When Workers Become Training Data

Simon Carver

Welcome to the show. This is The Human Workforce, and today’s quick take is simple and a bit unsettling: WHEN WORKERS BECOME TRAINING DATA, the real crisis is dignity. If this conversation hits you somewhere real, like, subscribe, and share it with someone thinking hard about work and AI. I’m Simon Carver, here with Lachlan Reed and our guest host, Jacques San Dimas. And Jacques, let me start with the question that kept me up: at what point does quietly doing your job become training the machine that may eventually do pieces of it too?

Jacques San Dimas

It becomes that moment when observation changes purpose. For many years, companies monitored work for security, compliance, fraud prevention, operational oversight. Reasonable things, mostly. But when the enterprise begins collecting clicks, keystrokes, shortcuts, navigation behavior, even periodic screen snapshots, not simply to supervise the present but to model the future... then something fundamental has shifted. The employee is no longer only performing labor. The employee’s behavior becomes raw material.

Lachlan Reed

Yeah, and that phrase, “raw material,” that’s the bit that sticks in my teeth. Because clicks and shortcuts sound boring, right? Mouse wobbles, keyboard taps, all that office wallpaper. But if a company’s studying those little moves to teach an AI agent how humans actually get through a system, then “just doing your job” isn’t just doing your job anymore. It’s more like leaving breadcrumbs and finding out later someone was mapping the whole bush track.

Simon Carver

And the Meta reporting is the sharp example here. Not sci-fi helmets, not mandatory GoPros, none of that. But Reuters described collection of mouse movements, clicks, keystrokes, shortcuts, navigation behavior, and periodic screen snapshots to improve AI agents that learn how people interact with digital systems. Jacques, that list matters to me because it’s so ordinary. A shortcut key is such a tiny human preference. And yet that tiny preference can teach a system competence, can’t it?

Jacques San Dimas

Yes. The small things are often the most revealing things. In a kitchen, you do not judge a chef only by the plated dish. You watch the hand placement, the timing, the sequence, the hesitation before a correction. Workflows are similar. A shortcut tells you familiarity. A pause before clicking “approve” tells you judgment. A screen snapshot may reveal context, exception handling, workaround behavior. The system is not merely counting outputs anymore. It is studying how a human thinks through the work.

Lachlan Reed

Wait, the “pause before clicking approve” bit — that’s the one I won’t forget. Because that hesitation might be caution. Or experience. Or someone thinking, “Hang on, this looks dodgy.” If software starts treating hesitation as a pattern to mine, that’s not measuring effort. That’s sniffing around inside judgment itself.

Simon Carver

Right. It’s one thing for a manager to know I answered ten emails. It’s another for a system to learn the rhythm of my corrections, the route I take through messy software, the exact moment I second-guess a decision. That feels less like performance management and more like... digital residue. Little flakes of me coming off while I work.

Jacques San Dimas

And this is where dignity enters the room. Because once the system learns from your behavior, it is not simply evaluating you. It is consuming patterns that came from your attention, your expertise, your caution, your accumulated mistakes. Organizations must be honest about that distinction. A thermometer measures temperature. It does not become the weather. But in these environments, the measurement itself is used to build the next operational actor.

Lachlan Reed

Most people signed up for a job description, not behavioral harvesting. That’s the plain-English version, isn’t it? If you joined a company in 2012 or 2018, you probably thought, “I’ll do the work, they’ll pay me, maybe I’ll grow, maybe I’ll get some stability.” You did NOT think, “Sweet as, my mouse movements might help train a system that reduces how many people do this later.”

Jacques San Dimas

Exactly. There is the formal contract, and then there is the psychological contract — the unwritten exchange of labor for compensation, expertise for opportunity, loyalty for some degree of continuity or advancement. In the generative AI era, a new exchange may be occurring without explicit acknowledgment: tacit knowledge extraction, workflow digitization, behavioral replication. That is a serious ethical problem because the assumptions changed, but the consent often did not.

Simon Carver

Let me try to say that back, and you tell me if I’ve got it wrong. The issue isn’t only that monitoring exists. Offices have had monitoring forever — time clocks, call times, warehouse scanners, all that. The newer issue is that the purpose has changed. My experience, habits, corrections, and judgment patterns may now be used to build future machine intelligence. So the old “this is for safety or compliance” explanation no longer covers the full reality.

Jacques San Dimas

Almost. I would sharpen one word: not “may” alone, but “may persist.” Behavioral data can be durable. Reusable. Repurposed. That permanence changes the moral weight. Every workaround, every escalation path, every unusual exception you handle can become reusable intelligence. That is why employees feel the ground shifting beneath them.

