AI Progress Outpacing Expert Forecasts, Study Finds

What happened

Leading AI experts have consistently underestimated how quickly artificial intelligence is advancing, according to a study from the Forecasting Research Institute, as reported by The Decoder.

The clearest example is mathematics. AI systems reached gold-medal level at the International Mathematical Olympiad five years ahead of the median expert forecast. Commercial momentum has also outpaced expectations, with Anthropic's annualised revenue reported at about five times what experts had predicted.

The picture is more mixed for real-world applications. The study's forecasts for areas such as self-driving cars are described as less accurate, which suggests that progress in digital, rules-based domains has moved faster than progress in physical deployment.

Why it matters

Most organisations build AI roadmaps around consensus timelines drawn from expert forecasts. If those forecasts have repeatedly lagged actual capability, plans that assume gradual adoption may be overtaken quickly. For leaders in digital transformation, the practical risk is a roadmap that is too slow for what the technology can already do, or one that commits to fixed milestones before the underlying capability is clear.

The uneven results for physical applications also matter for service design. Language, analysis and planning tools can be introduced into service flows relatively quickly, while autonomous or physical systems depend on regulation, infrastructure and safety validation. Journeys that mix digital and physical touchpoints will need to be designed around that gap, not around a single adoption curve.

By the numbers

• Five years ahead of the median expert forecast for AI reaching gold-medal level at the International Mathematical Olympiad.
• About five times the revenue experts predicted for Anthropic, based on annualised revenue.

The Renascence take

The headline is about forecasting, but the real lesson is about how organisations decide when to act. Expert timelines are a poor basis for customer commitments, and the study suggests they are consistently too conservative.

Stop planning AI around when experts think it will arrive and start planning around what your customers can already be served by today. Build small, reversible experiments into live journeys, and measure them against real customer outcomes rather than forecasts. The uneven progress in physical applications is a reminder that the hardest part of service design is often the handoff between a fast digital capability and a slow physical reality. Customer-obsessed operators will design those handoffs deliberately.