A different way to write about the future of AI
There are already countless articles about what AI will be able to do in five or ten years. Most of them are, at their core, an opinion with citations attached – a selection of impressive quotes from people who have their own stake in how their prediction lands. This article takes a different approach: it takes the question "how reliable are predictions about AI in the first place?" exactly as seriously as the question "what will AI be able to do?" – and grounds itself in research that has studied that reliability directly. The result is less tidy than a smooth forecast, but it's something you can actually plan against.
Why predictions about AI progress have historically missed so often
Overreaching AI predictions aren't a new phenomenon of the ChatGPT era. Psychologist and AI pioneer Herbert Simon said in 1965: "Machines will be capable, within twenty years, of doing any work a man can do." In 1970, Marvin Minsky told Life magazine: "In from three to eight years we will have a machine with the general intelligence of an average human being." Both were off by decades – not because they were incompetent (both rank among the most influential figures in AI's history), but because estimating the timing of a breakthrough works structurally differently from estimating its direction.
That could be dismissed as a historical footnote – if not for a much harder, much more recent piece of data. AI Impacts, one of the methodologically cleanest research groups on this topic, has surveyed thousands of active machine learning researchers for years on the expected timing of "human-level" AI. In the 2022 survey, the median estimate (50% probability) stood at 2060. One year later, in the 2023 survey of 2,778 researchers, it had jumped to 2047 – 13 years earlier, within twelve months.
The real insight isn't the number 2047 itself. It's the shift. A genuinely well-founded estimate, grounded in real knowledge, shouldn't move 13 years because of one year of new observations – unless the underlying uncertainty was enormous the whole time, and the number merely pretended to be more precise than it was. That instability is itself the finding. It tells you: anyone stating a confident date for an AI breakthrough today is selling you a certainty that even the world's best-informed researchers demonstrably don't possess.
A better test: what do the world's best forecasters say?
One could object that domain researchers are biased – too close to their own field, sometimes too euphoric, sometimes too cautious. That's what makes a second finding more interesting: the "Existential Risk Persuasion Tournament" research project and its successors pit AI domain experts directly against so-called "superforecasters" – people with no AI background, but a demonstrated, unusually strong track record on general forecasting questions (per psychologist Philip Tetlock's research).
In the most recent survey wave (2025/2026), both groups landed surprisingly close on whether "human-level" AI will exist before the year 2100: roughly 80% probability on both sides. Domain experts put the median at 2050, superforecasters at 2047 – a smaller gap than one might expect. On a more concrete, technically graspable question (whether an AI can autonomously complete an eight-hour work task, measured via the METR benchmark), the superforecasters were noticeably more optimistic: 2028 versus 2030 among the experts.
The genuinely remarkable part, though, isn't the gap between the two groups – it's the movement within each group. Between June 2025 and April 2026 – under a year – both camps shifted their estimates noticeably earlier, with the superforecasters moving even more than the experts. That means even the world's most demonstrably accurate, methodologically trained forecasters keep re-adjusting their AI estimate at short intervals. If even this group has no stable answer, a single CEO quote or keynote slide is no substitute – it's marketing with a date attached.
What can actually be predicted robustly: the cost curve
After all that well-founded scepticism toward timing predictions, it's worth looking at the one piece of data that has behaved remarkably reliably for years: not AI models' capabilities, but their cost. For a fixed level of capability (the same quality of answers, not whichever model happens to be newest), the cost per request has fallen roughly 280-fold since ChatGPT launched in late 2022, with a continuing trend toward roughly 1,000-fold over three years. The effect amounts to a cost halving roughly every six to eight months.
The difference from timing predictions is fundamental: cost reduction through efficiency gains, competitive pressure and better hardware utilisation is an engineering process made up of many small, independent contributions – not a single breakthrough you have to wait for. That's exactly why it can be extrapolated more reliably than "when does the next big capability arrive". For your own planning, this is the more important anchor: whatever AI can do today will, with high confidence, be available in 12 to 18 months for a fraction of today's cost – regardless of whether and when some speculative capability leap occurs.
The limit that became real in 2026: the data wall
A second, less-discussed factor is equally well-researched: the available amount of high-quality, human-written training text is finite. The research group Epoch AI puts the usable stock at roughly 300 trillion tokens and estimates, with 80% confidence, that this stock – depending on how aggressively models are trained on the same data multiple times – will be exhausted somewhere between 2026 and 2032.
