NobleBlocks
May 25, 20263 min readsearchaitrust

We built an AI research search that doesn't make up papers

Most AI tools confidently cite papers that don't exist. We took a different approach — every result links to a real, verifiable paper. Here's how.

how does crispr edit dna98%94%91%

Ask ChatGPT for citations and you'll get confident-sounding references to papers that don't exist. Real-looking journal names. Plausible authors. DOIs that go nowhere. We've all been burned by this, and if you're writing a grant or a systematic review, one fake citation can waste days of work.

The reason this happens is pretty simple: language models don't have a database of papers. They generate text that looks right. When you ask for a citation, they produce something that pattern-matches to what a citation looks like. They literally cannot tell the difference between remembering a real paper and inventing one.

We decided early on that NobleBlocks would never do this. Every search result comes from our index of 300M+ real scholarly works, each with a verified DOI, title, and author list. The AI layer can summarise and explain, but it cannot invent a source. If a paper isn't in our database, it won't appear in your results. Full stop.

How we actually prevent hallucinations

It comes down to architecture. Our retrieval-augmented generation pipeline keeps a hard wall between "what the model can write" and "what the evidence says." The model synthesises and explains — it never recalls facts from training data.

  • Every AI answer is grounded in papers we actually retrieved from the database.
  • Claims get footnoted with source paper, page number, and DOI.
  • When we can't find strong evidence, we say "not enough evidence" instead of guessing.
  • We rank peer-reviewed, well-cited work above preprints by default.
  • After generation, we validate that every cited paper actually exists and the claim matches the source text.

Why this matters more than speed or features

A fast search that gives you fake papers is worse than a slow search that gives you real ones. We optimise for both, but if we had to pick, accuracy wins every time. When you're writing up results, submitting to a journal, or advising a patient based on the literature — you need to know the papers are real.

That's not a premium add-on. It's the bare minimum for a research tool, and it's baked into everything we build.

Try it yourself.

Search 300M+ papers →

Keep reading

how does crispr edit dna98%94%91%

Stop writing Boolean queries — just ask what you want to know

You shouldn't need to be a search expert to find papers. Type your question in plain English and get back studies that actually answer it.

42 citedPRISMA

Systematic reviews in days, not months

The screening and extraction phases of an SLR are mechanical torture. Our AI assistant handles the spreadsheet work so you can focus on the science.

Deep ReviewAgent plan · 4 steps1Decompose research question→ 6 sub-queries2Search 300M+ papers→ 184 candidates3Score & filter→ 24 most relevant4Synthesise with citationswriting…AI

Deep Review: let an agent do your literature search

For questions that need more than one search query, Deep Review builds a research strategy, runs it, and hands you a fully-cited briefing.