The great AI hype correction of 2025: four keys to understanding the reset in expectations

The MIT Technology Review article argues that 2025 has been the year the narrative bubble around generative AI —not necessarily the technology itself— began to deflate. Since OpenAI launched ChatGPT in late 2022, the sector has seen a steady escalation of ever more grandiose announcements (voice, image, video, reasoning), and executives at the major AI companies went so far as to promise the technology would replace office work, usher in an era of abundance and accelerate scientific and medical discovery. That fear of missing out (FOMO) pushed companies around the world, especially in the Global North, to rush into adopting AI without the results matching those expectations, according to several studies cited in the piece.

The episode that crystallized the change of mood was the launch of GPT-5 in August, preceded by months of hype from Sam Altman, who described it as a model with "PhD-level" ability in any subject and hinted, with an image of the Death Star from Star Wars, that something of colossal power was coming. When the model was finally unveiled, the general feeling was one of continuity rather than a qualitative leap. Researcher and YouTuber Yannic Kilcher declared that "the era of boundary-breaking advances is over" and that AGI is not arriving, comparing the current moment for large language models (LLMs) to the maturing phase of smartphones, where annual updates no longer surprise the general public even if the underlying technology remains relevant.

The piece offers four ways to read this readjustment without falling into the opposite extreme of writing the technology off. The first is that LLMs are not the gateway to artificial general intelligence (AGI). Even Ilya Sutskever, co-founder of Safe Superintelligence and former chief scientist at OpenAI, now acknowledges the limitations of the technology he helped create: in a November interview with Dwarkesh Patel, he explained that these models generalize far worse than people, because they learn to solve specific tasks without grasping the underlying principles. The article attributes much of the overselling of expectations to the fact that the models' fluent use of language primes us to perceive a human mind where there is only a very sophisticated statistical pattern.

The second takeaway is that AI is not a magic, instant fix for every business problem. An MIT study published in July, which became a central argument for skeptics, found that 95% of companies that had tried AI had got no value from it. The article itself qualifies that figure, however: it refers only to companies that implemented bespoke AI systems and had failed to scale them beyond the pilot phase within six months, and it does not include informal chatbot use by employees outside official projects —a kind of AI "shadow economy" that, according to that same MIT study, existed at around 90% of the companies surveyed, though its value was not measured. A November study by Upwork also showed that agents based on models from OpenAI, Google DeepMind and Anthropic failed at many simple work tasks when acting alone, but improved markedly when collaborating with people who knew what they were doing. That clashes with Altman's prediction that in 2025 AI agents would "join the workforce" and materially change companies' output. Researcher Andrej Karpathy, who coined the term "vibe coding", sums up the situation by noting that chatbots beat the average person at many tasks (basic legal advice, code debugging, high-school math) but not a human expert, which would explain why they have caught on with individual consumers without yet upending the labor economy.

The third takeaway addresses the debate over whether there is an AI bubble and, if so, what kind. The article contrasts two precedents: the 2008 subprime mortgage bubble, which left nothing of value behind when it burst except debt, and the dotcom bubble of 2000, which did leave internet infrastructure and companies that went on to become giants such as Google or Amazon. With today's AI it is unclear which of the two scenarios applies, partly because there is still no established business model for LLMs nor an obvious killer app, and because the scale of investment in data center infrastructure is a worry given uncertain future demand, compounded by circular financial deals (such as the payments between Nvidia and OpenAI) that raise suspicion. Against that pessimism, investor Glenn Hutchins, co-founder of Silver Lake Partners, argued on the Technology Business Programming Network podcast that almost every large data center already has a creditworthy counterparty contracted to absorb its output —citing Microsoft, with its credit rating, as an example of a guaranteed buyer. The article also recalls that after the dotcom crash the companies with enough capital to hold out survived, and points to Synthesia, an AI-generated avatar startup that investor Nathan Benaich (co-founder of Air Street Capital) initially did not see as viable, and which today has some 55,000 corporate customers, bills close to $150 million a year and was valued at $4 billion in October.

The fourth takeaway is a reminder that ChatGPT was neither an absolute starting point nor the end of the road: it is the culmination of a decade of advances in deep learning, whose roots go back to the 1980s and whose field as such has been taking shape since the 1950s. Far from stalling, academic AI research is in a moment of ferment: the sector's leading conferences had to reject more quality papers this year than they could handle, even as preprint repositories such as arXiv were flooded with low-quality AI-generated content. Sutskever describes the current moment as a return "to the age of research", not a dead end. Benaich argues that the excess of enthusiasm also has an upside, because it attracts money and talent: just two or three years ago the people building these models were, in his words, "amateur researchers who stumbled on something that worked", whereas now every serious professional in the tech sector is working on it.

The piece concludes that this correction of the hype was necessary and predictable, that the pendulum should not swing to the opposite extreme of dismissing the technology, and that alongside the genuine "wows" of recent years —leaps in video generation, reasoning models, wins in programming and math competitions, or tools such as Google DeepMind's Nano Banana Pro, capable of turning a book chapter into an infographic for free— sit real limitations, considerable economic and environmental costs, and enormous uncertainty over whether today's spending will pay off. AI, the article sums up, is not going away, but even what has already been built is not yet fully understood, let alone what comes next.