Why predictions about AI progress have historically missed so often
Overconfident AI predictions aren't a new phenomenon. AI pioneer Herbert Simon said in 1965 that machines would be capable of doing any work a human could do within twenty years. Marvin Minsky said in 1970 that a machine with average human intelligence would exist within three to eight years. Both were off by decades - not from incompetence, but because estimating the timing of a breakthrough works structurally differently than estimating its direction.
A hard, current data point: the researcher survey
AI Impacts has surveyed thousands of active machine learning researchers for years on the expected timing of "human-level" AI. In 2022, the median estimate was 2060. One year later, in 2023, it had jumped to 2047 - 13 years earlier, within twelve months. The insight isn't in the number 2047, it's in the shift itself: a serious estimate grounded in real knowledge shouldn't move by 13 years from one year of new observations.
Even the world's best forecasters keep adjusting
You could argue domain researchers are biased. That's why a second finding is more interesting: tournaments pit AI experts directly against trained "superforecasters" - people without an AI background but a demonstrated, exceptional track record on general future questions. Within a few months, both groups shifted their estimates noticeably earlier, with the superforecasters moving even more than the experts. If even this group has no stable answer, a single quote or keynote slide is no substitute for one - it's a story with a date attached.
What can actually be predicted reliably: the cost curve
One data point has behaved remarkably reliably for years: not model capabilities, but their cost. For a fixed performance level, the cost per request has fallen massively since late 2022, roughly halving every six to eight months. Cost reduction through efficiency gains and competitive pressure is an engineering process with many small, independent contributions - not a single breakthrough to wait for. That makes it a more reliable anchor for your own planning than any capability date.
The most underrated finding: AI doesn't lift everyone equally
A controlled study with thousands of customer service employees found that AI support measurably improved the weakest and least experienced staff, while the already-strongest employees barely benefited. This leveling effect shows up again across several independent studies. The practical consequence: AI doesn't primarily make your best people even better - it raises the floor. The biggest lever tends to sit with newer staff and standard tasks, not top performers.
Why this matters for you as a decision-maker
Meet any fixed date with skepticism, instead of waiting for a "mature" moment - the cost curve argues for an early, small start to learn, not for waiting. And look for the biggest lever where your own processes show the most quality variance today, not with your already-strongest people. The full research, further evidence, and the economic framing are in the article "The future of AI: what can seriously be predicted - and what can't". What these principles mean concretely for your business is what "What this means for your business: the AI roadmap for decision-makers" pulls together at the end of this course.