They Were the Training Data

“The average intelligence of the people who are at this company is significantly higher than the average set of people that you can get to do tasks.”

That’s Mark Zuckerberg, in leaked audio from a Meta all-hands meeting, explaining why his company chose to surveil its own engineers rather than hire outside contractors to produce AI training data. Meta had been quietly tracking employee activity across Gmail, GChat, VSCode, and its internal AI assistant for months. The program was called the Model Capability Initiative. The next day, 8,000 of those employees received layoff notices.

The sequence matters. Track them. Extract value from their work product. Then eliminate the source.

The Asymmetry Nobody Talks About

When you use a platform, you produce data. Scroll patterns. Click timing. Reading duration. Emotional responses inferred from behavioral signals. Every platform collects this. Everybody knows this, vaguely, in the way you know your phone is always listening but choose not to think about it too hard.

What the Meta leak reveals is the math behind that collection.

Zuckerberg was explicit: he tracked employees because they were smarter than contractors. Their keystrokes, code patterns, and communication habits were worth more as AI training material than anything he could buy on the open market. He acknowledged the opacity was deliberate: “it is not strategically in your interest for us to communicate everything in all the detail that we normally would on this.”

Read that again. The head of one of the largest information companies on earth said, out loud, that transparency about data collection was against his strategic interest.

More than 1,000 Meta employees signed a petition demanding the program be halted. Fliers appeared on office walls. The petition arrived one day before the layoff emails.

Your News Feed Already Works Like This

Here is why this matters beyond Meta’s internal HR drama.

The AI systems that curate your news, select your recommendations, and decide what information reaches you every morning were trained the same way. Not on Meta employees specifically, but on behavioral extraction at scale. Your reading habits, your emotional triggers, your scroll-stop moments. All of it fed into models that learned one thing: what keeps you engaged longest.

The relationship between you and the AI that feeds you news is structurally identical to the relationship between Meta employees and the Model Capability Initiative. You produce behavioral data. The system extracts value from that data. The system’s optimization target is not your wellbeing. It is the company’s capability.

The difference is that Meta employees at least got a salary while they were being harvested. You get a news feed that’s been precision-engineered to activate your amygdala.

The Pattern Has a Simple Name

There is an old concept in economics called the principal-agent problem. It describes what happens when someone making decisions on your behalf has different incentives than you do. Your financial advisor recommends funds that pay him commission. Your real estate agent pushes the quick sale over the best price.

The AI version of this problem is worse because the asymmetry is invisible. A financial advisor at least has to disclose his fee structure. The AI that selects your morning news has no disclosure obligation. No transparency about what signals it uses. No accountability for the behavioral model it has built of you. And no regulation requiring any of it, after Musk and Zuckerberg personally convinced the President to shelve the only executive order that would have addressed this last week.

You are the training data. And the trainer answers to nobody.

What Opt-Out Actually Looks Like

Most news apps give you a choice that is not really a choice. Accept personalization (and the behavioral surveillance that enables it) or leave. There is no third option on the menu.

ntrl was built on a different premise: the AI analyzes the article, not the reader. No behavioral model. No engagement optimization. No profile of your emotional triggers being sold upstream to improve a model you never consented to train.

That is not a feature. It is an architectural decision. Once you build a behavioral profile, the incentive to monetize it is permanent. The only way to guarantee the relationship stays clean is to never collect the data in the first place.

The Leaked Audio Changed Nothing

A thousand employees signed a petition. The audio went public. The news cycle moved on. Meta’s stock barely moved.

The Meta leak is clarifying not because it will change Meta’s behavior, but because it said plainly what every data-extracting platform has calculated silently: you are worth more as a signal than as a customer. Your intelligence, your habits, your behavioral patterns are raw material. And raw material does not get consulted about how it is used.

The question is not whether your news app has made this same calculation. It has. The question is whether you have access to one that hasn’t.