Picture this: it’s 9:47am on a Tuesday. Somewhere in your organization, an algorithm has already flagged three employees as “low engagement” based on their keystrokes, assessed a fourth as “emotionally disengaged” from their morning standup, and generated a productivity risk score for an entire team — all before anyone has emptied their inbox. In many jurisdictions, none of this is illegal.

This is the new landscape of workplace surveillance: not the visible instruments of a previous era (think swipe-card readers or CCTVs) but a largely invisible architecture of observing that AI has made possible. Technological systems do not merely reflect existing social arrangements; they actively redesign them, encoding the assumptions of their creators while projecting “objectivity” that makes them hard to contest. When an algorithm decides who is “productive,” it is not making a neutral measurement — it is making a consequential and political one.

Furthermore, search and classification systems systematically surface discriminatory outputs while hiding the human choices that produced them. The workplace is not a different domain: it is the same dynamic with higher stakes for the employees themselves, given their need to remain employed.

The shift AI represents is not merely one of degree. It is a shift in kind. For ethics, compliance, and risk professionals, this distinction matters enormously. We are accustomed to navigating surveillance questions through the lens of legal permissibility: what does the Privacy Act require? What did employees consent to in their employment contracts? These are necessary questions, but they are not sufficient ones. The law, as it so often does with emerging technology, lags behind this technology’s capabilities and the ethical risks they create. We urgently need an ethical architecture for these decisions that precedes deployment, engages employees as moral stakeholders rather than data sources, and asks harder questions than compliance allows for.

What AI changes — not just scale, but kind

AI-enabled workplace monitoring ostensibly observes performance, manages productivity, and protects assets. But this idea obscures what is genuinely new.

To understand this, consider the panopticon, the prison design where inmates cannot know whether they are being observed, and thus behave as if they are. This is precisely the dynamic of contemporary AI-enabled surveillance. But modern surveillance is not a neutral technology that happens to be misused — it is a practice with racialized origins, developed to track, control, and discipline populations coded as deviant. To deploy AI surveillance in the workplace without dealing with this history is less a technical oversight than it is a systemic ethical failure.

Three features distinguish AI-enabled surveillance from its predecessors:

• Continuous inference. Where legacy systems record, AI systems interpret, generating probabilistic assessments of mental states, intentions, and risk that may bear no relation to the qualities they purport to measure.
• Aggregation. Daniel Solove’s “taxonomy of privacy” identifies how individual data points become ethically significant in combination. AI dramatically accelerates this.
• The digital poorhouse. Algorithmic systems claim objectivity while functioning to penalize those who are already least powerful. Practically, that means employees with disabilities, carers, and those who speak non-dominant languages.

Contemporary dilemmas in practice

The productivity-surveillance trap

Activity-based productivity metrics have proliferated rapidly in hybrid and remote work environments. But research suggests they do not measure what they claim to measure. Research demonstrates that electronic performance monitoring is associated with reduced job performance and increased stress, particularly when employees experience it as surveillance rather than support.

The deeper problem is that measurement shapes behavior: what gets monitored gets gamed, reducing the quality of work. Deep work, creative synthesis, mentoring, and the forms of informal collaboration that sustain organizational culture are mostly invisible to AI.

Emotion AI and affective inference

Tools claiming to assess employee engagement, wellbeing, or stress through facial analysis and voice tone detection during video calls have become mainstream. The validity of these systems is questionable because the assumption of universal, reliably readable facial expressions of emotion simply does not exist. Os Keyes’ (2019) analysis of automated gender recognition systems shows how unrepresentative training data makes outputs not only inaccurate, but actively harmful. Deployed in a performance management context, emotion AI does not measure engagement neutrally. It measures departure from an arbitrary, unknowable norm.

Network and communication surveillance

AI analysis of communication metadata — who speaks to whom, response latency, information flow patterns — is now embedded in standard enterprise platforms. In this, organizations rely on a largely fictional consent framework. After all, consent obtained as a condition of employment, in conditions of economic dependency, does not constitute meaningful agreement. It constitutes compliance under duress.

