A 2022 study built AI models that predict how people tend to judge faces—and can edit facial images to shift those impressions. The models do not reveal a person’s real personality, trustworthiness, competence, or identity. They reproduce patterns in human ratings, including shared stereotypes.
What did the researchers create?
In a paper published online April 21, 2022, in Proceedings of the National Academy of Sciences (PNAS), researchers combined deep generative image models with a crowdsourced dataset of human impressions. The dataset, called One Million Impressions, contains 1,020,000 judgments about 1,000 synthetic, naturalistic face stimuli across 34 attributes.
The attributes ranged from age and adiposity to trustworthiness, masculine or feminine appearance, and familiarity. These are ratings of what observers infer from an image—not objective measurements of a person’s character or identity. The study’s authors say that subjective and socially constructed impressions need not correspond to the person’s actual identity, attitudes, or competencies. Read the PNAS paper.
How does the face-judging AI work?
The model learns associations between facial image representations and the average ratings people gave. It can then predict the impressions observers are likely to report, generate synthetic faces along modeled attribute dimensions, or alter a face image to increase or decrease a perceived attribute.
#1 Best Overall
That distinction matters: the system models what people tend to see, not what is true about the person pictured. Calling an output a “trustworthiness” score, for example, does not make it a reliable measure of trustworthiness. It describes a population’s reaction to visual cues.
Did the edits actually change people’s impressions?
The paper reports that the model’s predictive accuracy approached human interrater reliability—the degree to which different people agree in their ratings. The researchers also tested image transformations in 20 preregistered experiments involving more than 1,000 participants. In general, the edits shifted ratings in the intended direction.
There was an exception: the reported significant positive linear trend did not hold for familiarity ratings of manipulated real faces. The findings therefore support the model’s ability to steer some impressions under the tested conditions, not a guarantee that every edit will produce the intended effect in every observer or setting.
Why call the AI “deliberately biased”?
The phrase refers to the deliberate inclusion and modeling of human biases in the ratings, not to a claim that the researchers set out to build a discriminatory decision system. Coauthor Joshua Peterson described the dataset this way: “Our dataset not only contains bias, it deliberately reflects it,” as quoted by Futurism.
Rank #3
The authors characterize the judgments as systematic biases and stereotypes shared by raters. A model trained to reproduce those impressions can encode them even when it predicts ratings accurately. Predictive performance is not proof that an impression is fair, correct, or useful for making decisions about an individual.
Whose impressions does the model represent?
The authors describe the model as predicting impressions in a general, mostly White, North American population. That limits how broadly its outputs should be interpreted. The paper also notes that Black faces may have been undersampled among the stimuli and reports gaps between model performance and rating reliability for some racial or ethnic attributes. The authors caution that the evidence does not establish why those gaps occurred.
Results from this dataset should not be treated as a universal account of how all people, cultures, or communities perceive faces. Nor does the study establish that the model can validly assess real-world individuals outside the tested research context.
What risks did the researchers identify?
Conventional image manipulation can put a person into a different scene or context. This method raises a different concern: it can alter a face itself in ways intended to change the impressions viewers form. The authors warn that subtle edits may be difficult to detect while still influencing social perception. Such changes could potentially boost or damage someone’s reputation.
The authors argue that methods, implementations, and supporting data “should be made transparent from the start” so that detection and defense protocols can be developed. The possibility of impression-shifting edits is a reason for scrutiny, not evidence that this particular model has been used to target real people.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the headline leaves out
“Judges you as brutally as your mother-in-law” is a provocative way to describe a model of first impressions, but the AI is not independently discovering who someone is. It estimates how people in the study tended to interpret generated facial cues, then can manipulate images to influence those interpretations. Those judgments reflect the raters and their biases—not a dependable diagnosis of the person in the image.
The primary paper is Peterson, Uddenberg, Griffiths, Todorov, and Suchow, “Deep models of superficial face judgments,” published in PNAS on April 21, 2022, volume 119, issue 17, article e2115228119. Futurism’s coverage by Noor Al-Sibai followed on April 25, 2022.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →
