The previous article ended on an asymmetry. Where influence is cheap, exit is expensive, and nowhere is exit more expensive than at work. A worker who dislikes a feed can close the tab. A worker who dislikes the sensor in the truck cab, the click counter on the laptop, or the pick-rate number on the handheld scanner cannot close the shift.
The tools have arrived quietly and in fragments, and they present themselves as management software. The interesting question is not whether an employer could one day read a mind. It is what happens now, with instruments that measure attention, fatigue, and the physical correlates of emotion, and with the organizational machinery that turns those measurements into decisions about pay, hours, and continued employment.
A ledger of what can actually be sensed
Separating capability from marketing is the first task, because vendors blur it constantly.
Measurable with reasonable confidence. Keystroke counts, click timing, application focus, location, badge events, vehicle telemetry, and heart rate are all direct physical signals. If a device records it, it happened. Nobody is inferring anything the machine did not observe.
Inferable with moderate confidence. Drowsiness is the best-supported inference in this category. Eye closure, blink duration, steering corrections, and lane position are reliable enough that regulators treat fatigue detection as a safety technology rather than a surveillance one. The European Union’s AI Act makes that judgment explicit: it excludes physical states such as pain or fatigue from its definition of emotion recognition, naming systems that detect “the state of fatigue of professional pilots or drivers for the purpose of preventing accidents” as outside the definition.
Inferable with weak confidence. Emotional state, engagement, motivation, attitude, and intent sit at the far end. These are not observed; they are constructed from facial movement, voice, typing cadence, or activity patterns by models trained on someone else’s labeled data. The construct is where the trouble starts.
Why the inference layer is the fragile part
In 2019, Lisa Feldman Barrett and four co-authors published a review in Psychological Science in the Public Interest that should have ended a commercial industry. Surveying the evidence behind the “common view” that faces broadcast emotion, they found that people do sometimes smile when happy and scowl when angry more often than chance would predict, and that how people communicate anger, disgust, fear, happiness, sadness, and surprise varies substantially across cultures, situations, and even across people within the same situation. Similar configurations of facial movement express more than one emotion category, and a given configuration, such as a scowl, often communicates something other than an emotional state.
For a vendor selling emotion detection, that is a problem without a workaround. A camera can measure the geometry of a face. The claim that the geometry means frustration requires a relationship between face and feeling that does not hold reliably enough to be diagnostic.
Fatigue is different, which is why the regulatory distinction is not arbitrary. Eye closure and steering drift are physical events with mechanical consequences. Being wrong about them means a truck leaves the road. Being wrong about engagement means a person gets a lower score for reasons nobody can reconstruct.
Motivation is the weakest signal of all, and the most consequential. Motivation is partly a story people tell about themselves, partly a drive that behavior samples poorly, and partly a response to being observed. Any system that claims to measure it is really measuring the behavior it can see and naming the residual. That naming decision carries weight, because a manager who receives a “motivation score” will treat it as information rather than as a model output.
The rules that already bind employers
A surprising amount of this is regulated already, mostly in Europe and mostly in the last two years.
The AI Act prohibits AI systems used to infer the emotions of a natural person in the workplace and in education, except where the system is intended for medical or safety reasons. The prohibition applies since February 2025. The Commission’s interpretive guidelines, adopted in July 2025, read the safety exception narrowly and treat “workplace” broadly, covering the whole employment relationship from recruitment through dismissal, and covering both physical and virtual workplaces.
Two details matter more than the headline. First, the physical-state carve-out is real and wide. Because the Act’s definition of emotion recognition excludes pain and fatigue, a system that watches for drowsiness does not fall under the emotion ban even though it reads a body without consent. Second, the medical and safety exception does not obviously cover wellbeing, stress, boredom, or job satisfaction. A vendor relabeling an engagement monitor as a wellness tool is making a legal argument, not a semantic one, and the argument is testable.
Alongside the prohibition sits a classification regime. Annex III of the Act treats AI systems used in employment and worker management as high-risk. That covers recruitment and selection, decisions on terms, promotion and termination, task allocation, and monitoring or performance evaluation. High-risk status brings documentation, logging, transparency, human oversight, and accuracy obligations. The posture is worth noticing: the legislature concluded that these systems work well enough to be dangerous and not well enough to be trusted.
The platform work directive adds harder edges for gig work. Directive (EU) 2024/2831, adopted in October 2024 with a transposition deadline of December 2026, restricts what platforms may process through automated monitoring and decision systems. Article 7 forbids processing personal data about a worker’s emotional or psychological state, private conversations, data collected while the person is not performing platform work, inferences about protected characteristics or trade-union activity, and biometric identity matching against a database. Article 9 requires transparency. Article 10 requires human oversight by staff with the training, competence, and authority to override automated decisions. Article 11 bars decisions to restrict, suspend, or terminate an account from being based solely on automated processing.
