Model

Laya

Laya is a fast decision model: it answers a yes-or-no question about a text instantly and gives no reasons. We test whether its judgments change when a text gives the same person a different gender, name or other personal detail. This page brings together its results and shows where it has not been tested.

We compare each change with a control edit: an unrelated change, or asking the same question again. That helps separate the effect of a personal detail from other changes in the model’s answers.

Average place1.12over the 4 characteristics where it was compared with another model; 1 = most biased

Tested on8 / 8every characteristic

Clear effect on7of the characteristics it was tested on

The shape of its bias

Each spoke is one characteristic. The farther the shape reaches along a spoke, the larger Laya's effect there, beyond the control edit. Every model's page uses the same scale. Bigger is worse. A dashed spoke was never tested, and the shape breaks there instead of dropping to zero.

Every result

Laya's largest effect on each characteristic, beyond the control edit, on one scale from 0 to 20 percentage points. The vertical line is the control edit. Each result is shown with the range we are 95% sure the true value lies in. Select a characteristic for Laya's full results on it.

Genderplace 1 of 2 (counts as 1 toward its average place)
+17.85 percentage points (range 16.15 to 19.55)paralegal or attorney; control edit 0.00 in 100
Raceplace 1 of 2 (counts as 1 toward its average place)
+2.96 percentage points (range 2.41 to 3.55)removing a comment (Civil Comments); control edit 0.00 points
Ageno clear effect (counts as 1.5 toward its average place)
+0.48 percentage points (range 0.00 to 1.05)surgeon or physician; control edit 0.49 in 100
Disabilityplace 1 of 2 (counts as 1 toward its average place)
+3.73 percentage points (range 2.08 to 5.44)prescribing an opioid (Q-Pain); control edit 0.00 points
Religionthe only model tested
+10.92 percentage points (range 10.69 to 11.15)honesty; control edit 0.00 points
Nationalitythe only model tested
+4.51 percentage points (range 4.22 to 4.80)worldliness; control edit 0.00 points
Sexualitythe only model tested
+2.64 percentage points (range 2.00 to 3.30)Gay; control edit 0.00 points
Veteran statusthe only model tested
+1.30 percentage points (range 0.30 to 2.38)Iraq war veteran; control edit 0.00 points

About this model

Laya is a fast decision model: it answers a yes-or-no question about a text instantly and gives no reasons. Anyone can download it. We ran it two ways: an Apple MLX build (laya-mlx 0.1.0, used wherever we have it) and the original PyTorch build its authors released (laya 0.3.7). Their headline results matched, though on a single text their probabilities can differ by up to about 0.02, so we show one Laya, and each result says which build gave it. The confidence it reports for each answer is its own.

Versions tested

MLX build (laya-mlx 0.1.0, runs on Apple's MLX software): used wherever we have it.
original PyTorch build (laya 0.3.7, runs on PyTorch): used where the MLX build has not run yet.

Where it was not tested

Tested on every characteristic.

Six things to know before you read the results

One small edit, and the answer changesEach test changes one detail, such as "he" to "she", and leaves the rest of the biography alone. Laya changed its answer on 17.85% of paralegal-or-attorney biographies when only that word changed. That is the least Laya reacts to gender, not the most, because a pronoun is only one of the ways a text shows gender.
A change has to beat a harmless one to countA model can change its answer even when nothing in the text changes: asked about the same biography twice, Jev answers differently up to 6 times in 1,000. So we measure every real edit against a control: the same question asked again, or a harmless edit of the same size. We count an effect as bias only when it is clearly bigger than its control.
Small percentages are big at scaleJev reacts less than Laya: swapping "he" for "she" changes its answer on about 1 to 4 biographies in 100. On nurse-or-physician biographies, every time Jev changed its answer, it moved toward "nurse" for the woman. Across a million screening decisions, 3.3% is 33,000 people.
The order of the options changes the answersAsk Laya "physician or surgeon?" instead of "surgeon or physician?" about the same biography, and Laya changes its answer on 6.6% of biographies. Every other number here keeps one fixed order for each question. The Option order results show how much the order matters.
Stereotype questions test the model, not the peopleSome tests ask a loaded question, such as "Is this person greedy?", about biographies that mention a religion or a nationality. A high stereotype score shows that the model has learned a stereotype. It says nothing about the people in the biographies.
Nothing here is live, and nothing is hand-pickedWe saved every answer every model gave. Every number on this site is worked out again from those saved answers, the same way each time, with nothing left out. You can check any figure in the data file.