By 2030, AI experts expect AI to score 75% on the FrontierMath benchmark; the public expects 50%

What the chart shows

The Forecasting Research Institute’s Longitudinal Expert AI Panel asks AI experts and members of the public to forecast specific milestones. For 2030, the median expert expects AI to score 75% on FrontierMath, a set of hard maths problems, against 50% for the public. Experts also expect more AI-assisted US work hours (18.2% vs 10%) and more US ride-hail trips in driverless cars (20% vs 12%).

Experts expect 30% of scientific papers to engage with AI by 2030, against 21.7% for the public. The one forecast here on which the public is higher is the chance of a ‘rapid’ general AI progress scenario: 23% against the experts’ 20%.

Key numbers

How to read it

• Median forecasts from the LEAP survey (Forecasting Research Institute, 2026) for questions resolving in 2030. Values are percentages: benchmark accuracy, shares, or probabilities.
• The data file holds 2030 and 2040 forecasts with no column for the year (rows 1–18 are 2030, rows 19–28 are 2040), and four measures appear in both. The chart shows the five measures forecast only for 2030. For 2040, experts’ medians are higher for daily AI companion use (30% vs 20%), AI-discovered drug sales (25% vs 15%) and a level-8 ‘technological Richter scale’ event (70% vs 40%).
• The AI Index text says the public predicts more electricity use by AI and a higher chance that AI solves a Millennium Prize problem; for 2030 the data shows 8% vs 7.4% and 27.2% vs 20.4%.