The more serious risk is the chilling effect on legitimate organizational speech: dissent, whistleblowing, grievance-raising, and union activity are all forms of communication that network surveillance systems may flag as anomalous. The employee who knows their messages are being scored for “positivity” will not raise feel that their rights are exercizable.

Hybrid work and the colonization of private space

Remote monitoring technologies extend surveillance into employees’ homes. This is not minor. It dissolves a spatial and ethical boundary that employment law, professional ethics, and common dignity have long treated as consequential. Moya Bailey’s concept of misogynoir — the specific, compounded experience of gendered anti-Blackness — is instructive here: the home surveillance context falls hardest on those whose domestic labour, caregiving responsibilities, and residential circumstances are most likely to diverge from the implicit “acceptable” professional norm.

An ethical framework — five questions for adjudication

The Kantian tradition offers the foundational principle: persons must be treated as ends in themselves, never merely as means. But Kant’s formulation was designed for a world of individuals in relatively symmetrical moral relations, and — as Charles Mills argues powerfully in The Racial Contract (1997) — liberal moral frameworks have historically extended their protections selectively, along lines of race, gender, and class. The dignity surveillance threatens is not an abstract universal; it is embodied, socially situated, and unevenly distributed.

Nora Berenstain’s (2016) concept of epistemic exploitation (i.e., the extraction of knowledge, testimony, and labor from marginalized people without recognition or reciprocity) maps precisely onto inferential surveillance: employees’ behavior, communication, and even emotional states are harvested to generate organizational value, often without their knowledge or consent, in ways they cannot contest. This goes beyond privacy violation and into exploitation.

With this grounding, five questions provide for ethical adjudication before deploying any surveillance system:

1.  Purpose limitation.  Is this surveillance actually necessary for a legitimate, provable organizational aim?  Helen Nissenbaum (2010) offers that information flows are ethically appropriate when they match the context in which information was originally shared. Employee communications shared in a work context carry different ethical character than data harvested to build behavioral profiles.

2.  Proportionality.  Is the level of intrusion proportionate to the risk or validated need? Would a less invasive means achieve the same end? The answer is, more often than not, yes.

3.  Power and dignity.  Does this system treat employees as moral agents deserving of explanation and contestation or as inputs to be optimized? The question is whether employees can meaningfully understand and challenge the systems that affect their livelihoods.

4.  Disparate impact.  Who bears the costs of this system disproportionately, and is that distribution justifiable? Benjamin’s (2019) analytical frame demands that we ask not only whether a system works, but for whom it works, and at whose expense. A productivity monitoring system that functions adequately for most employees while systematically disadvantaging neurodiverse workers or primary caregivers, for example, is a tool of discrimination.

5.  Meaningful consent and transparency.  Do employees know what is collected, how it is used, and by whom, and do they have genuine recourse? Consent is not an employment contract clause. It is an ongoing, informed, freely given agreement.

These questions are interdependent. An organization that satisfies four and fails the disparate impact test has not done enough.

The compliance gap

Ethics, compliance, and risk professionals have sophisticated tools for legal risk. The EU AI Act now classifies emotion recognition and biometric monitoring systems as high-risk, requiring conformity assessments and human oversight. Australia’s Privacy Act reforms are advancing. General Data Protection Regulation (GDPR) frameworks govern data minimization and purpose limitation. These are meaningful floors, but are minimum standards, not the basis for ethical aspiration.

Legal permissibility and ethical permissibility are not the same thing. Leaders who conflate them expose their organizations to harms that do not appear in legal risk registers: eroded trust, reduced discretionary effort, chilled organizational speech, and the quiet exodus of exactly the people whose experience creates organizational value.

From what we have learned about psychological safety, we know that teams that feel observed and evaluated, rather than trusted and supported, consistently underperform on the complex, collaborative, and creative tasks that generate real value. Surveillance-heavy cultures do not produce productivity. They produce algorithmic performance theatre.

