On the morning of Friday, June 12, 2026, two things happened at Meta, and only one of them was on purpose. Starting around half past nine Eastern, Facebook, Instagram, and Messenger fell over. Users were thrown out of their sessions and met with login loops and blank white screens. Downdetector logged something in the neighborhood of a hundred and ten thousand reports for Facebook alone, then briefly buckled under its own traffic and started serving error pages to the people trying to find out whether the error was theirs. The outage ran for roughly four hours before the last of the services limped back. That was the part nobody scheduled.
The part that was scheduled was the memo. Sometime that same day, Mark Zuckerberg sent a note to employees, later reviewed by Reuters, conceding what anyone who works at his company already knew. “Given the complexity of these changes, we’ve made mistakes and will almost certainly make more,” he wrote, adding that he was “focused on providing as much stability as possible” and did not expect further company-wide layoffs this year. He also said he did not want to over-promise, “because the world is changing in ways that are out of our control.”
Read those two events next to each other and you have the whole man in a single news cycle. The chief executive who cannot keep three apps online for a Friday morning is the same one issuing a memo about stability. The phrase he reaches for to explain the wreckage is “out of our control,” as if a force of nature had visited the layoffs and the surveillance and the revolt upon a passive company rather than a person making decisions. This is not candor. It is the most familiar product Meta ships, the one it has shipped for two decades: the apology that changes nothing, timed to be the story so that the conduct is not.
Zuckerberg Cried Wolf One Too Many Times
To understand why “we’ve made mistakes” should be read as a press release rather than a reckoning, you have to remember how recently the previous press release ran. In January 2025, Zuckerberg posted a video titled, without apparent irony, “More speech and fewer mistakes.” In it he announced he was dismantling Meta’s third-party fact-checking program, the most extensive operation of its kind ever built by a platform, on the grounds that the fact-checkers had become too politically biased and that moderation had curdled into censorship. He promised to move trust-and-safety teams from California to Texas to build trust in places with, as he put it, less concern about the bias of his teams. He pledged to work with the incoming administration to push back on governments pressuring American companies.
Set aside, for a moment, what you think of the politics of that decision. Hold onto the structure of it. Here is a man who built the machinery that decided what was true on platforms used by billions, ran it for years, and then, when the political weather turned, tore it out and recast its entire existence as a mistake he was now nobly correcting. And he did it months after admitting, in an August 2024 letter to Congress, that the company had in fact bowed to government pressure to suppress certain content during the pandemic, content he now said he regretted taking down. Conservatives had spent years arguing the apparatus throttled right-leaning speech and buried inconvenient stories. In killing it, Zuckerberg did not exactly dispute them; he conceded the over-reach and monetized the concession as a free-speech conversion narrative.
So the pattern is not new in June 2026. It is the only pattern. First you build the thing. Then you run the thing well past the point of public objection. Then, when keeping it costs more than confessing it, you announce that you have learned, you have grown, you are providing stability and restoring expression, and you would appreciate the benefit of the doubt. The confession is not the end of the behavior. It is a feature of the business model, the periodic ritual that buys another eighteen months before the next one. “More speech and fewer mistakes” in winter; “we’ve made mistakes” by summer. The words rotate. The man does not.
What “stability” looks like from inside the building
It is worth dwelling on the word stability, because the people receiving the memo know exactly what it is meant to paper over. In May, Meta cut roughly eight thousand jobs, about ten percent of its global workforce, and reassigned around seven thousand more employees into artificial-intelligence work many of them did not ask for. Put those numbers together and nearly one in five people at the company either lost their job or woke up doing something unrecognizable. Combined with the rounds of 2022 and 2023, Meta has now eliminated something on the order of thirty thousand positions in roughly four years, a fact that makes the promise of “stability” read less like a commitment and more like a punchline delivered to the survivors.
