Do AI and human forecasters make better predictions together than either does alone?

This explores whether combining LLM and human forecasts, either by averaging their predictions or by having people work with an LLM, produces more accurate predictions than either does alone. The short answer is that the collection has no study that directly pools the two, so it can't settle the question. What it does have are studies of the separate pieces: how good LLMs are alone, what happens when you pool many LLM judgments, and what happens when humans use LLMs as assistants. Those pieces point to a surprising answer: the most obvious way of combining them, putting an LLM in the loop with a human, doesn't automatically help.

Start with the LLMs alone. A retrieval-augmented system, meaning a language model that looks up current news before answering, forecast real-world events nearly as well as competitive human forecasters. It sometimes beat the human crowd, on questions published after its training data ended, so it couldn't have memorized the answers Can retrieval-augmented language models forecast like human experts?. In fields where experts only modestly beat chance, such as predicting which startup founders will succeed, LLMs already clear the human bar Can language models beat human venture capital experts?. In neuroscience, fine-tuned models beat experts at guessing which experimental results actually occurred. The same pattern-filling habit that causes hallucination when a model looks backward turns into real prediction when it looks forward Can LLMs predict novel scientific results better than experts?.

Pooling also works among LLMs. Individual LLM rankings of business strategies are noisy and biased, but averaging many of them produces rankings that look like expert consensus. The twist is that the gains come mostly from *more* evaluations, not from *different* models Can aggregated LLM rankings match human expert judgments?. That matters for human–LLM pooling. The classic case for mixing forecasters is that their errors differ and cancel out, yet LLM groups converge early, conform to each other, and surface less unique information than human groups do Do language model groups mimic human group reasoning patterns?. If LLMs bring little diversity of their own, human judgment may be exactly the independent signal a pool needs.

The hybrid evidence warns against the easy version, though. In a 348-person experiment, people who evaluated strategic decisions with LLM help considered a wider range of factors but predicted no more accurately. They also felt more overloaded and less ownership of their decisions Does using LLMs actually improve strategic decision making?. Assisting a human is not the same as pooling with one, because the LLM changes how the person thinks instead of adding a separate opinion that can be averaged in. A further problem is that LLMs are good at proposing possibilities but poor at judging how confident to be in them Can language models reliably judge their own candidate quality?, and well-judged confidence is exactly what a good pool needs from each member.

Taken together, the collection suggests the design question matters more than the yes-or-no question. LLM forecasting improves sharply when the workflow splits number-crunching from reasoning about context Can LLMs actually forecast time series better than we think?Can decomposing forecasting into stages unlock numerical and contextual reasoning?. By the same logic, the promising human–LLM setup is likely one that keeps the two independent and combines their outputs afterward, rather than one where the LLM sits in the human's ear. The open question is whether the human's independence survives once they've seen what the model thinks.