🎉
Revolution wins Sustainable Business of the Year 2026!
Revolution wins Sustainable Business of the Year 2026!
AGI is everywhere. The phrase appears in pitch decks, in funding announcements, in newspaper headlines, and in the mouths of CEOs whose companies happen to sell AI products. Predictions vary wildly. Sam Altman has suggested superintelligence might arrive “in a few thousand days”. Dario Amodei expects AI to wipe out half of entry-level white-collar jobs within five years. Daron Acemoglu, who won the 2024 Nobel Prize in Economics, estimates AI will add less than 1% to GDP over a decade.
The gap between these forecasts is too wide to be a difference of opinion. What it really shows is that AGI works as a story more than a stable technical concept. The version you hear depends on who is telling it and what they are selling.
Try to define AGI, and the first problem appears immediately. There is no agreed-upon definition.
OpenAI’s official charter describes AGI as “highly autonomous systems that outperform humans at most economically valuable work.” Sam Altman has used several other definitions over the years, in Moore’s Law for Everything, in Planning for AGI and beyond, and in The Intelligence Age. Sometimes economic, sometimes cognitive, often vague.
The clearest definition turned out to be contractual. According to reporting by The Verge, OpenAI and Microsoft agreed that AGI is achieved when OpenAI’s systems generate $100 billion in profits. That is the threshold written into the contract. No Turing tests. No benchmarks. Profit.
Researchers have taken notice. Melanie Mitchell argued in Science that the term has been used so inconsistently it has lost most of its scientific meaning. Gary Marcus has made similar arguments for years. The vagueness functions as a feature. A moving target is easier to claim victory over.
The skeptical view about current AI models reaching AGI is not coming from outsiders.
Yann LeCun, who leads AI research at Meta, has repeatedly argued that large language models will never reach human-level intelligence. His position is that current models cannot reason, plan, or understand the physical world. He has proposed an entirely different architecture (called JEPA) that he believes is required for real progress.
François Chollet, creator of the ARC-AGI benchmark, has consistently shown that LLMs perform poorly on tasks requiring abstract reasoning rather than pattern matching. His 2019 paper On the Measure of Intelligence remains one of the more rigorous attempts to define what general intelligence would actually require. In early 2025 he left Google to co-found Ndea, a research lab built on the explicit premise that LLMs are not the path to general intelligence.
Subbarao Kambhampati at Arizona State has published extensive research showing that LLMs cannot reliably plan or reason, even when their outputs look like reasoning. Melanie Mitchell at the Santa Fe Institute has made similar arguments.
In mid-2025, Apple’s research lab added a notable contribution. Their paper “The Illusion of Thinking” tested OpenAI’s o3 and Anthropic’s Claude 3.7 Sonnet on logic puzzles of increasing complexity. As complexity grew, accuracy collapsed toward zero, even when the models were given the solution algorithm directly in the prompt. The paper has been contested by follow-up work, but the core finding sits awkwardly with the AGI timeline narrative being sold by the same companies whose models were tested.
The interesting thing is not that AI experts disagree. The interesting thing is that the loudest predictions come from the people selling the product, and they tell two opposite stories at once.
OpenAI tells the wizard’s story. AGI will cure diseases, solve climate change, end scarcity. Altman’s The Intelligence Age frames the coming decade as one in which AI delivers prosperity at scale. The pitch is optimistic, expansive, and excellent for raising money from investors who want to be on the side of progress.
Anthropic tells the prophet’s story. Dario Amodei wrote an essay called Machines of Loving Grace describing a “compressed 21st century” in which AI cures most diseases and doubles human lifespans. Then in May 2025, he told Axios that AI could eliminate up to 50% of entry-level white-collar jobs, pushing unemployment to 10–20% within five years.
Both stances sell the same product. The optimistic frame attracts users, partners, and government goodwill. The catastrophist frame attracts safety funding, regulatory engagement, and the kind of attention that says “our technology is so powerful you must take us seriously.”
Peter Thiel, whose Founders Fund was an early OpenAI backer and whose Palantir is one of the largest beneficiaries of the AI boom, has been one of the more interesting sceptics. In a 2025 interview with Ross Douthat, he argued that AGI talk often functions as cover for actual technological stagnation. The more loudly we discuss future breakthroughs, the less we have to confront the absence of present ones. Coming from someone profiting from AI commercially, the position is harder to dismiss.
