AI flag accuracy
Statistically indistinguishable from the 89.0% expert-human baseline.
Open research + software
AI can make customer screening 10x cheaper.
We tested five AI models on customer screening for synthetic DNA orders. The best matched expert-human accuracy on four flag criteria at about one-tenth the cost. We published the data, prompts, and a tool you can try.
Results
Comparable accuracy. Lower cost.
Gemini 2.5 Pro with web, bibliographic, and sanctions tools was the strongest system in our 41-profile comparison across five information-gathering tasks.
AI-assisted cost
Per customer, compared with $14.04 for manual screening.
Information gathering
Average AI cost before human review, about 50 times cheaper than manual screening.
The study evaluated information gathering, source quality, source fidelity, and flags—not fulfillment decisions. Human reviewers retained authority over follow-up and fulfillment.
Try it
Try AI customer screening.
Enter a public, fictional, or authorized customer profile. This demo adapts the paper's prompts, searches public records, and returns cited evidence for review. The deployed demo itself was not evaluated.
Open the tool- 01Verify affiliation, institution, and email domain.
- 02Search U.S. sanctions and export-control lists.
- 03Find relevant publications and laboratory work.
- 04Review the evidence and source links yourself.
Open by default
Inspect or reuse every part.
Developed by
Hanna Pálya and Alejandro Acelas.
Hanna Pálya is a PhD student in mathematical epidemiology at the University of Warwick who has researched regulatory routes for DNA synthesis screening.
Alejandro Acelas is a data scientist and developer who has researched risks at the intersection of AI and biosecurity.
Run a screening