For workers subject to employee monitoring software in the age of AI, I imagine it must be like a pesky little brother you don’t want hanging around you all the time. Not only does it want to know what you’re up to when you’d rather it not, it’s now trying to learn from you. The thought of being a role model could be flattering, but you have every reason to expect your little brother will tattle on you at the slightest provocation, and that it may be trying to usurp your role as the kid good at that thing you do. Worse still, the little brother’s agenda changes: Tips shared in confidence one day may be used against you the next.
In April, Meta installed software on employee computers that monitored their apps, keystrokes, and mouse movements. The company had said it needed the data to train its AI models.
I’m honestly not sure if employees were more upset about potentially training their own replacements or about the indignity of being spied on. But they were not happy. In May, amid massive layoffs, they sued their employer. A primary complaint was that the company had engaged AI in the decision-making process, informed by the data collected by the monitoring tools about their work habits. Affected employees who had been away on family or medical leave believed they had been misrepresented by this data, and effectively discriminated against.
It will be for the courts to weigh the validity of the accusation. But the moral is clear: Even if your employer says they need your data to train AI (because you’re such an expert!!), and means it at the time, they might later use that data against you in a performance evaluation. Little brother at work!
Do you think your employer isn’t watching you? Don’t be too sure. In a Wall Street Journal column this week with tips about how to fool the monitors, Callum Borchers warns that many states don’t have laws requiring employers to tell you what they’re running on your work computer.
(Borchers’s tips, if you need them: Try to look only about 80% busy, because higher is suspicious; mark planned meetings on your calendar, lest you appear idle during those times; and prefer a physical mouse “jiggler” device over a software tool, because some monitors track the other apps you have running.)
Even if your employer actually isn’t tracking your work or your conversations on internal messaging servers, that doesn’t mean a little brother won’t be looking over them later. Case in point: Google paid $10 million at auction this week for the entire dataset of Spirit Airlines, which went bankrupt in May and ceased operations. Per CNN, this includes “emails and internal communications, spreadsheets, transactions with the public including bookings and frequent flyer information as well as human resources information on its employees.”
Google said only that the data “can be helpful in improving our products and AI models.” How so, exactly? The hosts of the Hard Fork podcast speculated yesterday that Google might not want this so much for standard text-prediction training data, but rather to create a simulated airline environment in which models might learn business skills. Simulated task environments are a booming sub-industry within AI right now.
The purchased data is supposed to have been stripped of individually identifiable information, but I wouldn’t bet on that protection holding against even present AI, which is already very good at making correlations to unmask the anonymous, let alone future AI. While I don’t think Google is interested in assessing the performance of former Spirit employees, let alone unmasking them, the data might become part of models that potentially could, and they may have incentive to do so: A model that learns more real-world context about the characters in its simulation might more masterfully navigate it.
Little brothers grow up. A potential scenario I don’t think I’ve seen even in sci-fi: a company’s AI turning you down for a job based on vibes about you from a simulation it only vaguely remembers from its childhood.
The analyses and opinions expressed on AI StopWatch reflect the views of the individual contributors and the sources they cover, and should not be taken as official positions of the Machine Intelligence Research Institute.