This limit isn't a distant future problem; it explains a shift that has already happened. Between 2024 and 2025, visible progress moved noticeably away from pure "more data, more training compute" toward so-called reasoning models, which instead spend more compute at the moment of the request itself (which is why some models visibly "think" longer before answering). That's a direct response by the research community to a foreseeable resource limit – and a good example of how technical limits don't stop progress, they redirect it. For companies, the practical takeaway is: the next wave of improvement is likely to come less from "bigger" models than from models allowed to spend more of their own intermediate steps on the same request – with corresponding effects on response time and cost per request, not just raw quality.
The most underrated finding: AI doesn't lift everyone equally
Arguably the most practically valuable finding from AI research so far has nothing to do with "the future" in the sense of timing at all – it describes how AI already measurably works today, and the pattern is more likely to strengthen than disappear as models improve. A controlled field study by Harvard Business School and Boston Consulting Group with 758 consulting professionals found that AI assistance doesn't improve tasks evenly, but runs along a "jagged frontier": some tasks a language model solves surprisingly well, while seemingly near-identical neighbouring tasks the same model reliably fails – without that boundary being obvious to the people using it.
An even more revealing, independent finding: in a study of 5,179 customer-service agents, the weakest and least experienced staff improved by roughly 34% with AI assistance – while the already-strongest performers showed barely measurable improvement. This levelling effect (sometimes called "skill-levelling") has since resurfaced in multiple independent studies across different tasks. The practical consequence is uncomfortable for the usual sales narrative: AI doesn't primarily make your best people even better. It raises the floor. Whoever is looking for the biggest lever is more likely to find it among newer staff, standard tasks and widely varying quality than among top performers.
Why even economists don't agree on what this means economically
Even the magnitude of AI's future economic effect is openly contested among serious economists – not out of ignorance, but because the decisive assumptions can't yet be measured. MIT economist Daron Acemoglu expects a "non-trivial but modest" effect from generative AI over the coming decade: roughly 0.7% additional productivity and 1.1% additional GDP growth. Goldman Sachs, using a different model, arrives at roughly 1.5% annual productivity growth – a substantially larger effect.
The difference doesn't come from better or worse arithmetic; it comes from different assumptions. Acemoglu assumes AI can profitably automate only around 5% of work tasks in the near term and counts mostly efficiency gains within existing tasks. Goldman Sachs assumes around 25% and additionally counts effects that are hard to quantify in advance – such as workers shifting into new tasks that don't exist yet. An independent, already-measured (not projected) figure comes from economists at the Federal Reserve Bank of St. Louis: they estimate that generative AI has raised US labour productivity by up to 1.3% since ChatGPT's launch so far – real, but far from the biggest promises.
The lesson here isn't which number is "right". It's that the range of serious estimates among top economists is itself enormous – and that any single number someone states with confidence is one of several plausible assumptions, not a fact.
What this means for your company's planning
All of this points to a different way of planning than "wait for the next big AI moment" or "switch everything to AI right now". Four concrete conclusions follow directly from the well-evidenced findings above – not from speculation:
- Waiting has a real cost, but hesitating because the technology isn't "mature" yet is the wrong reason: the cost curve is the most reliable trend in this entire field. What works today will be available at a fraction of the cost in 12 to 18 months – that's an argument for an early, small entry to learn, not for "waiting until it's finished".
- Don't expect even benefits – deliberately look for the jagged frontier inside your own processes: a blanket "adopt AI" announcement runs into a different, invisible boundary in every company, between tasks that work well and visually similar tasks that don't. That can only be found task by task, not through a general rollout.
- Expect the biggest lever where the biggest variation already exists: the best-evidenced effect of AI is raising standard tasks and less experienced roles – not multiplying already-excellent work. Anyone looking for the biggest return should look first at processes with high quality variance, not at their own top performers' work.
- Treat any prediction with a firm date with structural scepticism: anyone giving you a fixed year for an AI breakthrough either has information even the world's most rigorously tested forecasters don't have – or is selling you a story. The second option is by far the more likely one.
That exact task-by-task search for your own jagged frontier – where AI already helps reliably, where it doesn't (yet), and where the levelling effect pays off the most – isn't a theoretical exercise. It's a concrete first piece of analysis, one that can be done in a manageable timeframe before larger budget decisions get made.