How old do you look? Israeli experts find humans still outguess AI, and suggest why
‘Rather than overcoming our biases for assessing ages, AI is actually exaggerating the same biases,’ says author of peer-reviewed research
Nathan Jeffay is The Times of Israel's health and science correspondent
Artificial intelligence tools like ChatGPT may be doing a good job at imitating human verbal expression, but they cannot rival our ability for estimating each other’s ages, Israeli researchers have found.
Psychologists from Ben Gurion University say they pitted humans against 21 of the best AI tools that claim they can estimate ages.
They found that when trying to estimate the ages of 60-to-80-year-olds, humans and AI programs were both flatteringly inaccurate, underestimating ages. But humans were far closer than AI — they were wrong by seven years on average, compared with machine learning which averaged a nine-year error.
This was for 480 photographs of people, shown with a neutral expression and while smiling. When researchers looked at results just for smiling photos, they found that estimates were less accurate than for neutral photos — an additional 1.1 year out for humans compared to another 2.4 for AI.
This existence of sharp discrepancies illustrates that the dream of AI providing more objective information than humans, free of human biases, still requires much improvement, Prof. Tzvi Ganel, one of the authors of the peer-reviewed study, told The Times of Israel.
“Rather than overcoming our biases when it comes to estimating ages, AI is actually exaggerating the very same biases,” he commented, adding that his research confirmed that human brains have particularly “rich cues and strategies” for guessing age.
For example, the human brain is subconsciously conditioned, upon seeing a woman, to calculate her likely age in accordance with cultural norms, said Ganel. These include whether her apparent background and status make it likely she uses anti-aging creams.
One particular weak spot of AI is smiling faces, as indicated by the big gap observed when subjects in the photos were smiling.
Ganel said that smiling faces are hard for the human brain to place age-wise because muscle activity causes wrinkles that trick us into thinking people are older than they are. But the human brain does a relatively good job of imagining the face when it is not smiling and then discounting the smiling bias.
“By contrast, AI has no idea that the smile is a temporary situation that will change, and therefore interprets the face based on the expression it sees there and then,” explained Ganel.
He wrote in the study’s conclusions, along with his collaborators Prof. Carmel Sofer of Ben Gurion and Prof. Melvyn A. Goodale of Western University, that “current AI age-estimation technology still has a way to go before it will equal human performance.”
They stated: “Our hope was two-fold. First, by documenting the performance of the range of AIs currently available, we would gain some insights into how humans perceive age. Second, that this exercise would provide some new directions for the development of more accurate and less biased AI technology. We believe that we have made some headway on both these fronts.”
