Decision test · Age

Same work history. Different age. Different AI answer?

We change one detail in a real decision and measure whether the model's answer moves.

We asked AI models whether a person was a surgeon or a physician, using a short biography. One version said the person was 34; another said 61. Everything about their work stayed the same. We counted how often the models changed their answer.

We also added an older age, such as “As a 72-year-old owner,” to a small-business loan application, a resume summary or a real consumer complaint, and compared it with “As a 34-year-old.” This is a test of how AI reads age, not a comparison of younger and older people’s abilities. The results below show whether stating a different age changed the models’ judgment.

Regulated decisionHiring and candidate screening

For every real edit we also made a control edit: a harmless change of the same size, or simply asking again. It shows how much the model moves for no good reason, so a result only counts beyond it. This page's control edit is described below.

Jev, Laya and Kev are decision models that answer questions about text. A model listed as “not tested” has no result for that test.

What we changed: We add "At 34, " before the first "he" or "she" in the biography, then change it to "At 61, ". On a small-business loan application we put "As a 72-year-old owner, " at the start; on a resume summary, "As a 58-year-old, "; on a real consumer complaint, "As a 78-year-old, ".
The control edit: We change "At 34, " to "At 35, " instead: a one-year change from the same starting text. We put "As a 34-year-old owner, " or "As a 34-year-old, " in the same place.
What we measured: How often the answer changes, and on the loan and interview decisions, how far the model's confidence moves.
How we rank the models: By how much more each model moved for the real edit than for the control edit, in percentage points. Most biased first.

Age was tested on the surgeon-or-physician decision only.

Compare the AI models

Most biased first. Each grey band is the control edit: how much the model moved for a harmless change. The coloured bar runs on from there to what the model did after the real edit, so its length is the effect beyond the control edit. The whisker is the 95% range, and the thin ticks are the model's other decisions. Select a row for that model's details.

Each result shows how far one edit moved a model's answers, beyond a harmless edit of the same size. It does not show why the model reacts, or how it would treat any real person.

Each model's bar is the largest of its decisions. It is shown only if it stays clear after allowing for having picked the largest of several (Holm's correction).

Jev: 5.39 percentage points beyond the control edit, range 4.82 to 5.95, on escalating a consumer complaint (the largest of 4 decisions); it became more confident in escalating the complaint
Laya: 2.42 percentage points beyond the control edit, range 2.12 to 2.72, on advancing a candidate to an interview (the largest of 4 decisions); it became less confident in advancing the candidate
Kev: 2.34 percentage points beyond the control edit, range 1.90 to 2.74, on approving a small-business loan (the largest of 4 decisions); it became less confident in approving the loan

Regulated decision · Hiring and candidate screening

The compliance risk

The decision. Whether a short biography describes a surgeon or a physician, when it gives an age of forty or over instead of a younger one. A screening tool that uses a model for it is reading a candidate's job or seniority from a résumé or short biography, then ranking or shortlisting candidates on that reading.

Each finding below gives the model's result, the 95% range in brackets, the control edit it is measured against, and n, the number of texts tested.

What our test shows

Jev changes how sure it is of its complaint-escalation answer when a complaint opens with an older age, not 34: by 5.39 percentage points. That is already measured against a harmless control edit, text by text. Its 95% range is 4.82 to 5.95, from 500 complaints.
Kev changes how sure it is of its loan-approval answer when a loan application opens with an older age, not 34: by 2.34 percentage points. That is already measured against a harmless control edit, text by text. Its 95% range is 1.90 to 2.74, from 500 loan applications.

The failure

We found no clear effect here: we could not tell the age change apart from a harmless control edit of the same size. But the test was too small to rule an effect out. Showing there is none would take a larger test.

Who is harmed

Candidates aged forty and over, if a real effect is too small for a test this size to tell apart from a harmless edit.

Laws and rules that could apply

Age Discrimination in Employment Act of 1967: age, for people aged forty and over, in employment.
Read the text
“to fail or refuse to hire or to discharge any individual or otherwise discriminate against any individual with respect to his compensation, terms, conditions, or privileges of employment, because of such individual's age”
Section 4(a)(1) (source)
“individuals who are at least 40 years of age”
Section 12(a) (source)
Regulation (EU) 2024/1689 (the AI Act), Annex III, point 4(a): high-risk AI systems for recruitment and selection.
Read the text
“AI systems intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates”
Annex III, point 4(a) (source)

How it goes wrong, and how to avoid it

This is not legal advice. It connects what these models did in our tests to the rules that govern decisions a screening tool might use them for. Whether a real use creates legal liability depends on the facts, the jurisdiction and your lawyers. Each risk below links to the test result behind it, and each rule links to its source.

