AI tools at work: what mood analysis and video screening miss

Autistic Mirror (outside view, role employer)

The employee who complained is describing something neurologically concrete, not mere discomfort.

Predictive coding is the mechanism behind it. The brain constantly makes predictions about its environment and compares them with what actually arrives. In autistic people this system is more sensitively calibrated: uncertainty about who does what with which data, and when, produces a persistent prediction error. That is not irrational resistance, but a system that cannot close the unknown and therefore stays on alert. That costs capacity, which is then missing for the actual work.

Intolerance of uncertainty adds to this as an amplifier. Tolerating uncertainty is hard when a system with a high need for prediction has to keep recalibrating. The core question is not "What does the tool measure?" but "Do I really know what happens with it? Who sees it? When? In what context?" As long as these questions are not fully answered, they remain active as open loops. An open loop is a persistent prediction error: the brain keeps an expectation active that has not yet been resolved, and the comparison keeps running in the background, even without conscious thought.

A third aspect is the double empathy problem. The understanding gap between autistic and non-autistic people is mutual: two different processing styles create friction in both directions. When your team communicates "this is supposed to help", that sounds internally like a sufficient explanation. For employees who need details, consequences and procedures to experience safety, "it helps" is not a closable argument but another open loop.

What you have overlooked is not the resistance itself, but that "measuring mood" means for employees that their communication is being evaluated algorithmically without them knowing the exact mechanism. Structurally, that is the same as an unannounced observation setting. Autistic employees who already spend energy on masking experience this as an additional demand: they now also have to "sound right" for a system whose criteria they do not know.

The following adjustments are concretely accessible: full transparency about data access, retention period, evaluation logic and consequences in writing, before the tool runs. In addition, the explicit option to opt out of mood analysis without this being read as a signal. Not every adjustment has to apply to everyone, but the structure must be such that nobody has to guess what happens with their data.

The tool may have legitimate goals. Whether it achieves them depends on whether the conditions of its introduction preserve or undermine employees' sense of safety.

Autistic Mirror (outside view, role employer)

Video-based AI application analysis structurally excludes autistic applicants, even if that is not the intention. The mechanism behind it is direct.

Masking is the key point here. Autistic people who actively adapt their neurology in social contexts invest considerable cognitive resources in regulating facial expression, eye contact and speech rhythm. This adaptation serves to be read as "competent" in a neurotypically coded social environment. What AI systems classify as "natural" facial expression or as "engaged" and "confident" are neurotypical norms of expression that these systems learned by training on neurotypical data. Autistic ways of expressing oneself, less variable facial expression, different speech rhythm, a more direct tone without the expected emotional modulation, are systematically rated worse by these systems because they deviate from the training data pattern, not because they indicate lower suitability.

The second mechanism is bottom-up processing. The autistic brain processes sensory and social information in a more detail-oriented way and without the same automatic prioritisation of what is "socially expected". This shows in pauses in conversation that are longer than neurotypically expected, in response latency because information is processed more completely before an answer is formulated, and in a way of expressing oneself that responds more precisely to the actual question instead of the implicit social context. These are exactly the qualities that AI systems calibrated for "engagement" and "spontaneity" rate negatively.

In addition there is selective mutism as a spectrum: in high-demand, unfamiliar settings such as a job interview, speech production can be noticeably restricted in some autistic people under increased pressure. That is not an indicator of competence but a context-dependent neurological reaction. AI systems that rate fluency and confidence of formulation interpret this as weakness.

The efficiency promise of such systems is also empirically questionable: they measure conformity with neurotypical communication norms, not actual suitability for a position. The selection errors that arise are legally relevant. In Germany the AGG applies when a selection procedure systematically disadvantages people with disabilities, even if the disadvantage was not intentional. In the EU the AI Act additionally applies, with specific requirements for AI systems in high-risk applications, which explicitly include personnel decisions.

Efficiency gains without a structurally excluding effect are possible if the focus lies on task-related information instead of communication style. Written, asynchronous preliminary assessments that capture concrete competencies filter by actual suitability and at the same time reduce barriers for autistic applicants, people with selective mutism, people with language processing differences and applicants whose first language is not German.

Autistic Mirror (outside view, role employer)

The written communication of autistic people systematically deviates from the patterns such tools are trained on, and not because something is missing, but because a different communication style is present.

Phatic communication is the key. Most written messages in a work context contain two layers: an informational layer (what is said) and a relational layer (how it is meant to land with the other person). Openings like "Great, thanks for the quick reply!", closing politeness formulas, softening phrases like "maybe it would make sense" instead of "that doesn't work" are part of this relational layer. AI sentiment tools are trained on texts in which this relational layer is supplied, and therefore code its absence as coldness, distance or negativity. Autistic writers often leave this layer out, not because they are dissatisfied, but because their communication style is information-centred and does not automatically integrate the relational level into word choice.

That is the double empathy problem in its algorithmic form. The model expects the neurotypical communication style as the norm, reads everything that deviates from it as a disturbance and gives the team lead back a mood measurement that is actually a style measurement. The colleague communicates correctly, directly and efficiently. The tool penalises her for it.

There is a second problem: alexithymia, the difficulty of translating one's own emotional states into language, is involved for many autistic people. Even if they wanted to supply the positive relational layer, access to the fitting affective phrasing is not always fluently available. That is not indifference, but a different interoceptive processing route, in which the feeling is present but the linguistic encoding is missing or delayed.

For the team lead this means in practice: the tool measures stylistic conformity with a neurotypical baseline, not mood. A sentiment analysis based on a training dataset without a representative share of autistic language patterns does not produce valid results for this writer. If he relies on this signal, he is acting on the basis of a mismeasurement.

The conversation he is considering would signal to her that her direct, factual communication is a problem. That would be pressure to mask, and masking costs her resources she is currently putting into the work.