I lead an ML research team, so I spend a fair amount of time on the building side of AI systems, not the being-evaluated-by-them side. That's exactly why the workplace AI conversation nags at me — I know how easily a model that looks clean on a validation set turns out to be encoding something ugly about the population it was trained on, and I know how rarely the people affected by a deployed model get to see how it actually scored them. In a research setting, a biased model produces a bad paper. In a hiring pipeline, it produces a bad decision about someone's income and career, made partly by an algorithm and largely without their knowledge.
Companies are already using AI to screen resumes, analyze interview footage, monitor productivity, predict who's about to quit, and shape promotion recommendations. Most employees have no real visibility into any of it. That's not a hypothetical governance gap — it's the current default, and it's worth being specific about where the actual risk sits rather than treating "AI at work" as one undifferentiated concern.
Not All Workplace AI Carries the Same Risk
It helps to separate low-stakes AI from high-stakes AI, because the ethical bar should scale with the consequences of getting it wrong. Meeting transcription, writing assistance, code review suggestions, and calendar scheduling are genuinely low-risk — nobody's livelihood turns on whether the AI summarized a meeting slightly wrong. Resume screening, interview analysis, performance prediction, and promotion recommendations sit at the other end entirely: these are decisions with real, sometimes irreversible consequences for a person's income and career, made by a system whose internal logic the person being scored usually can't see or contest. Productivity monitoring and flight-risk prediction land somewhere in between — not usually a single make-or-break decision, but a steady stream of inputs that quietly shape how a person is perceived by their manager.
The mistake I see in a lot of corporate AI policy is treating all of this as one category requiring one uniform response. It isn't. The scrutiny a company applies should track the actual stakes of the decision, not a blanket "we use AI responsibly" policy that reads the same whether it's covering a scheduling bot or a hiring algorithm.
Where Bias Actually Comes From
The bias problem in workplace AI isn't usually about someone deliberately building a discriminatory system. It's that these models are trained on historical data, and historical hiring and promotion data reflects decades of human bias baked into outcomes. A resume screener trained on a company's past hires will learn whatever patterns produced those hires — including patterns that had nothing to do with actual job performance. Facial and voice analysis tools built to assess "communication skills" or "cultural fit" during interviews have a documented tendency to penalize accents and non-Western communication styles, treating cultural variation as a deficiency because the training data was culturally narrow. Performance prediction and promotion models trained on a company's historical advancement patterns tend to reproduce whatever demographic skew already existed in who got promoted — the model isn't correcting the bias, it's automating it and giving it the appearance of statistical objectivity.
None of this makes AI uniquely villainous — a human hiring manager carries plenty of unconscious bias too. What's different with an algorithmic system is scale and invisibility: the same bias now applies consistently to every applicant, with no manager to occasionally deviate from pattern, and with none of the visible reasoning a candidate could push back on. A rejected candidate can sometimes ask a human recruiter what went wrong. They can't interrogate a model's internal weights.
Transparency Is the Actual Fault Line
Most of the concerning workplace AI use I've come across isn't concerning because the underlying capability is dangerous — it's concerning because it's invisible. Employees usually don't know that AI screened their application, don't know how a productivity score gets calculated, don't know that a model influenced their promotion, and don't know what data was collected to feed any of it. Each of those gaps has a direct practical cost: you can't contest a decision you don't know was automated, you can't improve toward a metric you don't understand, and you can't trust a process you can't see into.
This is also where I'd push back hardest on employers, because transparency here is cheap relative to the trust it buys. Telling employees when AI is used in a decision that affects them, explaining in plain language what a score is based on, and giving a real appeal path to an actual human doesn't cost much technically. What it costs is admitting that the AI system might be wrong sometimes — which is precisely the admission a lot of companies seem reluctant to make.
The Legal Ground Is Already Shifting
This isn't purely a matter of corporate conscience anymore. The EU AI Act classifies AI used in hiring and employee evaluation as "high-risk," which triggers mandatory transparency and human-oversight obligations rather than leaving it to each company's discretion. New York City's Local Law 144 already requires employers to commission an independent annual bias audit before using an automated tool to screen or promote candidates — a real, enforceable requirement, not a best-practice suggestion. Other US states have started adding their own disclosure requirements, and the overall direction across jurisdictions is consistent: toward more mandatory transparency and auditing, not less. A company treating AI ethics as a purely voluntary, PR-driven exercise is going to find the regulatory floor rising underneath that assumption.
Monitoring Deserves Its Own Line
Employee monitoring gets bundled into the same conversation as hiring AI, but I'd treat it separately, because the ethical test is different. For hiring and promotion decisions, the core question is fairness and contestability. For monitoring, the core question is proportionality: does the level of surveillance actually match a legitimate business need, or has it crept toward tracking things — keystrokes, screenshots, bathroom breaks, webcam feeds — that have nothing to do with the actual work getting done. Tracking work hours and project completion is reasonable in almost any context. Tracking every keystroke and screenshot rarely is, because the marginal information it captures beyond outcome-based measurement mostly consists of things that are none of an employer's business. If a monitoring practice would feel invasive explained out loud to the person being monitored, that's usually a good enough signal that it's crossed a line worth reconsidering.
What I'd Actually Ask an Employer to Do
If I were advising a company on this rather than just writing about it, I'd keep the list short and concrete rather than building an elaborate governance framework nobody follows: disclose which decisions involve AI and which don't, in plain language rather than buried in a policy document nobody reads. Run bias testing before deployment and repeat it periodically afterward, since a model that was fair at launch can drift as the underlying population or behavior shifts. Keep a human in the loop for any decision with real consequences — a model can inform a hiring decision, but it shouldn't be the final word with nobody accountable for overriding it when it's wrong. And build a genuine appeal path, not a suggestion box that nobody actually staffs.
The pitch that AI removes bias from decisions humans used to make badly is seductive, and it's not entirely wrong — a well-audited system can in principle be fairer than an inconsistent human reviewer. But "in principle" is doing a lot of work in that sentence. What I've actually seen more often is bias laundered through a system that looks objective precisely because it's a black box, which makes it harder to challenge than the human bias it replaced, not easier. The technology isn't the problem. The absence of scrutiny is.