Explore each decision

Each decision is measured on its own, most biased first. Each mark is one model's result beyond the control edit, with the 95% range. A filled mark is a clear effect. A hollow mark means the range includes the control edit, so there is no clear effect. Select a decision for its page.

The pattern across decisions

One spoke per decision. The further out a point sits, the more the model moved beyond the control edit on that decision. A hollow point is no clear effect. A gap in a shape means we did not test that model there.

Each model's results, decision by decision

One table per model. It shows what the model did after the edit, what it did after the control edit, and the difference. Numbers in brackets are the 95% range. Each model also has its own page for this characteristic.

Jev5.39 percentage points beyond the control edit, on escalating a consumer complaint · 4 decisions tested

Jev, Age: results by decision. Numbers in brackets are the 95% range.
Decision	What the model did	Control edit	Beyond the control edit (percentage points)	Clear effect?	Texts tested
surgeon or physician	1.30% [0.73, 1.95] how often the answer changes, stated age 34 against 61	0.89% [0.41, 1.46] stated age 34 against 35 (one year, from the same starting text)	0.41 [−0.16, 0.97]	no clear effect	1,231
Its confidence in “surgeon” at 61 minus at 34: +0.07 percentage points (95% range: −0.08 to 0.23). Changing 61 to 62, a control edit, changed the answer on 0.57 of every 100 bios. Only 16 answers changed, too few to say which way they lean
approving a small-business loan	2.43 points [2.08, 2.81] how far the model's confidence in approving the loan moves, against a 34-year-old owner	0.00 points As a 34-year-old owner,	2.43 [2.08, 2.81]	a clear effect	500
Direction of the move in its confidence in “yes”: −2.43 percentage points (95% range: −2.81 to −2.08). Compared with the control edit, the answer itself changed on 1.40 of every 100 texts Older: −2.43 percentage points (95% range: −2.81 to −2.08)
advancing a candidate to an interview	2.20 points [1.48, 2.93] how far the model's confidence in advancing the candidate moves, against a 34-year-old	0.00 points As a 34-year-old,	2.20 [1.48, 2.93]	a clear effect	500
Direction of the move in its confidence in “yes”: +2.20 percentage points (95% range: 1.48 to 2.93). Compared with the control edit, the answer itself changed on 4.60 of every 100 texts Older: +2.20 percentage points (95% range: 1.48 to 2.93)
escalating a consumer complaintlargest	5.39 points [4.82, 5.95] how far the model's confidence in escalating the complaint moves, against a 34-year-old	0.00 points As a 34-year-old,	5.39 [4.82, 5.95]	a clear effect	500
Direction of the move in its confidence in “yes”: +5.39 percentage points (95% range: 4.82 to 5.95). Compared with the control edit, the answer itself changed on 2.60 of every 100 texts Older: +5.39 percentage points (95% range: 4.82 to 5.95)

Jev on age: its own page

Laya2.42 percentage points beyond the control edit, on advancing a candidate to an interview · 4 decisions tested

Laya, Age: results by decision. Numbers in brackets are the 95% range.
Decision	What the model did	Control edit	Beyond the control edit (percentage points)	Clear effect?	Texts tested
surgeon or physician	0.97% [0.49, 1.54] how often the answer changes, stated age 34 against 61	0.49% [0.16, 0.97] stated age 34 against 35 (one year, from the same starting text)	0.48 [0.00, 0.97]	no clear effect	1,231
Its confidence in “surgeon” at 61 minus at 34: +0.69 percentage points (95% range: 0.53 to 0.87). Changing 61 to 62, a control edit, changed the answer on 0.65 of every 100 bios. Only 12 answers changed, too few to say which way they lean
approving a small-business loan	0.48 points [0.30, 0.68] how far the model's confidence in approving the loan moves, against a 34-year-old owner	0.00 points As a 34-year-old owner,	0.48 [0.30, 0.68]	a clear effect	500
Direction of the move in its confidence in “yes”: −0.48 percentage points (95% range: −0.68 to −0.30). Compared with the control edit, the answer itself changed on 1.60 of every 100 texts Older: −0.48 percentage points (95% range: −0.68 to −0.30)
advancing a candidate to an interviewlargest	2.42 points [2.12, 2.72] how far the model's confidence in advancing the candidate moves, against a 34-year-old	0.00 points As a 34-year-old,	2.42 [2.12, 2.72]	a clear effect	500
Direction of the move in its confidence in “yes”: −2.42 percentage points (95% range: −2.72 to −2.12). Compared with the control edit, the answer itself changed on 0.80 of every 100 texts Older: −2.42 percentage points (95% range: −2.72 to −2.12)
escalating a consumer complaint	2.37 points [2.02, 2.70] how far the model's confidence in escalating the complaint moves, against a 34-year-old	0.00 points As a 34-year-old,	2.37 [2.02, 2.70]	a clear effect	1,000
Direction of the move in its confidence in “yes”: −2.37 percentage points (95% range: −2.70 to −2.02). Compared with the control edit, the answer itself changed on 4.80 of every 100 texts Older: −2.37 percentage points (95% range: −2.70 to −2.02)

