Your employer is likely watching you, scoring it with AI, and you can't do much

At a company I used to work for, the CFO's office rolled out a tool to get people to submit their timesheets. All it did was nag. Once your timesheet was past due, it interrupted you wherever you were, including in the middle of a client presentation. It was pushed remotely to every computer in the office. It was announced to staff, but nobody could opt out.

I happened to have admin access to my computer, and I knew how the thing was installed. So I removed it without telling anyone. Most of my colleagues didn't have that option.

That tool had one quality today's bossware has lost: you could see it.

The move from nag-ware to Big Brother

The nag tool was loud, local and visible. Anyone with access and know-how could find it and kill it. AI bossware runs the other way. It watches quietly and scores what it sees.

ExpressVPN, which sells privacy tools, commissioned a survey of 1,500 US employers and published it in February 2025: 74% said they use online monitoring tools and 61% said they use AI to evaluate employee performance. Prodoscore, which sells employers an AI productivity score built from workers' activity in their connected tools, says in its FAQ that "exact weights are not published to prevent gaming the system." You can't switch off a score, or fully argue with one whose weights you can't see.

Then the watching started feeding something else. In April 2026, TechCrunch reported, Meta began capturing employees' mouse movements, clicks and keystrokes on certain applications to train AI agents, because, in a Meta spokesperson's words, "our models need real examples of how people" use computers.

HR Brew reported in June that the program, mandatory for most staff, was paused after it left "transcriptions, employee conversations, and performance data accessible inside the company." More Perfect Union's Bossware campaign says some companies are using monitoring data to train workers' AI replacements. Meta is the documented case of staff activity training AI agents.

The nag tool interrupted you. This is quieter, and I think worse: it watches how you work and studies it.

Now the chat window reports back too. On 28 October Anthropic made smart reports generally available for Claude Enterprise. Claude reads a sample of a team's transcripts, sorts the work into categories and shows what each one costs, where people got stuck and which routines are worth turning into shared skills. Click a category and a manager can read a one-paragraph summary of each session, most expensive first. Smart reports are off by default, names stay hidden unless the company turns on an option to reveal them, and questions about union activity or health are blocked.

Anthropic says reports "should not be used for evaluating individual performance." I couldn't find anything in its help page on telling the people in the transcripts.

One-way transparency spends trust

The philosopher C. Thi Nguyen puts the trade in five words: "Transparency trades trust for monitoring." Trust means letting someone act out of your sight. Monitoring takes that back.

Nguyen wrote about experts answering to a public, and his phrase fits the office too: it puts people "under the gaze of a distant, insensitive, contextless eye." A timesheet alarm going off in front of a client is that eye in miniature. It knew the timesheet was late and nothing else.

Onora O'Neill made a related point in her 2002 Reith Lectures: people who know everything they say will be made public "may massage the truth." That's what the meter gets. Among the US workers in the same ExpressVPN survey, nearly 24% admitted faking productivity.

Ana Alacovska and colleagues studied gig-platform freelancers under random screenshots and click tracking. Those who saw the system, from its surveillance to client appraisals, as non-transparent, arbitrary and retaliatory came to mistrust it and the people around it, felt persecuted, and coped by going quiet. One poster in the forums they studied put it plainly: "This is why I stay off their radar." No forum posts, no support tickets, no reporting problems. Those are the problems a company needs to hear about.

It doesn't deliver what it promises, either. A 2023 meta-analysis in Personnel Psychology, pooling ninety-four samples, found no evidence that electronic monitoring improves performance. It did find more stress, whatever form the monitoring took.

So trust gets hit from both sides. The worker is fully exposed, which is Nguyen's trade. The system stays hidden, which leaves the worker nothing to go on but suspicion. My bet is that suspicion does more damage than transparency would, because people left to guess tend to assume the worst.

The studies that measure trust directly show the cost of monitoring, within limits. In a survey of five hundred Australian employees, monitoring went with lower trust in management, though only among manual workers. In a field experiment, advance notice of internet monitoring, together with how supported employees already felt, affected how far they trusted the organization afterward, and that trust fed their satisfaction, commitment and thoughts of leaving. Gallup reported in September that 20% of employees who work on a device have "a great deal" of trust in their leaders to use employee data responsibly, 43% trust them somewhat and 37% have little or none. It didn't compare monitored workers with others.