A final note on the business case framing: arguments for ethical restraint that rest primarily on ROI (e.g., surveillance is bad for productivity, therefore limit it) are strategically useful but ethically insufficient. They imply that if surveillance were demonstrably profitable, the ethical objections would dissolve. They would not. Dignity is not contingent on commercial return, and the question of whose dignity is treated as commercially relevant is one we are all accountable for addressing.

Practical commitments toward ethical deployment

For ethics, compliance, and risk professionals, the issue is not whether to engage with AI-enabled surveillance: these technologies are already shaping workplaces and generating harms. The issue is to engage with rigor and build meaningful ethical architecture for it.

Four commitments are achievable now:

• Build ethics review into procurement. Surveillance technologies should face the same scrutiny as any high-risk business decision before deployment, not after the first grievance.
• Create genuine employee voice mechanisms. Not notice-and-comment, but substantive participation in the design of monitoring regimes that affect working conditions. Workers subject to algorithmic assessment have a right to understand and contest it.
• Embed surveillance transparency in ESG disclosure. As ISSB and CSRD frameworks extend human capital reporting requirements, organisztions that cannot account for how they monitor their employees are carrying undisclosed risk.
• Build in sunset clauses and mandatory review. Surveillance systems approved in one context (i.e., pandemic emergency, etc.) have a documented tendency to persist long after their justification has expired. Governance structures that require affirmative reauthorisation rather than passive continuation are a safeguard against harm.

The organizations that will navigate this landscape with their reputations, their cultures, and their workforces intact are not those that have the most sophisticated monitoring capabilities. They are those that have asked seriously and repeatedly whether and how those capabilities should be used. They’ve also been honest enough to answer.

References

Bailey, M. (2021). Misogynoir Transformed: Black Women’s Digital Resistance. New York University Press.

Barrett, L. F., Adolphs, R., Marsella, S., Martinez, A. M., & Pollak, S. D. (2019). “Emotional expressions reconsidered: Challenges to inferring emotion from human facial movements.” Psychological Science in the Public Interest, 20(1), 1–68.

Benjamin, R. (2019). Race After Technology: Abolitionist Tools for the New Jim Code. Polity Press.

Berenstain, N. (2016). “Epistemic exploitation.” Ergo, 3(22), 569–590.

Bhave, D. (2013) “The Invisible Eye? Electronic Performance Monitoring and Employee Job Performance.” Personnel Psychology, 67(3), 605-635.

Ravid, D. M., Tomczak, D. L., White, J. C., & Behrend, T. S. (2020). “EPM 20/20: A review, framework, and research agenda for electronic performance monitoring.” Journal of Management, 46(1), 100–126.

Browne, S. (2015). Dark Matters: On the Surveillance of Blackness. Duke University Press.

Edmondson, A. C. (1999). “Psychological safety and learning behavior in work teams.” Administrative Science Quarterly, 44(2), 350–383.

Eubanks, V. (2018). Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. St. Martin’s Press.

Floridi, L. (2016). “On human dignity as a foundation for the right to privacy.” Philosophy & Technology, 29(4), 307–312.

Foucault, M. (1977). Discipline and Punish: The Birth of the Prison. Pantheon Books.

Keyes, O. (2018). “The misgendering machines: Trans/HCI implications of automatic gender recognition.” Proceedings of the ACM on Human-Computer Interaction, 2(CSCW), Article 88.

Mills, C. W. (1997). The Racial Contract. Cornell University Press.

Nissenbaum, H. (2010). Privacy in Context: Technology, Policy, and the Integrity of Social Life. Stanford University Press.

Noble, S. U. (2018). Algorithms of Oppression: How Search Engines Reinforce Racism. New York University Press.

O’Neil, C. (2016). Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. Crown Publishing Group.

Solove, D. J. (2006). “A taxonomy of privacy.” University of Pennsylvania Law Review, 154(3), 477–564.