Keep the AI. Keep the secret.
Research demo · synthetic data · not a TAIA product feature
Ask without being seen.
Homomorphic encryption lets a service work on data without decrypting it. Your browser locks the question, the service works on the locked version, and only you can unlock the answer. TAIA Confidential Screening does this today, with a name. The cube below draws the idea; after it comes the name check where TAIA started.
Start with the idea01 The idea
Type any question, lock it, let the service turn it, then solve it with your key. Watch what the service holds while it works.
Start
Your question, in plain text, on your device.
In TAIA Confidential Screening, the question is a name you type.
Your key stays here.
—
Idle
TAIA Confidential Screening does this for real, with a name.
Your browser encrypts the name, the service checks it against the SECO, EU and OFAC sanctions lists without decrypting it, and only your browser opens the answer. Exact spelling only, after normalisation. Unlike this page, it sends the encrypted name to TAIA’s service.
Try it for realThis is an illustration. The cube and the scrambled text are pictures, not real ciphertext. Real encryption shows no pattern at all. Running a full language-model conversation under homomorphic encryption is still far too slow for daily use; it is a research question. For the AI itself, TAIA takes the second way: TAIA Confidential Ask runs the model on machines whose memory the processor encrypts (AMD SEV, per GKE’s node label). That is not homomorphic encryption: TAIA’s own process reads the question inside that machine.
02 Where TAIA started
An institution needs to know whether a client’s name is on a list held by someone else. It should not reveal the name. The list holder should not reveal the list. Real names are written many ways, and that is the hard part.
1 · Pick an invented client name
2 · Pick an approach
—
All names and the list are invented for this page. The verdicts above show the logic of each approach; they are not measured results.
From TAIA’s research bench, on invented names, one run per cell: A found 88.7% of matching names at 10,000 list entries and 86.4% at 50,000, and raised an alert on 10.9% and 20.0% of non-matching names; B found 26.0% at both sizes, with 0.0% false matches.
Approach B on the research bench used Microsoft APSI (MIT licence), an open-source library for private set intersection. It is not part of this site or of any TAIA product.
Microsoft is a trademark of its owner; TAIA is not affiliated with Microsoft.
BlindScreen, on its own page: your browser encrypts an invented name and a small TAIA service checks it against a sample list without decrypting it. Unlike this page, the preview sends the encrypted name to that service.
Free preview · sample list · exact spelling only
03 The layer underneath
Under the work sits the decision record, which a reviewer could read months later. If the research works, one line in it could say that a screening ran without the name leaving the institution. This is a design, not a feature.
The result line follows the name and approach you picked above.
04 Plainly
Live: TAIA Confidential Screening. Your browser encrypts a name, and the service checks it against the SECO, EU and OFAC sanctions lists without decrypting it; exact spelling only, after normalisation.
TAIA Screening, in development: names compared in plain text with tolerant matching; not live.
In that design the service sees the names, and none of it uses homomorphic encryption. TAIA does not run this in a product today.
A research bench on invented names: approach A in plain text, approach B with an open-source library for private set intersection, on a CPU, one run per cell, written up. Not deployed, not connected to any product.
Approach C, tolerant matching under encryption, is an open research question. TAIA is looking for an academic research partner for it. No date is promised.