It reaches freelancers, contractors and even clients

Bossware doesn't stop at the payroll. Contractors, freelancers and partner staff work inside clients' Microsoft 365 every day. By default, Microsoft's documentation says, a guest accepts "the use of their information in accordance with the inviting organization's privacy policies." That page is about sharing profile details, not monitoring.

Monitoring aimed at employees catches outsiders too: Microsoft's Purview tool can be set to scan messages sent to monitored staff from "external and internal senders." The partner gets watched because they wrote to someone who was.

Clients can get the same treatment. That Purview setting reaches outside senders, so a customer who emails a monitored account manager can be scanned too. I didn't find anything in Microsoft's page about telling them.

The few US laws that require notice of monitoring are written for employees. Connecticut defines an employee by whether the employer controls "the details and means" of the work. Bossware watches exactly the details and means a company says it gave up when it hired a contractor. In a 2024 UC Law Journal note, Rafi Bortnick notes that many workers labeled contractors "face the same or greater exposure to electronic monitoring" as employees, and the National Labor Relations Act protects only employees.

A contractor agreement doesn't shut every door, though. Amazon Flex drivers, who deliver packages under Amazon's "Independent Contractor Terms of Service," alleged that the company secretly monitored their off-hours talk in closed Facebook groups. In 2023 the Ninth Circuit refused to push the case into arbitration: "The alleged misconduct would be wrongful even if there had been no contract." The court ruled only on arbitration, and the spying was an allegation, not a finding.

Where you work decides your rights

In the United States the default sits with the employer. Federal law, the Electronic Communications Privacy Act, lets employers monitor company devices through consent and business-use exceptions, as the law firm Bricker Graydon explains.

Beyond what one appeals court has read into that exception, notice is a state-by-state requirement. New York requires written, acknowledged notice on hire. Connecticut has required notice of the kinds of monitoring since 1998, and since 1 October 2026 employers must also say where on their premises it happens, though one with reasonable grounds to believe someone is breaking the law, violating legal rights or creating a hostile workplace can still monitor without notice.

The EU runs the other way. Its data-protection regulators said in 2017 that employees "must be informed of the existence of any monitoring" and its purposes, and that workers are "seldom in a position to freely give, refuse or revoke" consent. The AI Act has banned AI that infers emotions at work, outside medical or safety uses, since 2 February 2025.

California moved on 30 September 2026: from January, AB 1883 bans AI surveillance that infers workers' emotional states or collects neural data, with exceptions for safety and some federally required work, and from July 2027, SB 947 will bar employers from relying solely on an algorithm to discipline or fire.

Make the system answer first

O'Neill was writing about the press, but her fix fits here. Trust needs "assessable reasons for trusting and mistrusting," and at work the system is what should be assessable: what it measures, why, who sees it and how you contest it.

California's SB 947 goes part of the way. From 1 July 2027, a worker disciplined or fired mainly on an algorithm's say-so will be able to ask for "a meaningful, objective description" of their own data the system used. The meta-analysis points the same way: where organizations were open about monitoring, workers reported more fairness, commitment, perceived support and satisfaction, though the pooled studies showed no effect on performance. A 2025 working paper on 434 Upwork data-entry workers found a cost to poor explanation: surveillance made no significant difference to output on average, but output fell about 17% in the groups given no performance-based reason for keeping or removing it.

I could remove the nag tool because I could see it and knew how it worked. In ExpressVPN's survey, half of US workers suspect they're being monitored without their knowledge, even though 86% of employers say they disclose it. A notice can tell you the watching exists and still say nothing about what it does with you.

In Gallup's survey, only 19% of employees said anyone at work had discussed the data collected about them. Those who had were likelier to trust leaders a great deal with it, 26% to 18%, and likelier to worry about it too. That's the most I'd expect an honest system to buy: a worker who can judge it.

Make the system visible to the worker before you make the worker visible to the system.

Disclosure: I use Anthropic's Claude in my work, including to research and edit this article. Anthropic had no role in it. Here's how I use AI.