In the United Kingdom, the Information Commissioner’s Office published guidance on monitoring workers in October 2023 requiring that monitoring be lawful, necessary, proportionate, transparent, and achieved through the least intrusive means available, with a data protection impact assessment and worker consultation expected. The enforcement record shows the guidance has teeth. In February 2024 the ICO ordered Serco Leisure, Serco Jersey, and seven associated community leisure trusts to stop using facial recognition and fingerprint scanning for attendance across 38 leisure facilities, covering the biometric data of more than 2,000 employees. Workers had not been offered a clear alternative and the system was presented as a condition of being paid. The Commissioner’s summary was blunt about why that failed: you cannot reset someone’s face or fingerprint the way you reset a password. Less intrusive options, in this case identity cards, were available the whole time.
Consent inside a hierarchy
The Serco case exposes the mechanism that makes workplace sensing different from consumer sensing, and it is the same mechanism Huxley identified in Brave New World Revisited when he argued that control through reward outlasts control through punishment.
Formal consent is close to meaningless when one party sets the terms of employment. A worker can agree to a wearable, and the agreement can be genuine in the sense that the worker signed it and prefers signing to being fired. The relevant question is what refusal would cost. If declining removes someone from the pool considered for promotion, moves them off preferred shifts, or marks them as difficult, the option to refuse exists on paper and not in the wage relationship.
Huxley’s insight was that this kind of arrangement is more stable than coercion precisely because nobody has to be threatened. The environment is arranged so the preferred behavior is the convenient one. A dashboard that ranks workers by activity makes being visible the path of least resistance. A scheduling algorithm that rewards responsiveness to off-hours messages makes responsiveness normal. Over time, the baseline shifts, and the worker who once chose monitoring for an edge finds that the unmonitored state is now the suspicious one.
The same logic explains why the safety argument deserves to be taken seriously rather than dismissed as a pretext. Detecting a drowsy long-haul driver protects the driver and everyone sharing the road. Refusing that technology on principle would be a strange position for anyone who has read the fatigue literature. The question is not whether the sensing is legitimate. It is who holds the data, what else it gets joined to, and whether the person being measured can see the measurement and contest it.
Those three questions have answers in the strongest regulations. Article 10 of the platform work directive requires human oversight with actual authority to override. Article 11 bars solely automated termination. The ICO’s necessity and proportionality test asks whether a less intrusive method reaches the same goal. Each of these is a way of saying that the sensor can exist while the decision about the person stays with a person who can be held responsible.
What monitoring reveals about productivity metrics
Workplace sensing is usually justified by productivity, so it is worth asking what productivity research actually measures.
Gloria Mark, Daniela Gudith, and Ulrich Klocke ran a controlled interruption experiment reported at CHI 2008. Forty-eight participants handled a simulated email workload. Interrupted participants finished faster than uninterrupted ones, roughly 20.3 to 20.6 minutes against 22.8, with no significant difference in errors or politeness in their replies. The cost showed up in how the work felt: reported stress rose from about 6.9 on a 20-point scale to 9.5 for same-context interruptions and 9.1 for different-context ones, alongside higher frustration, time pressure, and effort.
The widely repeated claim that it takes 23 minutes and 15 seconds to recover from an interruption does not come from that paper. It traces to a 2006 interview, and it has been misattributed for years. What the CHI 2008 study shows is stranger and more useful: people compensate for interruptions by working faster, and the compensation is paid in stress rather than in output quality. An activity monitor that counts keystrokes and window switches would see the faster completion and none of the cost.
Mark’s earlier field study with Victor González and Justin Harris, presented at CHI 2005, observed 24 information workers for more than 700 hours of timed observation. They spent an average of 11 minutes 4 seconds in a working sphere before switching. Fifty-seven percent of their working spheres were interrupted. Of the interrupted work, 77.2 percent was resumed the same day, after an average of 25 minutes 26 seconds and 2.26 intervening working spheres. Just over half of the interruptions were self-initiated.
That last number matters for anyone designing an attention metric. People interrupt themselves to manage their own work. A monitoring system that penalizes context switches is penalizing a behavior that the workers in the study were using to cope with a workload the system had already made fragmented. The metric would punish adaptation and reward the appearance of continuous focus, which is not the same thing as focus.
What the evidence says about the people being measured
The clearest findings here are about how workers experience monitoring rather than about its productivity effects, and the experience is consistently negative.
