AI paradoxes give us a peek into its future
Where will AI be in 2030? For an informed view, we can unpack three paradoxes that have come to symbolise the uncertainty.
I cannot recall another time in my career when “paradoxes” were mentioned so often in meeting rooms. Demand moves the wrong way, productivity disappears, and the tech is both awesome and underwhelming. If we can’t explain basic stuff, how can we plan for it?
Paradoxes live in the fault lines between logic and language, between our common sense and reality. AI evolves so fast that its implications are surprising, even counterintuitive.
Beyond good discussion points, paradoxes are a good crystal ball for the near future. As trends move from counterintuitive to probable to evident, they put a spotlight on how expectations gradually adjust and how they, in turn, shape the trends.
How will AI adoption play out in the next few years? No one knows for sure. But these three frequently cited paradoxes might point us in the right direction.
Key ideas
Three paradoxes can help us see where AI goes next. Jevons says efficiency breeds demand: make knowledge work cheaper, and companies will invest more in it. Solow warns that the productivity payoff arrives late, perhaps exposing unsound economics. Moravec suggests that AI interacting with the physical world was the real prize all along.
Jevons
When AI was able to generate marketing text at scale, we all expected its value, and therefore the demand for it, to plummet. Instead, it skyrocketed, flooding the internet with slop and em-dashes. Why do we ask for more of something when we need less of it?
William Jevons coined his paradox while studying the steam engine. When Watt’s design made engines far more efficient, each one burned less coal for the same work. Demand for coal soared anyway, as cheaper power pushed industry to run more engines and find new uses for them.
Jevons’s paradox is at the heart of AI optimism. If we make human labour more efficient, companies will hire more, not less. Whilst the transition will undoubtedly be hard, Jevons gives us a useful playbook: use AI to expedite tedious work, allocate time to more valuable tasks, and companies will naturally invest in more jobs. The optimism holds only if the savings are real and arrive on time.
Solow
Anyone who uses AI daily can point to a myriad of annoying tasks that are now done faster. So why aren’t we measurably more productive? “You can see the AI age everywhere apart from productivity” is how Robert Solow might have put it if he were observing LLMs in 2026 rather than PCs in 1987.
The key to Solow’s productivity paradox is the prolonged period between the emergence of a technology and its implementation. The web only took off when it graduated from brochures and journals into the ecosystem of all organisational interactions. Similarly, AI productivity will not show in statistics until organisations use it to rewire their workflows and value chains.
This paradox is central to AI pessimism: transformation will move too slowly for Big Tech to recoup its massive CapEx investment. Revenues will falter, investors may get spooked, and markets will correct sharply. Jevons and Solow are locked in a race against time, but both assume that knowledge work is the main event.
Moravec
How is it that AI can solve seemingly impossible maths problems, but cannot match a toddler in basic physical tasks? Come to think of it, who asked for digital “agents”? We wanted flying cars and robot butlers!
Hans Moravec observed that high-level reasoning tasks are computationally cheap, but everyday sensory and motor skills require massive power. According to Yann LeCun and others, language models are a suboptimal technology for artificial intelligence. Disconnected from the physical world, LLMs cannot handle unpredictability and are therefore highly inefficient outside the comfortable parameters of the digital world.
New approaches like World Models propose a whole new approach to AI, in which a model can learn from physical examples and simulations. If they are successful, motor skills will be as cheap as generating text, paving the way for the era of physical AI and robotics.
Where will AI be in 2030?
How can AI be so clever and so stupid at the same time? If we finish tasks faster, why can’t we reallocate the time or measure the gains?
These paradoxes are prime examples of how AI moves faster than common sense. As confusion gives way to newly established perceptions, we get some clues to the state of AI in 2030.
As of summer 2026, Solow is beating Jevons. AI adoption is hitting one obstacle after another: turf wars, compliance and budgets are only the obvious ones. By 2030, we will see the first new workflows and truly novel apps begin to emerge. Whilst I think the transition can be managed, the current business models are built on unsound economics and are unsustainable.
By 2030, the world may realise that knowledge work is overrated and that real transformation arrives only when AI interacts with the physical world. World Models, Vision-Language-Action models, World Action models and others may replace chatbots and agents. In time, we might see LLMs as a mere enabling tech that allows robots to reason and interact with language.
No one can be sure how AI will play out. But unravelling the strands of what makes no sense today is a good place to start.
Reading list
The Messy Middle (Kinder Futures) – Molly Kinder’s map of the long, uneven jobs transition that sits beneath the Jevons optimism.
Bananas, cups and peelers: Robots learn how to handle curved objects (Tech Xplore) – the other half of Moravec, on why a toddler still out-handles a robot.
World Models: Enabling the next AI revolution (Yann LeCun, YouTube) – LeCun’s argument that language models are the wrong substrate for physical intelligence.
World Action Models: The Next Frontier in Embodied AI (arXiv) – the survey behind the closing bet that embodied models supersede chatbots and agents.
Vision Language Models Explained (Hugging Face) – a primer on the VLA lineage the conclusion points toward.
Let AI Burn (Where’s Your Ed At) – Ed Zitron’s sharpest statement of the Solow-driven pessimism: the economics may break before the payoff lands.
How long will the AI mania last? (The Economist) – the counterweight, on why the transition might still be managed.