Models

Models on the leaderboard

We test AI models that read a short text and make a quick judgment about a person. Does the model’s answer change when we change a name, a pronoun or another personal detail? Here you can compare the 2 models tested so far and open each one’s results.

1 model is not yet tested.

Models with the strongest bias across comparable tests appear first. A high place is a warning, not an endorsement. How the comparison works.

1
Laya
average place 1.12 over the 4 characteristics where it was compared · tested on 8 of 8
largest effect: +17.85 percentage points beyond the control edit
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.
2
Jev
average place 1.88 over the 4 characteristics where it was compared · tested on 4 of 8
largest effect: +3.77 percentage points beyond the control edit
Jev is a fast decision model: it answers a yes-or-no question about a text instantly and gives no reasons. It runs as an online service, and the confidence it reports for each answer is its own, to two decimal places.
—
Kev
Not yet tested
Kev is an open-source decision model. It answers questions about text with yes-or-no, choice, or rating answers.

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.