brain isn’t a master algorithm — it’s a tangled, adaptive system where the same neural real estate can serve multiple functions. So when companies promise “general intelligence,” the first question should be: what exactly do they mean?
Without a clear or consistent definition, AGI becomes a shape-shifting target — always just around the corner, always defined in a way that makes current progress seem more impressive than it really is.
AI can pass tests — but doesn’t understand them
Benchmarks are seductive. It feels like intelligence when large language models outperform humans on bar exams or logic puzzles. But these successes often hinge on brute-force pattern recognition or statistical tricks — not comprehension.
A model can solve Ph.D.-level problems and still fail at counting letters in “strawberry.”
This is the test-passing illusion. Models optimize for outputs, not meaning. And in many cases, success comes down to exposure to similar test material during training. That’s not general intelligence — it’s high-performance mimicry.
Benchmarks don’t measure what they claim to.
Many AGI benchmarks aren’t psychometrically validated. Some have been included in training data. Others can be gamed through brute-force sampling or simple pattern-matching. Emergent behaviors — often cited as signs of growing generality — sometimes disappear with better metrics or continuous scoring.
What looks like a leap in ability may just be an artifact of how we measure progress.
AGI definitions now serve business goals
This may be the most overlooked point: the definition of AGI has changed to align with what today’s models can do — and what companies can sell.
Where AGI once meant human-level reasoning across any domain, it’s now framed as the ability to perform “most economically valuable” work. That reframe conveniently centers on office tasks, coding, and language — things current LLMs do well.
When AGI is defined as “whatever generates value,” it becomes more of a pitch deck than a scientific concept.
It’s no coincidence that OpenAI, Anthropic, and others use the AGI language just as they raise billions and sign enterprise deals. The mythology of AGI fuels valuation, even when the technology falls short of the rhetoric.
Final thought
The dream of AGI might one day come true. But right now, much of the conversation is shaped less by breakthroughs and more by branding. If we want to get serious about artificial general intelligence, we need more precise definitions, better benchmarks, and the humility to admit what current systems can’t yet do.
