Mimir is now in open beta.
Ask a materials science question, if the answer is in the literature, Mimir can find it.
Free for .edu & .gov emails
Try it out with the link below.
Tomorrow we're opening beta access to Mimir, an AI research tool for materials science that only cites real papers from reputable journals.
Ask a question, get an answer with citations from over 1 million papers.
There are fabricated citations to catch, the ones that don't show what they're cited for, and the methods section is missing the detail you need.
The literature keeps growing while the day stays the same length, so the tax keeps rising.
Many tools have gotten better at finding papers faster but almost none have touched the reading and verifying underneath, which is the part that actually takes much time.
The sharp rise in fake references doesn't line up with ChatGPT's launch, it starts about 18 months later.
The authors tie it to a specific shift, the arrival of AI search and agentic research assistants that generate citations automatically from live web results.
The fabrication is increasingly produced by tools built to help with research, sitting inside papers that look completely normal, which makes it both harder to spot and harder to stop.
The sharp rise in fake references doesn't line up with ChatGPT's launch, it starts about 18 months later.
The authors tie it to a specific shift, the arrival of AI search and agentic research assistants that generate citations automatically from live web results.
The fabrication is increasingly produced by tools built to help with research, sitting inside papers that look completely normal, which makes it both harder to spot and harder to stop.
Every researcher has done this:
Asked ChatGPT for verified references on a technical question. Got 10 papers that looked perfect. Real-sounding titles. Real-sounding authors. Real-sounding journals.
Then you tried to verify them.
6 don't exist. 2 have the wrong authors. 1 is from a journal that was never published. The last one is real but says the opposite of what ChatGPT claimed.
You just wasted 45 minutes fact-checking an AI that confidently lied to you. And you're back to doing the literature review manually.
This is the single biggest problem with using AI for research. The models are trained to sound right, not to be right. They generate plausible-looking citations the same way they generate plausible-looking code. Sometimes it works. Sometimes it's fabricated from scratch.
Someone finally built a tool where this is structurally impossible.
Mimir searches millions of real scientific papers. It reads them. It understands your question in plain language. And it returns a cited answer where every single reference is a verified, real publication.
Not "probably real." Not "we tried to check." Architecturally impossible to hallucinate.
The system indexes actual papers, books, patents, and industry reports. It can only cite what exists in its corpus. If a paper isn't real, it can't appear in your results. The architecture won't allow it.
Ask a complex technical question. Get an answer with 10+ verified references in seconds. Each citation links to the actual paper. Real title. Real authors. Real journal. Real data.
What it covers right now: materials science, chemistry, physics, engineering. Expanding into new domains continuously.
What used to happen: 2 days reading papers, cross-referencing citations, verifying sources, discovering half your references are garbage.
What happens now: ask a question. Get a cited answer. Every reference is real. Move on.
Founded by a materials scientist and a computer scientist who spent too many years on the wrong side of this problem. They built Mimir because they've been the person spending two days answering a question that should take minutes.
1,000+ researchers already using it.
Free for .edu and .gov emails.
Every researcher has done this:
Asked ChatGPT for verified references on a technical question. Got 10 papers that looked perfect. Real-sounding titles. Real-sounding authors. Real-sounding journals.
Then you tried to verify them.
6 don't exist. 2 have the wrong authors. 1 is from a journal that was never published. The last one is real but says the opposite of what ChatGPT claimed.
You just wasted 45 minutes fact-checking an AI that confidently lied to you. And you're back to doing the literature review manually.
This is the single biggest problem with using AI for research. The models are trained to sound right, not to be right. They generate plausible-looking citations the same way they generate plausible-looking code. Sometimes it works. Sometimes it's fabricated from scratch.
Someone finally built a tool where this is structurally impossible.
Mimir searches millions of real scientific papers. It reads them. It understands your question in plain language. And it returns a cited answer where every single reference is a verified, real publication.
Not "probably real." Not "we tried to check." Architecturally impossible to hallucinate.
The system indexes actual papers, books, patents, and industry reports. It can only cite what exists in its corpus. If a paper isn't real, it can't appear in your results. The architecture won't allow it.
Ask a complex technical question. Get an answer with 10+ verified references in seconds. Each citation links to the actual paper. Real title. Real authors. Real journal. Real data.
What it covers right now: materials science, chemistry, physics, engineering. Expanding into new domains continuously.
What used to happen: 2 days reading papers, cross-referencing citations, verifying sources, discovering half your references are garbage.
What happens now: ask a question. Get a cited answer. Every reference is real. Move on.
Founded by a materials scientist and a computer scientist who spent too many years on the wrong side of this problem. They built Mimir because they've been the person spending two days answering a question that should take minutes.
1,000+ researchers already using it.
Free for .edu and .gov emails.
If you work in materials, physics, or chemistry, the new hallucination rate data looks almost reassuring.
The highest fabricated citation rates are in social science and CS. Physics adjacent fields sit far lower, around a fifth of a percent, that a fifth of a percent across millions of references is still thousands of fake citations, and the study only caught references to papers that don't exist.
The real papers cited for the wrong claim, which careful technical fields are not immune to, weren't counted at all.
3.4 million scientific papers were published worldwide in 2025.
The number of indexed research studies rose 48% between 2015 and 2024.
The volume of published science is growing faster than the number of researchers who can read it.
(Source: Web of Science, UNESCO Science Report)
@daforerog The fix isn't better prompting or more careful editing. If the system generates citations rather than retrieving them, hallucination is likely.
@mindinpanic The version that changes their day already exists for some fields and most people just haven't found the right tool for their specific workflow yet.
@serxzsz@Polymarket It's not just legal. Researchers hit the same wall every day. The answer sounds right, the citation looks right, and then you spend an hour confirming it doesn't exist.
If you've asked ChatGPT for sources, you've had it hand you one that looked perfect and didn't exist.
With title that sound real, etc.
That happens because a general model predicts what a citation should look like instead of pulling a real one.
Mimir builds its answers from a corpus of indexed peer-reviewed papers, so the citations point at literature that actually exists, and you can open every one.
On a screen, a confident answer and a correct answer look exactly alike, and that's a whole problem.
A model can write something clean and well sourced looking on top of a paper that doesn't actually support it, and nothing in the formatting tells you.
The better it's written, the easier it is to wave through.
A Lancet study audited about 2.5 million scientific papers.
They found ~146,900 AI-generated fake citations in 2025. By early 2026, 1 in every 277 published papers contained a hallucinated reference.
Mimir only retrieves from indexed, peer-reviewed papers. If a paper isn't in the corpus, it can't be cited.