Why we built PsiPi

Researchers increasingly lean on generative AI to help write their papers — and sometimes to draft a manuscript almost end to end. That got us curious about the other side of the desk: what if the journal used AI too? If the editor and the reviewers were themselves language models, what would come back? PsiPi started as our attempt to actually see it happen — a curiosity experiment turned into a tool.

And yes, the name is a joke. “PsiPi” (ΨΠ) is a play on “SciFi”— a wink at the fact that an AI reviewer, like good science fiction, can say things that sound authoritative but are really just plausible-shaped invention. Keep that in mind and the tool is genuinely useful; forget it and you'll take a confident hallucination at face value.

Getting real peer review on a manuscript takes months. Most researchers only find out their paper has a fixable problem — a missing control, an overclaimed conclusion, a reproducibility gap — after sending it out and waiting. PsiPi compresses that feedback loop from months to minutes,so you can catch the obvious stuff yourself first. It's purpose-built for the natural sciences— physics, chemistry, biology, earth & space science, and adjacent fields.

You paste your own Gemini API key and your preprint. Three independent AI reviewer personas — a methodologist, a domain expert, and a reproducibility reviewer — read it in parallel, each scoring it against the same five criteria a real referee report asks about. An AI editor persona then synthesizes their reviews into a decision-letter-style document you can download as a PDF. Everything runs directly in your browser; there is no PsiPi server, and nothing you submit is stored anywhere by us.

What it's useful for: a fast, specific, narrative first pass — the kind of read a colleague might give your draft over coffee, at the depth of an actual referee report — before you send it somewhere that matters. Because the reviews are written by language models, they can be wrong, incomplete, or hallucinated, so treat them as a starting point for revision, not a final verdict. See example output on real, already-published preprints to get a sense of what comes back.