AI can't see the future, but it now estimates the odds of real events close to the best human forecasters. What the 2026 research shows, and its limits.
By Kristian FagerlieFounder & CEO, Sikt Intelligence
Short answer: AI cannot know the future, but it can now estimate the odds of real-world events close to the level of the best human forecasters. Not with certainty, and not on every kind of question. But the gap between "AI that guesses" and "AI that forecasts well" has closed faster than almost anyone expected.
Here is what the research actually shows, where the limits are, and how to judge any claim that an AI "predicts the future".
The phrase "predict the future" suggests a yes or no answer: this will happen. Good forecasting never works that way. It gives a probability: this is 30% likely. That difference matters, because a probability can be checked.
If a forecaster says 30% on a hundred different questions, about thirty of those events should happen. That is called calibration, and it is the only honest way to judge anyone, human or machine, who claims to see ahead.
So the real question is not "can AI predict the future?" but: can AI give probabilities that are both calibrated and informative?
When researchers first tested large language models on questions that resolved after their training cut-off, most were barely better than guessing. Asked cold, without research tools, the models had no reliable way to reason about events they had never seen.
A team at UC Berkeley built a system that searched for news, reasoned over it and combined several forecasts. On questions from competitive forecasting platforms, it came close to the aggregated forecasts of the human crowd, and in some settings beat it (Halawi et al., 2024).
The same year, a study with 991 people found that forecasters who could consult an AI assistant were 24–28% more accurate than a control group (Schoenegger et al.). AI was already useful as a forecasting partner.
By mid-2026, the Forecasting Research Institute reported that the top AI systems on its ForecastBench benchmark were statistically indistinguishable from superforecasters (FRI). Good Judgment pushed back that the human numbers were collected two years earlier on different questions (Good Judgment), and on Metaculus, top human forecasters still lead the best bots head to head (Metaculus). That lead is shrinking fast: in Metaculus's spring 2026 tournament (January to April), ten pro forecasters beat the top ten bots by just 1.25 points per question, a gap too small to be statistically significant, down from about 20 points in the second quarter of 2025 (Metaculus, September 2026).
We cover that debate in AI vs. superforecasters: who predicts better?, how the main benchmarks work in AI forecasting benchmarks explained, and the whole field, weather included, in AI forecasting, explained.
Three changes did most of the work:
Together these turn a language model into what is now called an AI superforecaster.
We put this to the test ourselves. For one week in October, Sikt forecast eight stock, gold and oil markets, with $1,000 on every big disagreement with the market. Seven of ten bets lost, and it still finished ahead, thanks to one long shot. This five-minute explainer shows how the test worked; the full results are in our one-week AI stock market test.
The best way to judge an AI forecaster is to watch it forecast events that have not happened yet, and then check. Sikt Intelligence is doing exactly that on the 2026 US midterms:
A live AI forecast, scored in public. In the Sikt Midterm Bench, Sikt forecasts the 2026 midterms next to Kalshi and Polymarket. Latest round, Nov 3, 2026: Senate: Democrats 75% (markets then 76%), every round; House: Democrats 91% (markets then 94%), every round. Every forecast will be graded after election day, misses included.
Ask four questions:
If any answer is no, treat the claim as marketing.
It can estimate probabilities for clear, checkable events such as an election result by a fixed date, and some systems do this well (see can AI predict elections?). Short-term market moves are much closer to random, and no forecaster, human or AI, should present them as predictable; we look at the research in can AI predict the stock market?. Nothing here is investment advice.
Not reliably on its own. A chatbot answering from memory knows nothing about events after its training data ends, and early tests found models asked cold were close to chance. Systems built for forecasting add live research, combine several forecasts and check them against a track record, and those now come close to the best human forecasters on some benchmarks. Either way, the output is a probability, not a prophecy.
Better than the average person, clearly. Compared with the very best human forecasters, it is close, and the evidence is mixed: some benchmarks show parity, others still show top humans ahead.
Look for a published track record scored with a proper rule such as the Brier score, on questions that were in the future when the forecasts were made, including every miss.
The full record of Sikt's first election benchmark. 13 timestamped rounds on the House, the Senate, ten key Senate races and their margins, next to the markets, from September 30 to election morning.
Read the article
Sikt Bench tests how well AI models forecast real events, using only what was knowable at the time. The first preview ranks six models on 232 events.
Read the article
