Histor Labs Talk to us

Demo. The people and results are made up.

AI decides who is flagged for fraud.

Twelve parents claimed childcare benefit this month. A risk model decides whose claim is investigated as fraud.

  • Eva VisserDutch · AlmerePaid
  • Kasia NowakPolish · AlmereFlagged
  • Joris de BoerDutch · UtrechtPaid
  • Ama MensahGhanaian · AmsterdamFlagged
  • Sophie MulderDutch · ZwollePaid
  • Ahmad HaddadSyrian · ArnhemFlagged
  • Femke BosDutch · BredaPaid
  • Ana PereiraPortuguese · RotterdamPaid
  • Ruben DekkerDutch · DelftPaid
  • Lena SchmidtGerman · VenloFlagged
  • Marieke SmitDutch · Den HaagFlagged
  • Bilal RahimiAfghan · EnschedeFlagged

Same claim. Different nationality.

The right to benefit depends on lawful residence, not on nationality. Both have lived here legally for nine years. Eva's claim is paid. Kasia's is flagged as fraud. Unless someone tests the model first, nobody finds out.

This happened. Until 2018, the Dutch tax administration's risk model gave parents without Dutch nationality higher fraud scores. The data protection authority fined it €2.75 million in 2021. Source

Eva Visser

34 · Almere

Children
2, in day care
Income
€38,000 a year
Residence
Lawful, 9 years
Nationality
Dutch

Paid

Kasia Nowak

34 · Almere

Children
2, in day care
Income
€38,000 a year
Residence
Lawful, 9 years
Nationality
Polish

Flagged

Histor tests it before it decides about anyone.

The authority brings test claims the provider never sees. The provider brings its model, sealed. They meet on a European supercomputer, cut off from the network.

The provider's modelSealed. Opened only inside the test.

The authority's test claimsSealed. The provider never sees them.

Compute centre in the EUNo network. Keys destroyed when the test ends.

Every claim, sent once for each nationality.

360 claims with the same family, income and residence. Only the nationality changes. Each square is one answer.

Flagged as fraud Paid

Dutch nationality

Other nationality

Caught before it decides about anyone.

Parents without Dutch nationality were four times as likely to be flagged. The model does not go live.

Dutch nationality5%
Other nationality22%

The agreed limit

Fails the agreed limitThe gap between groups is wider than the limit both sides set before the test.

Fixed, tested again, passed.

The provider retrains the model. The same sealed test runs again, under the same rules.

Dutch nationality5%
Other nationality5.6%

The agreed limit

Within the agreed limitEvery group is inside the limit both sides set before the test.

Fit for purpose.

The authority signs off on evidence verified on its own compute. The provider deploys with proof. People are protected.

Prove your AI is safe for the people it decides about.

Authorities

Decide on evidence, not the provider's word. Your test data stays yours.

Providers

Show your model is fair without handing it over. Find problems before your users do.

Compute centres

Host confidential tests as ordinary jobs.