The Institute for the Future of Work published a report in December 2024, Data on Our Minds: Affective Computing at Work, led by Phoebe Moore and Gwendolin Barnard. It named the emerging practice “algorithmic affect management” and documented how affective technologies inferring emotion are being wired into systems that make decisions about shift patterns, rates of pay, and performance ratings. Among the report’s survey findings, nearly 40 percent of workers reported being subject to technologies collecting information about their affective states in connection with health, safety, and wellbeing, and across the measures surveyed a majority did not agree that those technologies had a positive psychosocial impact. The authors observed that monitoring introduced on wellbeing grounds can itself produce technostress.
Big Brother Watch’s September 2024 report on bossware documented the practical forms: biometric sign-ins in construction, AI fatigue monitoring of National Express coach drivers at the wheel, keystroke and click logging from tools such as Teramind, pick rates tracked on handheld devices at Royal Mail and supermarkets, productivity scoring in Microsoft Teams, and emotion detection in hiring platforms. It also cited survey evidence on how this lands. A 2022 Trades Union Congress poll suggested at least three in five workers had been monitored by an employer. UNISON Scotland research found workers in electronically monitored jobs overwhelmingly describing the experience as demeaning, with more than half reporting stress and anxiety, 17 percent reporting depression, and 52 percent considering resignation. The report was careful to acknowledge that some monitoring is justified, including safety-critical fatigue detection and proportionate administration, which is what makes its criticism credible.
The prevalence picture remains murky. A 2022 review by the UC Berkeley Labor Center, by Lisa Kresge and colleagues, found that firm-level adoption studies are uneven and that understanding of how widespread workplace management technologies actually are remains weak. Claims that a majority of employers now run sophisticated monitoring rest on weaker evidence than their confident repetition suggests. The Institute for the Future of Work cites a 2019 survey of 239 large firms in which more than half reported some form of non-traditional monitoring, and notes a tripling in bossware use over five years. That is suggestive of a real trend in large firms rather than a description of the average workplace.
The policy literature has begun converging on a specific diagnosis. A July 2025 report from the National Employment Law Project, by Irene Tung and co-authors, argues that bossware intensifies existing job quality problems — harmful disciplinary practices, job precarity, lost autonomy, unfair scheduling, discrimination, and suppression of collective action — rather than creating a wholly new category of harm. That framing is more useful than a privacy-only framing, because it points at labor law, health and safety law, and anti-discrimination law as the instruments that already exist.
Where the speculative version begins
It is worth being explicit about what has not happened, because the vocabulary of neurosurveillance invites inflation.
Demonstrated. Cameras, wearables, and device telemetry can reliably record physical signals; drowsiness can be inferred from them well enough to prevent accidents; affective inference from faces and voices performs poorly against any rigorous standard.
Plausible engineering. Closed-loop systems that sense a state and respond to it — adjusting a schedule, prompting a break, dimming a screen — are buildable now and partly deployed. The interesting risk in this category is not mind reading but automated management: a system that infers something imprecise and then acts on it with the authority of a decision.
Speculative. Direct reading or writing of neural states through consumer-grade hardware in the workplace has no deployed basis. It is a research direction and a plausible medium-term engineering problem for some applications, and it should be described that way rather than as a present capability. Articles about “the AI boss inside your nervous system” are usually about the measurement layer, not the nervous system.
The practical consequence of that gap is that the debate about neural interfaces distracts from the systems that are already making decisions. A crude attention metric that triggers an automated warning is affecting a worker’s week today. A hypothetical electrode array is not.
Concrete tests for the next version of this workplace
The useful questions here are the ones with evidence attached on both sides.
Whether the inference holds up. An emotion or motivation system should be able to show that its output tracks the thing it claims to measure, in the population being monitored, under the conditions of the actual workplace, with published accuracy rather than a vendor case study. Given the Barrett review’s findings on cross-cultural and situational variability, any such claim needs to survive testing outside the population it was trained on. Most deployed systems have not been evaluated this way.
Whether refusal carries cost. The Serco enforcement notices turned on the absence of a real alternative. A workplace that offers monitoring alongside an equally viable unmonitored path is a different institution from one that offers monitoring as the price of full participation. This is testable by looking at what happens to the people who decline.
Whether a person can contest the output. Article 11 of the platform directive bars solely automated termination decisions, and the ICO’s guidance limits solely automated decisions with legal or similarly significant effects. The operative question is whether there is a human with the standing, the information, and the authority to reverse the system. A human who rubber-stamps the score is a formality.
