AI-based emotion detection has become a $20B industry, but some experts say the foundational science behind the tech is flawed and can adversely affect society
Machines can now allegedly identify anger, fear, disgust and sadness. ‘Emotion detection’ has grown from a research project to a $20bn industry Tweets: @dorotheabaur and @pernillet Tweets: Dorothea Baur / @dorotheabaur : #facialrecognition cannot differentiate between causation and correlation. If I'm refused a job because I have a ‘fake smile’ at an interview, maybe the smile has to do with back pain rather than my attitude towards the job. HT everyone #MachineLearning http://www.theguardian.com/... Pernille Tranberg / @pernillet : What are the most efficient methods to no be recognised by this? I am not talking about hiding from authorities in eg airports, but in general hiding from emotion detection. Guess we need masks #SurveillanceCapitalism http://www.theguardian.com/...
Context & Ripple Effects
Emotion detection began as a research project — Rana el Kaliouby's Affectiva built software to sense emotions from video — and has since scaled into a $20B industry selling anger, fear, disgust and sadness readings to buyers. The critiques in this piece are specific: Dorothea Baur argues facial recognition cannot separate causation from correlation, so a candidate's 'fake smile' at an interview might reflect back pain rather than attitude, while Pernille Tranberg asks how people can avoid being read at all.
The skepticism has hardened since. Later coverage reports [[a:963356|studies showing facial expressions match a person's actual emotions only 20%-30% of the time]], and the Wall Street Journal reports there are no universal expressions of emotion to train on — undermining the field's core premise. Yet adoption keeps spreading: emotion AI is now seeping into white-collar jobs to track workers' moods, raising privacy and bias concerns.
First-order effects
- Job candidates and employees are already being scored by systems whose readings can misattribute physical causes — back pain, fatigue — as emotional signals, exactly the failure mode Baur describes.
- Vendors in the $20B market are selling conclusions their own scientific base cannot reliably support, since expressions match felt emotions only 20%-30% of the time per the later studies.
Second-order effects
- Employers deploying these tools in hiring and workplace monitoring inherit the accuracy and bias problem as their own legal and reputational exposure, pushing some toward Tranberg-style countermeasures or avoidance rather than adoption.
Third-order effects
- If the no-universal-expressions finding holds, the industry faces a structural split: uses that can be validated against outcomes survive scrutiny, while facial-inference products sold as mind-reading become the sector's regulatory flashpoint — a pattern already visible as emotion AI moves into white-collar workplaces.
The trend: Emotion AI is scaling from lab demo into workplace enforcement infrastructure even as the underlying science of reading emotion from faces collapses under replication.