Can AI Actually Come Up With New Ideas?

Ask the internet and you will hear that AI has crossed a line from repeating what it read to inventing things no human thought of. The claim shows up in product launches, in investor decks, and in the quiet fear that human creativity is about to be automated away. It is worth slowing down, because the evidence for genuine originality is thinner than the confidence around it.

Start with a test that was built to be hard to game. A group of researchers released a benchmark called Reconstruction in mid-August, posted to arXiv on August 17 and 19. The setup is clever. They give a model only the bibliography of a real research paper, the list of works it cited, and ask it to recover the paper's central idea. Crucially, they use a hard date cutoff and anonymized references so the model cannot simply have memorized the answer from its training data.

The results were humbling. Across 643 papers in six scientific fields, seven frontier models recovered the actual research idea only about 3 to 15 percent of the time. A more elaborate setup, where several models cross-checked each other in a tournament, lifted that to somewhere between 23 and 42 percent. Better, but still a long way from the picture of an AI that reliably dreams up the next breakthrough.

What the number does and doesn't say

Read carefully, this is a measure of one thing: can a model reconstruct an idea that already exists, working from the same references the human author had? If it struggles to rediscover known ideas under controlled conditions, the burden of proof sits heavily on anyone claiming it routinely produces genuinely new ones.

That does not mean models are useless at invention, and honesty cuts both ways here. There are real results where AI systems produced something that worked and was not obvious. Anthropic's Claude designed working protein binders for 14 of 15 targets, a genuinely useful piece of molecular design. In a blind test, readers preferred AI-written fiction to the human kind. Models are very good at recombining what they know into forms that are new to the person asking.

The gap is between novelty and originality. Generating a protein sequence by searching a vast space of possibilities is a real capability, and it is not the same thing as having an idea. Much of what looks like AI creativity is skilled recombination, fast and tireless and occasionally surprising, but still built from patterns already in the training data. That is valuable. It is not the same as the leap the hype describes.

So the fair verdict is a split one. Can AI produce outputs no human has produced before? Clearly, yes. Can it originate ideas the way a researcher does, from a hunch that turns out to be right? The best controlled evidence we have says not reliably, not yet. The honest position is to be impressed by what these systems combine, and skeptical of anyone selling the leap from combination to genuine invention as though it has already happened.