The cost of watching how you feel just collapsed

The story being told about emotional surveillance at work is that AI got smart enough to read faces and voices. That is not quite right. The technology has existed for twenty years. What changed is the price. Vision and audio models have been commodified into cloud APIs, and inferring a worker's affect from an existing call recording or video feed is now a near-free add-on to infrastructure most employers already own. The Atlantic's feature this week, the ILO's psychosocial-risk report from April, and the Meta keystroke disclosure last week are not three separate stories. They are one story about what happens when a surveillance capability stops being expensive.

I want to be specific about who this lands on, because the distributional pattern is the argument.

The cost barrier was never technical. Affectiva, the company that pioneered commercial emotion recognition, was founded in 2009. Cogito, whose real-time emotional guidance product now runs across major US insurance and financial services call centres, including MetLife's, was founded in 2007.¹ For most of that history, deploying these tools meant bespoke integration, dedicated hardware, and per-seat licensing that only the largest call-centre operators could justify. The constraint was economics, not capability. That constraint is gone. Any employer with existing video conferencing or call-recording infrastructure can now bolt on emotional inference at marginal cost, because the foundational vision and audio models doing the heavy lifting are commodity cloud services priced by the API call.

This is the same dynamic as every previous surveillance escalation. CCTV became ubiquitous when cameras got cheap. GPS fleet tracking became ubiquitous when chips got cheap. The capability existed long before the deployment; what changed was the unit cost. Framing the current moment as "AI novelty" misses that the constraint was always economic, and obscures the fact that the policy window for shaping deployment is the window where costs are falling, not the window where capabilities are being invented.

That is the workplace surveillance software market, 2023 to 2032 projected. A compound annual growth rate of roughly fourteen percent. The growth is not coming from existing customers spending more on existing tools. It is coming from the marginal employer for whom these tools were not previously economic.

Who actually pays

The workers most exposed are the workers least able to refuse. The Atlantic feature, and the vendor case studies, and the ILO's evidence base all concentrate on the same populations: call-centre agents, warehouse workers, gig drivers, delivery couriers. Cogito's customer list is insurance and financial services contact centres. Amazon's Netradyne camera system, which captures driver facial expressions in delivery vans, was on 100,000 vehicles by 2022.² Uber, Lyft, and Amazon Flex use facial recognition check-ins and in-trip monitoring as a condition of platform access.

These are not the populations most likely to be in the room when deployment is decided. Call-centre worker unionisation in the US sits at roughly three to four percent. Warehouse worker unionisation around five percent. Gig platform workers under one percent in most jurisdictions.³ The workers least exposed to emotional surveillance are senior knowledge workers and executives, whose work is harder to monitor in this way, and whose cultural and political cost of being monitored is high enough that employers do not try.

This matters because the public framing of workplace AI tends to assume a generic worker. The generic worker, in most coverage, is a software engineer worried about whether the model will write better code than they do. The generic worker, in the surveillance story, is the call-centre agent whose live voice is being scored for emotional tone by a system the agent cannot inspect, audit, or appeal.

A 2023 Cornell ILR study of call-centre workers under AI-assisted monitoring found a twelve-percentage-point increase in self-reported anxiety and a nine-point increase in supervisor distrust.⁴ That is observational, and I will come back to the limits of the evidence. But the directional finding is consistent with what the ILO is now naming explicitly as psychosocial risk.

The consent fiction

Employment consent was always thin. Workers "agree" to surveillance by taking the job; specific terms appear in onboarding paperwork most workers do not read and could not negotiate if they did. For keystroke counting, this was arguably sufficient, in the narrow sense that a worker can comprehend what is measured even if they cannot refuse it. Keystrokes are countable things. Emails sent are countable things. Idle time is a clock.

Affective inference is not like this. A system that classifies a worker's emotional state from facial micro-expressions or vocal pitch is making inferences the worker cannot verify, from signals the worker cannot consciously control, against benchmarks the vendor does not publish. The gap between what the worker is told ("we use AI to support quality and coaching") and what is actually inferred ("this agent's frustration index in Q3 was 0.42, above the team median") is so large that consent becomes a legal fiction.

