Citation: Csaszar, F. A., Peterson, A., and Wilde, D. (2026). The strategic foresight of LLMs: Evidence from a fully prospective venture tournament. arXiv:2602.01684.
Cite
Paper highlights
Strategy depends partly on judging uncertain outcomes before they occur. We tested whether LLMs could do this by ranking the fundraising prospects of live technology ventures. The best LLM correctly ordered about 79% of venture pairs; the best expert reached 67%. Neither crowds nor the human–AI combinations we tested improved on the best standalone LLM.
This is prospective evidence, collected before outcomes were known, that frontier LLMs can outperform experienced managers and MBA-trained investors on a bounded strategic-foresight task.
Study design
- Thirty U.S. Kickstarter technology ventures launched after every model’s knowledge cutoff.
- LLMs and humans assessed them while fundraising was still in progress; realized funds raised supplied the ground truth.
- Each LLM completed 870 comparisons. Human groups included 346 experienced managers and three MBA-trained investors.
- We compared standalone LLMs, individual humans, crowds, and human–AI teams.
Results at a glance
- Gemini 2.5 Pro correctly ordered about 79% of venture pairs, and GPT-5 Mini reached about 78%. Random choice would produce 50%.
- The strongest MBA-trained investor reached 67% pairwise accuracy. The aggregated Prolific crowd reached 56%, and another expert reached about 52%; the latter two were not statistically distinguishable from chance.
- Several LLMs captured more than 83% of the funds in the true top five ventures; the best human captured 50%.
- Human-only, LLM-only, and human–AI ensembles generally fell below the best standalone LLM. Combining the best human with the best LLM also reduced performance.
Why it matters
The tournament shows how to test strategic foresight prospectively: select live decisions, hold information constant, record predictions before outcomes, and score every evaluator against the same realized criterion. It also identifies an augmentation trap: adding a weaker human judgment can pull a strong LLM forecast away from the better signal. Human contributions may matter more in framing the decision, supplying proprietary information, checking applicability, and acting on the result.
How to use this paper
Cite this for
- A fully prospective benchmark comparing LLMs, managers, investors, crowds, and human–AI teams on the same strategic-foresight task.
- Evidence that frontier LLMs can outperform experienced human evaluators when ranking uncertain venture outcomes from standardized information.
- The augmentation trap: simple human–AI aggregation can reduce forecast quality when it pulls a strong model toward weaker human judgment.
Useful for teaching
- Why prospective tests matter when evaluating AI in strategy.
- How to design a fair comparison between human judgment, crowd judgment, model judgment, and hybrid judgment.
- Where human work remains in AI-assisted strategy: framing the question, curating information, checking applicability, and acting responsibly on the forecast.
Careful claim
In one live venture-ranking tournament with standardized information and realized fundraising outcomes, frontier LLMs outperformed the tested human groups; the result is strongest as evidence for prospective benchmarking and bounded AI-assisted evaluation tasks.
Abstract
Can artificial intelligence outperform humans at strategic foresight—the capacity to form accurate judgments about uncertain, high-stakes outcomes before they unfold? We address this question through a fully prospective prediction tournament using live Kickstarter crowdfunding projects. Thirty U.S.-based technology ventures, launched after the training cutoffs of all models studied, were evaluated while fundraising remained in progress and outcomes were unknown. A diverse suite of frontier and open-weight large language models (LLMs) completed 870 pairwise comparisons, producing complete rankings of predicted fundraising success. We benchmarked these forecasts against 346 experienced managers recruited via Prolific and three MBA-trained investors working under monitored conditions.
The results are striking: human evaluators achieved rank correlations with actual outcomes between 0.04 and 0.45, while several frontier LLMs exceeded 0.60, with the best (Gemini 2.5 Pro) reaching 0.74—correctly ordering nearly four of every five venture pairs. These differences persist across multiple performance metrics and robustness checks. Neither wisdom-of-the-crowd ensembles nor human–AI hybrid teams outperformed the best standalone model.
Public discussion
- Can AI outperform humans at strategic foresight? explains the design and results for a general audience.
- Join the discussion on LinkedIn.
- Michigan Public’s Stateside interviewed me about what the result does—and does not—say about AI and human judgment.
- Humans + AI with Ross Dawson is a longer conversation about venture evaluation, foresight, and human–AI collaboration.
Other links
Last updated 2026-06-21