The computer can see the micro-expression. That doesn't mean we know what it means.
This post is a response to a recent literature review: Deception Detection Based on Video-Based Facial Micro-Expression Using Deep Learning and CNN by Deepthi Raj, Department of Computer Science and Engineering, Sahrdaya College of Engineering and Technology, Kerala, India. The paper is available at the link I've included with this post. I encourage you to read it alongside this response.
I want to start by saying something plainly: I have no quarrel with the researcher doing this work. The paper is technically competent, genuinely interesting as an engineering problem, and represents real intellectual effort. The author is solving a real problem — can a computer reliably detect subtle, involuntary facial movements that the human eye tends to miss?
That problem is worth solving. My issue is with the frame around it. Because when that paper gets read — and it will be read, and cited, and eventually quoted in procurement documents and training curricula — the takeaway will not be "interesting computer vision work." The takeaway will be: "micro-expressions are reliable windows into deception, and now machines can read them better than humans."
That conclusion does not follow. Not even close.
"The computer's accuracy is measured against human labels. But those labels are not ground truth. They are a theory."
What the technology is actually measuring
The deep learning models in this literature are trained on benchmark datasets — CASME II, SAMM, SMIC — where human coders have labeled facial movements using FACS, the Facial Action Coding System. FACS describes muscle activations: which units fired, in what combination, for how long. That part is relatively objective. The problem comes one layer up, when those muscle activations get assigned emotional meanings. "AU6 plus AU12 equals happiness." "AU9 plus AU17 equals disgust."
That interpretive leap — from muscle movement to emotional state — is Paul Ekman's theoretical contribution. It may be a reasonable approximation in some contexts. But it is a theory, not a fact. And when the CNN learns to classify "disgust," it is learning to replicate human coder judgments built on that theory. The model's reported accuracy — 87%, 97%, whatever the number — is measuring agreement with the labels. It tells you nothing about whether the labels correctly identify what a person is actually feeling, let alone whether what they are feeling means they are lying.
The paper solves a well-formed mathematical problem and then imports an entirely different, much harder claim. That is the framing problem.
The universality assumption that never quite held
Ekman's foundational argument was that basic emotional expressions are universal — cross-cultural, biologically hardwired, readable by anyone. The evidence for this was always more contested than the popular version suggested. The original studies used forced-choice methodology: show a photograph, pick the emotion from a list. Forced choice inflates agreement substantially compared to open-ended identification. When you remove the forced choice, cross-cultural agreement drops in ways that are hard to dismiss.
Lisa Feldman Barrett's constructionist work has pushed this further, arguing that facial movements do not carry fixed emotional meanings but are interpreted through cultural and contextual frameworks (I find her theory persuasive, and I say that as someone who has spent decades working across roughly 40 countries watching people communicate under pressure). You do not have to fully accept her position to recognize that the universality premise is genuinely disputed among people who study this for a living — and that building classification systems on top of it as though it were settled is a significant epistemic risk.
The frequency problem and the math that follows
In one of the most rigorous studies on this question, Porter and ten Brinke (2008) coded 700 high-stakes genuine and falsified emotional expressions. They found roughly 14 micro-expressions in the entire dataset — about 2% of all expressions coded. And those 14 were not clustered in the deceptive condition. They were distributed across both truthful and deceptive subjects at essentially equal rates.
Sit with that for a moment. In 700 high-stakes emotional expressions, the signal that is supposed to reveal hidden truth appeared 14 times — and showed no preference for deception when it did appear. Separate research has confirmed that micro-expression training has no measurable effect on deception detection accuracy. The impressive accuracy figures in the deep learning literature are measuring something real. They are just not measuring what the application context requires them to measure.
"Even if the computer correctly identifies a micro-expression of disgust — disgust at what? The question? The interviewer? A memory from last Tuesday?"
The meaning problem has no technical solution
This is the issue I keep returning to. Even granting everything else — that the classifier works, that the emotional label is correct, that the expression is genuinely involuntary — you still have not answered the question that matters. What does it mean in this interaction, with this person, in response to this specific stimulus?
An expression of contempt could mean the subject is lying. It could mean they find the question insulting. It could mean they hold the investigator in low regard. It could mean something happened last week that has nothing to do with your interview. The signal is informationally ambiguous in any real context, and resolving that ambiguity requires exactly the kind of human judgment, contextual knowledge, and relationship that no camera can replicate.
This is not a problem the next generation of models will solve. It is a structural feature of what micro-expressions are.
The circular reasoning that holds the whole thing up
Here is what concerns me most about how this literature accumulates. Paper A assumes the deception-microexpression link and cites Papers B, C, and D. Papers B, C, and D also assumed it, and cited Papers E through M. At no point in that citation chain does anyone go back to the foundational claim and ask whether it has been independently validated in ecologically valid conditions with adequate controls. The assumption gets laundered through volume. Enough papers citing each other creates the appearance of an established fact.
I have worked in this space for more than two decades. I helped develop interview and interrogation protocols that are now used across the intelligence community. The practitioner literature has been moving steadily away from behavioral cue approaches — including micro-expression reading — precisely because the reliability evidence does not support them at the level required for consequential decisions. The science-based interviewing movement exists in direct response to this problem.
When someone reads a paper like this and thinks "maybe Colton was wrong about micro-expressions, the technology has caught up" — that is the framing problem made real. The technology has genuinely advanced. The underlying validity question has not changed. And the danger is that sophisticated engineering wrapped around a weak foundation makes the foundation look stronger than it is.
What should be done differently
None of this means the research is worthless. Automated micro-expression detection could be genuinely useful — in clinical emotion research, in human-computer interaction, in studying communication patterns in contexts where ground truth is available and the stakes of error are low. The engineering work deserves to be evaluated on its own merits.
What it should not do is import the deception detection frame without defending it. That means explicitly acknowledging the contested status of the universality claim. It means distinguishing between classifier accuracy and predictive validity for deception. It means grappling with base rates and the reality that 14 occurrences across 700 expressions — none of them diagnostic — is not a foundation on which to build an investigative tool. And it means recognizing that the meaning of an expression — the question that actually matters for any investigative application — remains an open problem that the technology does not and cannot close.
The computer can see the micro-expression. We still do not have ground truth for what it means. Those are two very different things, and the gap between them is where real harm gets done.