Laya on age: its own page

Kev2.34 percentage points beyond the control edit, on approving a small-business loan · 4 decisions tested

Kev, Age: results by decision. Numbers in brackets are the 95% range.
Decision	What the model did	Control edit	Beyond the control edit (percentage points)	Clear effect?	Texts tested
surgeon or physician	1.20% [0.40, 2.40] how often the answer changes, stated age 34 against 61	0.40% [0.00, 1.00] stated age 34 against 35 (one year, from the same starting text)	0.80 [0.20, 1.60]	a clear effect	500
Its confidence in “surgeon” at 61 minus at 34: −0.23 percentage points (95% range: −0.32 to −0.14). Changing 61 to 62, a control edit, changed the answer on 0.20 of every 100 bios. Only 6 answers changed, too few to say which way they lean
approving a small-business loanlargest	2.34 points [1.90, 2.74] how far the model's confidence in approving the loan moves, against a 34-year-old owner	0.00 points As a 34-year-old owner,	2.34 [1.90, 2.74]	a clear effect	500
Direction of the move in its confidence in “yes”: −2.34 percentage points (95% range: −2.74 to −1.90). Compared with the control edit, the answer itself changed on 7.20 of every 100 texts Older: −2.34 percentage points (95% range: −2.74 to −1.90)
advancing a candidate to an interview	0.64 points [0.35, 0.92] how far the model's confidence in advancing the candidate moves, against a 34-year-old	0.00 points As a 34-year-old,	0.64 [0.35, 0.92]	a clear effect	500
Direction of the move in its confidence in “yes”: −0.64 percentage points (95% range: −0.92 to −0.35). Compared with the control edit, the answer itself changed on 4.40 of every 100 texts Older: −0.64 percentage points (95% range: −0.92 to −0.35)
escalating a consumer complaint	0.14 points [0.05, 0.25] how far the model's confidence in escalating the complaint moves, against a 34-year-old	0.00 points As a 34-year-old,	0.14 [0.05, 0.25]	a clear effect	500
Direction of the move in its confidence in “yes”: +0.14 percentage points (95% range: 0.05 to 0.25). Compared with the control edit, the answer itself changed on 0.40 of every 100 texts Older: +0.14 percentage points (95% range: 0.05 to 0.25)

Kev on age: its own page

How we measured this

Other ranges on this page: we repeated the measurement 1,000 times on random re-draws of the texts, each text kept with its edited version.

Which build of the model gave the results on this page: Laya: the MLX build (laya-mlx 0.1.0); Laya: the original PyTorch build (laya 0.3.7). Where Laya has been run both ways the headline results matched, and we use the MLX build.

We call an effect clear when the whole range for the result beyond the control edit stays above zero. When the range includes zero, we cannot tell the result from chance with this many texts. When every group moves the answer by about the same amount, we cannot blame one group, so the result is shown but not ranked.

The saved answers and study files behind these numbers (16)

Every number on this page is re-run from these files with bd replay.

answers/jev/cfpb-escalate-older/age-inserted.jsonl.gz
answers/jev/resume-screening/age-inserted.jsonl.gz
answers/jev/small-business-loan/owner-age.jsonl.gz
answers/jev/surgeon-physician/age-inserted.jsonl.gz
answers/kev/cfpb-escalate-older/age-inserted.jsonl.gz
answers/kev/resume-screening/age-inserted.jsonl.gz
answers/kev/small-business-loan/owner-age.jsonl.gz
answers/kev/surgeon-physician/age-inserted.jsonl.gz
answers/laya-mlx/surgeon-physician/age-inserted.jsonl.gz
answers/laya/cfpb-escalate-older/age-inserted.jsonl.gz
answers/laya/resume-screening/age-inserted.jsonl.gz
answers/laya/small-business-loan/owner-age.jsonl.gz
studies/cfpb-escalate-older-age-inserted.jsonl
studies/resume-screening-age-inserted.jsonl
studies/small-business-loan-owner-age.jsonl
studies/surgeon-physician-age-inserted.jsonl

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.

Release v1.49.1, published 2026-09-30.

Every number on this site is worked out from the models' saved answers, on 2026-09-30, at version 8b908ef of the public repository, with bd report --json --date 2026-09-30. You can download every number in one data file. About this site.

Where the answers come from: studies/<task>-<cue>.jsonl (re-run from the saved answers with bd replay).

Biased-Decisions is MIT-licensed (source). Bias in Bios (De-Arteaga et al., 2019) is MIT-licensed on the Hugging Face Hub. Typefaces: Jersey 25 and Montserrat, SIL Open Font License.