And the survivors are not quiet about it. At the center of the storm is a three-month-old unit called Applied AI Engineering, roughly sixty-five hundred engineers and product managers assembled, in large part, from the ashes of the metaverse dream that consumed something like eighty-three billion dollars before the company pivoted. Employees inside it have a name for themselves. They call themselves draftees. They have a name for the unit, too. “It’s literally the gulag,” one of them told Wired. Another said the work, which consists largely of hand-generating puzzles and coding problems to feed Meta’s models, is soul-crushing. At its inception the management structure was so flat that as many as fifty people reported to a single manager, which is not a structure so much as a queue. An internal company livestream was reportedly hijacked. The chief product officer, Chris Cox, told staff on a call this week that the last few months had been brutal and urged leadership to get back in touch with the company, which is a remarkable thing for the chief product officer to feel he needs to say out loud.
Against all of that, the memo offers its remedies, and the remedies tell you everything. There will be bigger budgets for team offsites. There will be a company-wide hackathon in July. Assigned desks will return to the offices before the year is out. This is the corporate equivalent of a man who has set the kitchen on fire promising the family a nicer set of plates. If you are an engineer who has just watched a fifth of your colleagues get cut or conscripted, who has been told your new job is to manufacture training data for the systems meant to make your old job unnecessary, a hackathon and a return of assigned seating is not stability. It is a distraction with a calendar invite attached.
The Employee Data Extraction Factory
The single ugliest detail in the entire affair is the one Zuckerberg’s memo does not dwell on, and it is the detail that most cleanly exposes how he actually thinks about the people who work for him and the people who use his products, who turn out to be governed by the same logic. In April, Reuters first reported on an internal effort called the Model Capability Initiative. It installs tracking software on the work laptops of Meta’s United States employees. According to internal memos viewed by CNBC, that software captures mouse movements, keystrokes, clicks, menu and dropdown navigation, and periodic snapshots of what is on the screen, across hundreds of applications and websites that include Google Workspace, GitHub, Slack, and LinkedIn. The stated purpose is to gather authentic data on how skilled humans operate a computer, so that Meta’s AI agents can learn to do the same. At an all-hands, in remarks captured on leaked audio, Zuckerberg made the pitch in his own words: the way the models get better, he said, is by “watching really smart people do things.”
Sit with the proposition. The plan is to record your most talented engineers in granular detail, keystroke by keystroke, in order to train the very systems intended to automate them, and to do it, at least at the outset, with no opt-out for anyone on a company-issued device, a condition the chief technology officer reportedly confirmed in writing. The workers understood the arrangement perfectly, because unlike the average Facebook user, they can read the code and they can read the org chart. They taped flyers to the walls of the California and New York offices, on vending machines and, the reporting noted, atop toilet-paper dispensers. “Don’t want to work at the Employee Data Extraction Factory?” the flyers asked, citing the National Labor Relations Act and pointing colleagues to a petition. The petition, which carried around a thousand signatures in May, has now passed sixteen hundred. In the United Kingdom, staff began organizing a union.
Only after that pressure did the company blink. On the second of June, a vice president in the superintelligence group circulated a memo softening the program: employees could now pause data collection in thirty-minute increments, and a narrow set of workers could opt out entirely, specifically remote employees with limited bandwidth, those who routinely handle sensitive information, and those who cannot reliably keep a laptop on a charger. The team also shipped performance fixes, because, it emerged, the surveillance software had been so heavy it was degrading the machines it was spying on. That is the tell within the tell. The monitoring was not only invasive; it was badly engineered, taxing the very productivity it claimed to be studying.
Here is why this matters beyond Menlo Park. The logic Meta just turned on its own staff is the same logic it has run on the rest of us for twenty years. Surrender your behavioral data. Trust that it is being used for your benefit and not against your interests. Accept that there is no real opt-out, only the theater of a setting buried three menus deep. The difference this time is the subject. When the data being harvested belonged to two billion users who could not see the mechanism, the company called it personalization and the objections never reached critical mass. When the identical mechanism was pointed at the engineers who can see exactly what it is and exactly what it is for, sixteen hundred of them signed their names to a document invoking federal labor law. The people closest to the machine took one look and refused. That should tell the rest of us something about the machine.
