Elicit.

The reading has been automated

Somewhere right now, a PhD student is staring at a search that returned 4,300 papers. She needs to read every abstract, decide which 60 are relevant, and pull a specific number out of each one. The old plan was three months. The new plan is to ask Elicit.

Elicit is an Oakland-based AI platform that does the slow part of science - searching, screening, and extracting data from academic papers - at a scale no human team can match. It sits on a corpus of more than 125 million papers, plugs into PubMed and ClinicalTrials.gov, and answers questions with tables instead of vibes. Every claim comes with the quote it came from.

Two million researchers use it. Pharma firms that advise the biggest drug companies on earth use it. And it started life inside a nonprofit that was, frankly, worried about what AI would otherwise be used for.

Science knows a lot. Finding out what is the hard part.

Here is the uncomfortable thing about modern research: the answer to your question probably already exists. It is sitting in a paper published in 2014, in a journal you have never heard of, behind a figure caption nobody indexed. The bottleneck is not curiosity. It is reading speed.

A proper systematic review - the gold standard of evidence synthesis - can take a year and a small team. Screening thousands of abstracts. Extracting hundreds of data points by hand. Reconciling two reviewers who disagree. It is careful, important, and almost comically slow. The volume of published research keeps doubling. The number of hours in a researcher's day, stubbornly, does not.

So the field had a choice: read less, or read differently.

Two people who thought reasoning should be abundant

Andreas Stuhlmuller started thinking about how to automate reasoning as a teenager - a hobby most people grow out of. He didn't. He published his first paper in 2010, earned a PhD in brain and cognitive sciences at MIT, did a postdoc at Stanford, and in 2017 founded a nonprofit called Ought to work on the question of how machines could reason carefully rather than just confidently.

Jungwon Byun joined in 2019, trading a Head of Growth role at the lending company Upstart for a research lab with no product and a very long time horizon. Her bet was simple and slightly contrarian: building a real product, used by real people, was a more direct path to impact than writing papers about impact.

Their shared thesis: the world would be measurably better if high-quality reasoning were abundant rather than scarce. AI was the lever. Researchers - people who reason for a living and feel the pain most acutely - were the place to start.

A research assistant that shows its work

The trick with AI and science is that confidently wrong is worse than slow. A chatbot that invents a plausible-sounding citation is not a research tool; it is a liability with good grammar. Elicit's whole design answers this objection before you raise it: every extraction is backed by a quote or a figure from the actual source. You can click through and check. It is reproducible, traceable, and auditable at every step.

That last part matters more than it sounds. It is the difference between a tool a curious grad student plays with and a tool a regulated pharma team is willing to put its name on.

Numbers, because researchers ask for them

Claims of accuracy are easy. Elicit went and tested itself against 994 Cochrane reviews - the most rigorous evidence syntheses in medicine - and published the scores. Skeptics, this section is for you.

The real-world numbers hold up too. In a systematic review run by VDI/VDE to inform German education policy, Elicit correctly extracted 1,502 of 1,511 data points - a 99.4% hit rate. Formation Bio used it to map how end-stage knee osteoarthritis had been defined across decades of trials. Oxford PharmaGenesis, which advises 8 of the top 10 global pharma companies, uses it to deliver literature reviews at a scale that used to be impossible.

Built to augment judgment, not replace it

Elicit is a public benefit corporation, which is a legal way of saying the mission is written into the paperwork, not just the pitch deck. The roots run back through Ought into the AI-safety community, where the operating worry is not that AI will be too weak but that it will be confidently, scalably wrong.

So the product is deliberately the opposite of a magic-answer machine. It hands you the evidence and the sources and expects you to think. The goal was never to take the researcher out of research. It was to take the tedium out, and leave the judgment where it belongs.

For a company that could have built yet another chatbot, choosing to build a tool that argues for being checked is a quietly radical decision.

Back to that 4,300-paper search

Return to the PhD student. The pile of 4,300 papers has not gotten smaller - if anything, by the time she finishes her thesis it will have doubled. What has changed is what that pile means. It used to be a wall. Now it is a query.

That is the bet Elicit is making about the next decade: that the rate-limiting step in human knowledge is not how fast we can produce research, but how fast we can absorb what we already have. If clinical decisions, policy choices, and the next experiment can all rest on more of the existing evidence rather than less, the compounding effect is hard to overstate.

The reading has been automated. The thinking, pointedly, has not. That was always the plan.

Interviews & demos