I Asked ChatGPT for Studies and Got an Essay. I Asked Elicit and Got Fifty of Them.

The task was modest by review standards: find whether interventions of type X improve outcome Y, with actual studies I could cite. I started where everyone starts — ChatGPT — and got a beautifully written answer with a reading list that was part real, part ghost.

The ghost problem is the whole story. ChatGPT's citations looked right, sounded right, and a third of them didn't survive a DOI lookup. The summaries were excellent; the underlying scholarship was a coin flip. For brainstorming, great. For a bibliography, disqualifying.

Elicit flipped the shape of the work. Instead of prose about the literature, it returned a table — fifty actual studies, each row with the method, sample size, and the extracted finding, in one screen. The paper I eventually built the argument around was one I'd never have found through conversation, because I didn't know to ask for it.

I still use ChatGPT daily and this isn't a takedown — it's a scope correction. ChatGPT is the colleague who explains the field at lunch. Elicit is the research assistant who actually pulls the papers. Confusing the second job for the first is how fabricated citations end up in print.