Fix the Feeling, Not the Face: When “Emotion Recognition” Becomes Emotional Erasure
They call it a win:
A recently published meta-analysis from Hunan Normal University finds that virtual reality, desktop games and wearable apps “significantly improve” social-emotional skills in autistic people. Their metric? A 0.897 effect size in teaching autistic participants to recognize, express and regulate emotions — the right ones, in the right way, at the right time.
But for whom?
The paper, published this week in Journal of Intelligence by Yunshan Liu, Sirao Li, Yaping Huang and Dan Li, begins with a claim that frames everything: “Individuals with autism spectrum disorder have deficits in socio-emotional competence.” Not differences. Not divergent ways of being. Deficits.
According to these scientists, I'm socio-emotionally incompetent. Ouch. That hurts. But here's the reality: If you judge a spoon by how well it cuts steak, it's going to disappoint you every time. That doesn’t mean the spoon is useless. It means you misunderstood what it was built for. It means the measuring tools are broken.
From the frame used in this study, every tech tool becomes a treatment. Every facial expression becomes a test. Every misalignment becomes a malfunction.
The authors cite improvements in “emotion recognition” — but define it as matching facial expressions to neurotypical norms. They praise tools like Emotiplay and Mind Reading, both built from Baron-Cohen’s model of emotional correctness. They elevate desktop interventions over VR because the results are more controllable, less messy, more easily monitored. And they celebrate professional implementers because trained researchers are more likely to stick to the script.
This is not emotional support. It’s emotional erasure.
None of the 25 studies included asked autistic participants what emotional safety means to them. None questioned the premise that autistic affect is inherently flawed. Most used children. Most measured outcomes through compliance-based checklists. The goal was never autonomy — only acceptability.
When autistic people resist social performance, we’re told it’s dysregulation. When we mask it well enough for the data to look clean, it’s called success. But there's a steep cost associated with masking.
This is the diagnostic-industrial complex in motion: pathologize difference, build tools to correct it, call it care. But “high efficacy” in this context means high adaptability to non-autistic expectations. The more we contort ourselves into mimicking neurotypical emotional fluency, the more the system congratulates itself.
It’s no accident that Hunan Normal frames tech as the answer to therapy shortages. When real care is out of reach, correction becomes a substitute. This isn’t innovation. It’s austerity in disguise. It’s about lowering the cost of control.
Before we celebrate efficiency, we have to ask: efficient for whom? And what’s the cost to those whose movements — or feelings — can’t be translated into training sets?
If a tool teaches an autistic child to smile on cue but never asks whether they feel safe, it’s not support. It’s gamified neurotypical simulation.
Let’s be clear: we don’t need more machines telling us when to emote. We need fewer systems that punish us when we don’t.
If you’re a researcher reading this: you may have entered this work with care and curiosity. But care requires reflection. We need new questions — and new metrics — rooted in safety, autonomy, and lived experience. That work is still possible. But it begins by listening.