Lachlan Reed

“Reusable” — yeah, that’s the kicker. If I teach a new teammate how to do something, that’s normal human work. If I accidentally teach an invisible system forever, with no say in where it goes next, that’s a different animal. Like, one’s mentoring. The other’s extraction wearing a helpful little lanyard.

Simon Carver

“Extraction wearing a helpful little lanyard” is painfully good. And we’ve seen signs of this broader trend, not just one company. Reporting has mentioned firms like JPMorgan Chase tracking AI adoption patterns, KPMG measuring AI utilization metrics, and more enterprise interest in workflow telemetry and behavioral analytics. Which, to be fair, does not automatically equal abuse. But it does tell us the modern workplace is instrumenting people at a finer grain.

Lachlan Reed

And that’s where I reckon the distinction matters: augmentation versus extraction. If AI removes repetitive admin, helps draft the boring first pass, reduces burnout, catches safety issues — beauty. That’s a mate helping carry the heavy box. But if the whole game is absorb human expertise, compress labor costs, and reduce dependency on actual humans with judgment... well, that’s not a teammate. That’s a vacuum cleaner with a quarterly earnings target.

Jacques San Dimas

A very expensive vacuum cleaner. And incentives matter. The technology is not inherently moral or immoral. Governance, disclosure, and purpose decide whether it elevates workers or quietly devalues them. If the system assists the nurse, analyst, claims reviewer, or customer support worker, that can be deeply humane. If it quietly harvests their methods in order to thin out the very people who made the system useful, trust will collapse.

Simon Carver

I keep thinking about a period years ago — not AI, just ordinary metric-heavy work — where every response time was visible, every delay had a color code, every pause needed explanation. And what it did to me wasn’t motivation. It made me narrower. Less generous. Less willing to think out loud. I didn’t bring my best judgment; I brought the version least likely to trigger a dashboard. And if that’s what simple metrics can do, imagine what happens when people suspect their judgment itself is being repurposed behind the scenes.

Lachlan Reed

The “least likely to trigger a dashboard” line... yeah. That’s not trust, that’s survival behavior. Once workers start performing for telemetry instead of reality, the company’s not even getting the good data anymore. It’s getting the scared data.

Jacques San Dimas

So what do organizations owe workers now? First, transparency — clear disclosure about what is being collected, why, for how long, and whether it will be used to train or improve AI systems. Second, boundaries — some categories of human behavior should not become collectible operational fuel simply because the technology allows it. Third, governance — documented rules, oversight, accountability, and consequences for misuse. And fourth, explicit protections: workers should understand their rights, the limits of secondary use, and the economic purpose attached to their data.

Simon Carver

The “economic purpose” part is so important. Because if a company says, “We collect screen snapshots to improve tools,” that sounds almost harmless. But if the real use is building systems that replicate exception handling, approval patterns, or workflow judgment, then the stakes are entirely different. Same snapshot, different moral universe.

Lachlan Reed

Also, more data doesn’t automatically mean better decisions. Every mechanic knows this, every rider knows this — if your gauge is dodgy, staring at it harder doesn’t fix the engine. And in workplaces, if people don’t trust the purpose, they change their behavior. They stop experimenting. They stop flagging edge cases. They stop being candid. So the company ends up measuring a weird, flattened version of work and calls it truth.

Jacques San Dimas

Yes. Trust and dignity are not soft ideals floating above operations. They are operational assets. When people feel respected, they disclose risk earlier, collaborate more honestly, and use better judgment under pressure. When they feel reduced to telemetry sources, they become guarded. Fear contaminates systems the way smoke contaminates a kitchen — quietly at first, then all at once.

Simon Carver

And maybe that’s the larger stake for the future of work. Not simply how intelligent AI becomes, but whether organizations remember what kind of thing a worker is. Not a productivity vector. Not a bundle of hesitation patterns. Not a mine to strip for behavioral ore. A person. A person whose humanity is not an obstacle to efficiency, but the source of the very judgment these systems are trying to learn from.

Lachlan Reed

And once people stop feeling trusted, they stop bringing that humanity in the first place. The shortcuts, the careful pause before “approve,” the weird little workaround that saves the day — all that good stuff dries up. You can’t harvest your way to a healthy culture. Bit of a mess, really.

Simon Carver

Jacques, Lachlan, thank you both. And thanks to you for listening. If you enjoyed this episode of The Human Workforce, subscribe and share it with someone who cares about where work is headed — because the future may be built with AI, but it will still be judged by how we treat people.