The clearest signal that the labour-replacement narrative deserves scepticism is operational. The same companies forecasting mass white-collar displacement are hiring tens of thousands of credentialed professionals to manually train their models.
Mercor, Scale AI’s Outlier, Surge AI, and Invisible Technologies recruit lawyers, doctors, PhDs, senior engineers, and professional writers to produce the training data and reinforcement signals that frontier models depend on. The work is sometimes called “ghost work” and has been widely documented in the New York Times, Wired, MIT Tech Review, and Bloomberg.
So the message to white-collar workers is: AI is coming for your job. The actual market behaviour is: please come and teach the AI to do your job, by the hour. That contradiction rarely appears in the same article.
None of this dismisses AI as a tool. In the actual present, AI is genuinely useful for specific tasks.
Research from MIT’s Erik Brynjolfsson and colleagues found a 14% average productivity gain among customer service workers using GenAI assistance. The detail that matters: bottom-tercile workers, those with the least experience, gained around 35%. Top-tercile workers gained close to nothing. The AI essentially extracted expert tacit knowledge and surfaced it to novices.
A follow-up study from Harvard and BCG produced what the authors called the “jagged frontier” finding. For tasks inside the AI’s effective zone, consultants saw a 40% quality improvement. For tasks just outside it, quality went down. The model was confidently wrong, and humans tended to trust it anyway.
For digital marketing, this maps directly to practical experience. AI is reliable for first-draft copy, schema markup, audience segmentation analysis, structured data tasks, and generating variations for A/B testing. It is unreliable for strategy, original creative direction, anything that requires reading market context the model has not seen, and anything where being confidently wrong costs more than being slow.
The right framing is that AI works as a power tool. It accelerates workers who already know the job. The judgment of knowing what to do still belongs to humans.
A second-order question is what AI does to its users over time. The research here is recent and not yet conclusive, but it is converging.
A working paper on AI assistance and persistence found that participants given AI help on a first task were more likely to give up sooner on later tasks attempted without help. A Microsoft and Carnegie Mellon survey of 319 knowledge workers, published at CHI 2025, found that higher confidence in GenAI was associated with lower self-reported critical thinking. An MIT Media Lab study using EEG measurements observed reduced neural connectivity and weaker memory recall in participants who wrote essays with ChatGPT compared to those who wrote unaided.
The broader pattern is not new. The foundational Science paper on the “Google effect” showed in 2011 that people remember less information they expect to be able to look up later. AI is the same dynamic at a much higher resolution.
This matters for anyone building a team. A workflow that produces faster work in the short term while gradually reducing the staff’s capacity to think through problems independently is a tradeoff that should be made deliberately rather than by accident.
The most grounded estimate of AI’s economic impact comes from someone with nothing to sell. Daron Acemoglu of MIT, the 2024 Nobel laureate in economics, published a paper called The Simple Macroeconomics of AI using a standard task-based framework. His central numbers: AI will add 0.53% to total factor productivity and 0.90% to GDP, cumulatively, over ten years. He estimates that less than 5% of tasks will be both technically and cost-effectively automated by AI in that period.
These figures are roughly an order of magnitude smaller than the projections coming from McKinsey, Goldman Sachs’s industry analysts, and AI company leadership.
Acemoglu is not alone among mainstream sceptics. Goldman Sachs’s own head of global equity research, Jim Covello, has openly questioned whether AI can deliver returns commensurate with the roughly $1 trillion in capital expenditure the buildout is expected to require.
Where does this leave a business owner trying to make sensible decisions?
AGI may eventually be answered as a technical question, though not on the timeline its loudest promoters suggest. The technical experts who would benefit most from it being imminent are publishing papers arguing that it isn’t. The largest economic study from a Nobel-winning economist says the productivity gains will be modest. The companies marketing AGI as a job-replacing force are simultaneously hiring professionals by the thousand to train their models.
Treating AI as the most useful new tool in a decade is correct. Treating AI as a force that will replace your team next year is a marketing claim. The first one informs how agencies are integrating AI into client work. The second is the noise around the signal.
The work of distinguishing the two is, as it has always been, the actual job.