The UK's ICO (Information Commissioner's Office, the data protection regulator) said as much in its 2023 employment monitoring guidance: inferring emotional states from biometric signals likely constitutes processing of special-category data under UK GDPR, requiring explicit consent at a high bar.⁵ How widely vendors meet this bar is not documented. In the US, there is no equivalent federal standard. Illinois' BIPA (Biometric Information Privacy Act) is the strongest existing hook, but courts are still settling whether real-time voice sentiment inference counts as a covered biometric identifier.⁶ California and Colorado's AI employment laws cover automated employment decisions; they do not clearly cover real-time monitoring. NYC Local Law 144 covers hiring, not continuous assessment.

The EEOC (Equal Employment Opportunity Commission, the federal employment discrimination regulator) has issued no guidance specifically on affective AI since the previous administration's AI employment guidance lapsed. Federal posture is effectively permissive.⁷

So the regulatory map looks like this: a strong European standard that vendors may or may not be meeting, a patchwork of state laws that cover adjacent but not exactly the same use cases, and a federal vacuum. The gaps in the patchwork sit exactly where the affected workers are concentrated: gig workers in states without biometric privacy law, warehouse workers in states without AI employment statutes.

What the vendors say back

I want to be fair to the counter-case, because there is one.

Cogito and similar vendors contest the "surveillance" framing directly. Their argument is that the tool is worker-facing, not manager-facing: the agent sees real-time coaching prompts about how they are coming across to the customer, while supervisors see aggregated quality scores rather than the underlying emotional inferences.¹ In some deployments this is genuinely the architecture. Whether the distinction holds at scale, under pressure from operations managers who would like more granular data and from sales teams that would like to offer it, is contested.

There is also a legitimate labour argument that documented, consistent performance measurement can benefit workers by replacing arbitrary managerial judgment. If a supervisor's complaint that an agent "has a bad attitude" can be checked against a system that scores tone, the worker has at least a paper trail to contest. I take this seriously. It is the strongest version of the case for these tools, and it is not nothing.

But it depends on workers having genuine access to the data, genuine ability to contest it, and genuine alternatives if the contest fails. None of these conditions hold for the modal call-centre agent or warehouse worker in 2026.

And the evidence on accuracy is not where the marketing implies. A 2021 Cambridge study found that emotion recognition tools performed seven to twelve percentage points worse on darker-skinned faces and non-native English speakers.⁸ Vendors point to internal validation showing improvement; those studies are not independently peer-reviewed. The honest position is that the accuracy gap may have narrowed but the demographic disparity has not been demonstrated to be solved.

What to watch

The European AI Act is the closest thing to a constraint. It classifies many workplace AI uses as high-risk and explicitly restricts emotion recognition in employment contexts. The pressure point will be whether US vendors maintain different product configurations for EU and non-EU customers, and whether that asymmetry becomes politically visible to non-EU workers.

State attorneys general are the active enforcement layer in the US. With the federal level quiet, BIPA-style litigation in Illinois and statutory enforcement in California and Colorado are doing the work. The number of BIPA class actions hit a record 1,270 cases in 2023.⁶ Watch whether voice sentiment inference becomes the next contested category.

The ILO's psychosocial-risk framing opens a second regulatory door. Occupational health and safety regulators in many jurisdictions have authority that data protection regulators do not. If continuous affective monitoring is documented as a workplace health hazard rather than only a privacy concern, the enforcement geography changes.

Efficiency for the operator is precarity for the worker, and the gap between the two is now cheap enough to industrialise.

I do not think the case against affective workplace surveillance requires strong causal evidence of harm, though the evidence is directionally clear. The case rests on the asymmetry. The worker cannot inspect what is being inferred about them. The worker cannot refuse without losing the job. The worker cannot meaningfully consent to inferences they cannot anticipate. And the workers most exposed are the workers with the least voice in deciding whether this is the kind of workplace they want to be in.

The cost of building that kind of workplace has just collapsed. That is the story. Whether the people on the receiving end have any say in it is the question the next year of policy and litigation will answer, mostly without them in the room.

Footnotes

¶