Whether the goal survives a cheaper test. The necessity and proportionality standard asks whether a less intrusive means would achieve the same purpose. Applied to an attendance system, that ruled out biometrics where a fob would do. Applied to productivity analytics, it would require showing what the invasive version delivers that a supervisor conversation or a work sample would not.
Whether the environment pushes back at all. Australia’s right to disconnect, in force since August 2024 for employers with 15 or more staff and since August 2025 for smaller employers under the Fair Work Act 2009, gives employees a right to refuse to monitor, read, or respond to employer or third-party contact outside working hours unless the refusal is unreasonable, with the reasonableness test weighing the reason for contact, its disruptiveness, compensation, the employee’s role, and their personal circumstances. The right is a protected workplace right, and disputes escalate from workplace discussion to the Fair Work Commission. It is a modest reform and an instructive one: it does not ban contact, it gives the employee a defensible position from which to say no. That is the structural equivalent of what the workplace sensing debate needs. Not a prohibition on measurement, but a floor under refusal.
Where this leaves the argument
The instrumentation of work is proceeding on the strength of two claims that do not hold equally well. The first, that monitoring improves productivity, is weakly supported and partly contradicted by research showing that people absorb interruptions as stress rather than as lost output. The second, that emotional and motivational states can be read from behavior, is the foundational claim of an entire vendor category and is close to unsupported.
What is well supported is narrower and more actionable. Physical signals can be measured, and some of that measurement saves lives, which is why safety monitoring earns its exception. Organizations with the power to measure also have the power to make refusal expensive, which is why consent is a weak shield. And the experience of being measured is itself a health and safety variable: the evidence on stress, anxiety, and intention to resign is among the more consistent findings in this literature.
There is also a deeper cost that does not fit neatly into a compliance framework. A workplace that quantifies attention teaches people to produce the appearance of attention. A workplace that scores engagement cannot tell the difference between someone who is invested and someone who has learned what investment looks like to a model. Those systems do not just observe the workplace; they change what competence means inside it.
The Huxley comparison lands here in a specific way. The soft-control apparatus in a workplace does not need to compel anything. It needs a metric, a default, and a promotion process. The result is a workforce that has adapted to being measured, mostly without noticing the adaptation, and a set of institutions that would struggle to explain what the measurement was for. The remedy is correspondingly unromantic: oversight with real override authority, necessity tests that survive the existence of a cheaper alternative, transparency about what is collected and what it feeds, and a defensible way for a person to say no. None of that requires settling whether machines can read minds. It only requires deciding who is allowed to act on what a machine says about one.
Sources and further reading
- Regulation (EU) 2024/1689 (AI Act), Article 5 prohibited practices, Annex III high-risk uses, and Recital 18 on emotion recognition and physical states: AI Act Service Desk
- Directive (EU) 2024/2831 on improving working conditions in platform work, including restrictions on automated monitoring and decision systems: EUR-Lex and EU-OSHA summary
- Information Commissioner’s Office, “Employment practices and data protection: monitoring workers” (2023): ico.org.uk
- Information Commissioner’s Office, enforcement notices against Serco Leisure and associated trusts over biometric attendance monitoring (23 February 2024): ico.org.uk
- Lisa Feldman Barrett, Ralph Adolphs, Stacy Marsella, Aleix M. Martinez and Seth D. Pollak, “Emotional Expressions Reconsidered: Challenges to Inferring Emotion From Human Facial Movements,” Psychological Science in the Public Interest 20(1), 2019: journals.sagepub.com
- Gloria Mark, Daniela Gudith and Ulrich Klocke, “The Cost of Interrupted Work: More Speed and Stress,” CHI 2008: ics.uci.edu — and Gloria Mark, Victor M. González and Justin Harris, “No Task Left Behind? Examining the Nature of Fragmented Work,” CHI 2005: ics.uci.edu
- Phoebe V. Moore, Gwendolin Barnard and Anna Thomas, Data on Our Minds: Affective Computing at Work, Institute for the Future of Work, December 2024: ifow.org
- Silkie Carlo and Jake Hurfurt, Bossware: The dangers of high-tech worker surveillance, and how to stop them, Big Brother Watch, September 2024: bigbrotherwatch.org.uk
- Irene Tung, Paul K. Sonn, Maya Pinto, Sally Dworak-Fisher and Josh Boxerman, “When ‘Bossware’ Manages Workers,” National Employment Law Project, July 2025: nelp.org
- UC Berkeley Labor Center, “How Common is Employers’ Use of Workplace Management Technologies? A Review of Prevalence Studies,” November 2022: laborcenter.berkeley.edu
- Fair Work Ombudsman, “Right to disconnect” (Fair Work Act 2009, sections 333M–333Q): fairwork.gov.au
Loading comments…