The previous pages showed how fast AI is improving and why some experts expect general intelligence perhaps within the coming decades. But the field is far from speaking with one voice on this. Many respected researchers — including people who built today's AI themselves — consider superintelligence distant, or argue that today's methods do not lead there. Their arguments deserve as much attention as the warnings, because they help judge what is real progress and what is just an impressive number.
Language models, they say, are not enough
Yann LeCun, one of the pioneers of neural networks and a recipient of the 2018 Turing Award (together with Geoffrey Hinton and Yoshua Bengio), has long argued that models trained mainly on text will not reach human-level intelligence. His argument starts from children and animals: neither a cat nor a toddler has read billions of sentences, yet both understand how things behave in the world — that a mug falls when you let go of it. According to LeCun, machines need so-called world models that learn from observation and can predict the consequences of actions. He set out his vision in A Path Towards Autonomous Machine Intelligence (2022).
The cognitive scientist Gary Marcus argues along similar lines. In 2018 he wrote down ten concerns about deep learning: it needs huge amounts of data, transfers what it has learned poorly to new situations and struggles with abstract rules. He does not deny its successes, but argues that reaching general intelligence requires supplementing it with other approaches, such as explicit work with symbols and rules.
Success at a narrow task is not a step towards generality
The computer scientist Melanie Mitchell, in her essay Why AI is Harder Than We Think (2021), describes four fallacies that, in her view, lead to AI progress being overestimated again and again:
- Narrow intelligence lies on the same path as general intelligence. When a machine masters chess or translation, it easily looks like a first step towards general intelligence. Mitchell points out that this “first step” has been announced many times in the history of the field.
- Easy things are easy and hard things are hard. In reality it is often the other way round: beating a chess champion was achieved before teaching a robot to reliably fold laundry. What people do without thinking tends to be hardest for machines.
- The lure of wishful names. When a test is called “reading comprehension”, it is tempting to conclude that a model that passes it understands text. But the name of a test is not proof of an ability.
- Intelligence is all in the brain. Human intelligence, Mitchell argues, is also bound up with the body, emotions and the culture we grow up in. Computation alone may not be enough to reproduce it.
Measuring how quickly a machine learns something new
François Chollet, author of the popular deep-learning library Keras, proposed a different view of what intelligence is in On the Measure of Intelligence (2019). In his view, what matters is not how many skills a system has — those can be “bought” with vast amounts of training data. What matters is how quickly, and from how few examples, it learns something entirely new. To test this he created ARC-AGI: a set of picture puzzles that people solve without much effort but that remain hard for AI. Anyone who wants to spot a real shift towards general intelligence should, according to Chollet, watch tasks like these, not records on tests that can be trained for.
Fluent speech is not understanding
The linguist Emily M. Bender, the computer scientist Timnit Gebru and their co-authors compared large language models in 2021 to stochastic parrots: they string words together according to how they occur together in texts, with no link to what they mean. Fluency, they argue, tempts people to credit models with understanding they do not have. The authors also highlighted problems that exist today and have nothing to do with superintelligence: biases taken over from training data, energy consumption and the concentration of power in a few large companies.
This is one of the main disputes in the field. Other researchers counter that models build useful internal representations of the world during training (see How AI works). Who is right has not been settled — and it also depends on what exactly we mean by “understanding”.
Technology is adopted slowly
The computer scientists Arvind Narayanan and Sayash Kapoor of Princeton propose seeing AI as “normal technology” (2025), similar to electricity or the internet. They distinguish three processes running at different speeds: the invention of a new method, the development of a useful application, and its actual adoption in people's work. The last is the slowest — companies, public bodies and people take years to adapt, and safety rules and legislation develop gradually. That is why, in their view, large economic impacts will come over decades, not suddenly. In their book AI Snake Oil (2024) the same authors show how often products that cannot do what they promise are sold under the AI label — especially systems meant to predict human behaviour.
The economist Daron Acemoglu, winner of the 2024 Nobel Prize in Economics, reached a similar conclusion. In The Simple Macroeconomics of AI (2024) he estimates that over the next ten years AI will raise the economy's total productivity by less than 0.53%. He does not deny that AI is useful for individual tasks — he shows how small a share of the whole economy those tasks make up so far.
Resources are not unlimited
Progress so far has also rested on models getting more and more text. Researchers at the organisation Epoch AI estimated that models will be trained on an amount equal to all the publicly available text written by humans sometime between 2026 and 2032. The authors themselves also list possible ways out: synthetic data, learning from other sources and more efficient training. So it is a limit the field will have to deal with, not necessarily the end of the road. Electricity and computing power play a similar role, as discussed on Benefits and risks.
Where sceptics and the worried agree
Sceptical voices do not say AI is harmless. Bender and Gebru focus mainly on harms happening today. Marcus publicly supports stricter regulation. The dispute is chiefly about pace and path: whether today's methods are enough, how quickly they will spread, and whether it makes sense to prepare for superintelligence or rather to tackle the concrete problems of current systems.
Both sides agree, though, that claims about AI need to be checked, not taken on trust. That is why When and how it might arrive sets out how to tell real progress from hype — and that applies to optimistic and sceptical predictions alike.