Training AI on stereotypical facial expressions is bound to mislead because scientific evidence indicates that there are no universal expressions of emotion
Lisa Feldman Barrett / Wall Street Journal : Bluesky: @tylerking.bsky.social . X: @carnage4life , @naimulkhan , and @lfeldmanbarrett LinkedIn: Dr. Jeffrey Funk , Rafael Brown , and Charles L Mauro Bluesky: Tyler King / @tylerking.bsky.social : They keep trying to find the commercial use case that works and keep failing. Because the technology is a fraud based what it might one day become rather than what it is now. [embedded post] X: Dare Obasanjo / @carnage4life : Every AI system that claims to either read people's emotions or other characteristics like their political affiliations from facial expressions is a scam. There's no underlying science here so there's no underlying “good data” to have trained an AI on to predict these qualities [image] Naimul Khan / @naimulkhan : Forget about face, my lab has been at it for 6+ years and we can't get past 72% even with physiological signals (ECG/EEG). Emotion recognition research has a long way to go, and should NOT be released in the wild like this irresponsibly. https://www.wsj.com/... Lisa Feldman Barrett / @lfeldmanbarrett : If you're interested to view the generative AI output I mention in my @WSJ op-ed, visit http://www.how-emotions-are-made.com/ .... LinkedIn: Dr. Jeffrey Funk : Despite OpenAI's claims, “The best available scientific evidence indicates that there are no universal expressions of emotion.” … Rafael Brown : Microsoft walked away from emotion detection and facial recognition as did IBM and Amazon. Here is a reading list: … Charles L Mauro : Dr. Jeffrey Funk My research team utilizes Micro Facial Expression Analysis research tools that do this exact set of functions. …
Context & Ripple Effects
This critique extends a long-running dispute over whether facial signals can reliably be translated into inner states. Earlier coverage noted that researchers questioned both the efficacy and ethics of emotion-recognition systems, including evidence of weak correspondence between expressions and emotions in a prior assessment of emotion-recognition accuracy.
The stakes have broadened from products marketed as emotion sensing to automated inference of sensitive traits. Concerns around AI systems classifying race or ethnicity from faces show why a weak scientific premise can become a discrimination risk when deployed in market research or other decision-making settings.
First-order effects
- Developers training models on standardized facial-expression labels face a validity problem: output may reflect stereotypes embedded in labels rather than a person's actual emotional state.
- The reported retreat by Microsoft, IBM, and Amazon from related efforts makes emotion-inference vendors’ scientific substantiation and deployment claims more consequential.
Second-order effects
- Buyers of emotion-analysis tools may shift validation toward context-specific, human-reviewed use cases—or avoid using inferred emotion as an input to consequential decisions—because facial-only signals are not a dependable proxy.
- Suppliers may try to supplement video with physiological data, but the reported difficulty reaching high accuracy even with ECG/EEG signals limits the case that additional sensing alone resolves the underlying inference problem.
Third-order effects
- If this evidence continues to shape procurement and policy, the category could move from broad claims of reading emotion toward narrower tools that describe observable facial movements without asserting hidden mental states.
- The episode reinforces a governance divide between models that classify visible features and systems that infer sensitive personal attributes; the latter are likely to face stronger demands for scientific validity, bias testing, and use restrictions.
The trend: Emotion-sensing AI is being pushed from expansive human-state claims toward narrower, evidence-bound uses as scientific and discrimination concerns converge.