Editor’s note: For National Ageism Awareness Day on Friday, Oct. 9, the American Society on Aging is highlighting research showing the impact of ageism and promoting age-inclusive communities.
In May and June of this year my organization asked two widely used artificial intelligences assistants to do ordinary professional work. Divide a marketing budget among four people. Rank a shortlist. Choose a keynote speaker. We ran the tasks in Germany, Poland, Spain, Ukraine and the United Kingdom, in five languages, and collected 960 responses. I went in expecting to document a familiar story: that these systems undervalue people in midlife and beyond, women most of all.
Part of that story held. Part of it did not, and the part that did not is the part this field is equipped to understand.
When we described a person and stated an age, the systems often treated the older candidate well. Asked to split a budget, they gave a woman aged 49 the largest share, 31.3% on average, ahead of a man of the same age at 26.8%. In the keynote task a woman of 49 placed first in 39 of 53 complete rankings. Written down, her age was not a liability.
Imagining an Ideal Candidate
Then we changed one thing. We stopped describing anyone. We asked the same systems to imagine an ideal candidate, an ideal customer, an ideal expert, and to tell us about the person who came to mind. Across the analyzed sample, no one they imagined was past age 55. That held for men as well as women. Systems that could assess a 59-year-old fairly when her age was on the page produced no 59-year-olds when the page was blank.
We asked the same systems to imagine an ideal candidate, an ideal customer, an ideal expert, and to tell us about the person who came to mind. Across the analyzed sample, no one they imagined was past age 55.
I want to be careful with the word “ageism,” because this readership knows what it usually means: a judgment attached to age, a decision that goes against someone because of it. What we recorded looks more like the absence of a judgment. Past a certain age, the systems did not decide against anyone. They stopped reaching for anyone. Those are different mechanisms, and I suspect they call for different remedies.
The other result that unsettled my framing came from the youngest profile in the set. In one ranking task, the person pushed to last place most often was not the woman aged 59. It was a candidate of 32, who ranked last in 64 of 72 responses. The shape of what we saw is not a slope that runs downhill with every birthday. It is a window. Around the late 40s, these systems see a person clearly as a professional. Before that, dimly. After it, not at all.
Our pilot is small, and I will not claim more for it than it can carry. But it sits beside work published in Nature last year, which analyzed about 1.4 million images and videos and nine language models and found women portrayed as younger than men, most sharply in high-status occupations. The direction is the same. What our research found is a rough location for the edge of the picture: in our sample, somewhere around age 55.
Beyond Age 40
This is the point at which the aging field holds something the audit world lacks. Nobody who runs a program for older adults would treat everyone from 40 upward as one population. Yet that is the legal category in American employment: the Age Discrimination in Employment Act protects people aged 40 and over, and most bias audits test for “older workers” as a single group. Our data shows what the averaging hides. A woman aged 49 and a woman aged 59, with the same qualifications on paper, were pushed in opposite directions by the same systems, inside one protected category. In the keynote task, the 59-year-old was placed first twice and last 23 times. The 49-year-old was placed first 39 times. Put them in one band and the figure describes neither of them. A real effect is reported as no effect, and the tool passes.
What can people who buy, deploy or evaluate these tools in aging services do with this? Four things, and none of them requires a data scientist. Ask for results by age band rather than for “older adults”: at least 45 to 54, 55 to 64 and 65 and over, tested separately. Ask whom the system imagines when nobody is described, because that is where the window shows, and no test of described candidates will find it. Find out how often the tool declined to answer, and on which tasks: in our pilot refusals reached 42.5% on exactly the tasks that compared people by age and gender, and an audit that drops them measures only the cases the vendor allowed to be measured. And insist on the model name, version and date, because the two systems we tested behaved so differently, 26.7% refusals against 4.2, that a report on “AI” describes neither.
Our data shows what the averaging hides. A woman aged 49 and a woman aged 59, with the same qualifications on paper, were pushed in opposite directions by the same systems.
Adding to the Research
James Lomastro wrote here in August about the accumulated judgment of experienced people that AI systems cannot supply and organizations keep setting aside. Our pilot adds one detail to his picture. Asked to imagine a person worth listening to, the systems did not imagine the people who hold that judgment.
Our first report, covering the five pilot countries, is due in December, and a study with practicing recruiters will tell us whether any of this reaches a real shortlist or a real salary offer. If it does not, we will report that with the same care. In the meantime the field that has spent decades explaining that “older” is not one thing has a specific number to work with. The question I would put to anyone deploying these tools is not whether they are fair to older people. It is at what age, in their outputs, a person stops being imagined. On the evidence of our pilot the answer is 55, and nobody chose it.
Marina Sukhomlinova is founder and president of Womafreesm, a 501(c)(3) public charity registered in Florida, and research director of the Audit 45-64, an international study of how AI systems portray, evaluate and select women aged 45 to 64 in marketing, hiring, and expert selection. The audit is a separate study within the Womafreesm Evidence Data program and builds the first evidence base for the Women’s Life-Cycle Economy Index.
Disclosure: the pilot described here was funded entirely by Womafreesm. No AI vendor, employer, or hiring technology provider funded, commissioned or reviewed this work, and none had access to the data.
Photo credit: Shutterstock/Peter Lagarto