Reliability is downstream of the thing he broke
Return now to the outage, because as a practitioner I do not think it was unrelated to everything above, and I do not think it was bad luck. When multiple flagship services fail at once, the cause is almost always shared backend infrastructure, the deep plumbing that everything else sits on. Meta declined to specify what broke. But you do not need the post-mortem to understand the conditions that make a Friday-morning collapse more likely, because those conditions are the subject of the memo itself.
You cut eight thousand people. You force seven thousand more out of the systems they understood and into a data-labeling unit they describe as a gulag. You demoralize the engineers who remain, surveil their keystrokes, and watch sixteen hundred of them sign a petition and start a union. You let institutional knowledge walk out the door in successive waves over four years. And then you act surprised when the lights flicker. Reliability is not a feature you can buy back with a hackathon. It is downstream of retention, of morale, of the unglamorous tenure of the people who know where the bodies are buried in a codebase that has been accreting for two decades. You cannot gut that organization and expect the backend to hold, and on June 12 it did not hold. The outage was not an interruption of the story Zuckerberg was telling that day. It was the most honest sentence in it.
The record the apology is meant to make you forget
An apology works by narrowing your field of view to the present tense. We made mistakes, we are fixing them, let us move forward. So it is worth widening the lens, because this publication has spent a good deal of ink on the parts of Meta’s business that no internal memo will ever concede, and the present confession sits on top of a long and well-documented pattern of telling people numbers that are not true.
Consider the company’s relationship with its own metrics, which is not a matter of opinion but of court record. In the long-running litigation over the “potential reach” figure that Meta showed to advertisers, plaintiffs alleged the number was inflated by as much as four hundred percent, that it counted accounts rather than people, that it knowingly included duplicates and fakes, and that in some American markets the supposed reach exceeded the actual population. Unsealed internal documents quote a Facebook employee describing revenue the company should never have made because it rested on wrong data. The chief operating officer, by the suit’s account, had been aware of the problem since 2017. A product manager proposed a fix that would have made the metric honest; managers rejected it because the revenue impact would be, in the company’s own word, significant. The exposure in that case has been pegged north of seven billion dollars, and the Supreme Court declined to hear Meta’s attempt to escape it. This is not ancient history; a federal judge ruled as recently as late 2025 that the company had forfeited its bid to force the dispute into private arbitration by sitting on the right for seven years.
That episode was not the first. A decade ago Meta conceded it had overstated the average time users spent watching video, the metric that publishers and advertisers used to reorganize their entire strategies around so-called pivots to video. The company settled a related advertiser suit for forty million dollars, and the plaintiffs, having reviewed the internal records, argued the inflation of those view times was far from an honest mistake. So when the same company now asks you to accept its account of its own candor, you are being asked to trust the numbers of an enterprise with a documented, evidence-backed habit of lying about numbers whenever the truth would have cost it money. The memo about mistakes is itself a metric. Treat it the way the courts have taught us to treat the others.
Then there is the war on the open web, which we have covered as it has unfolded. The mechanism is simple and the same everywhere AI summarization has touched: the model answers the question on the surface that used to send a reader to the source, and the publisher is left with the attribution and none of the traffic. The damage is no longer speculative. By the measurements that came out this spring, small publishers have lost on the order of sixty percent of their search referral traffic over two years; zero-click rates have climbed past two-thirds; click-through nearly halves when an AI summary sits at the top of the page. Meta’s contribution to this is of a piece with its character: an AI layer that intercepts the click, alongside years of quietly throttling news and political content in the feed, a suppression reversed only at the precise moment it became politically convenient to court the right. The phrase one industry report used cuts to the bone. Attribution does not pay the bills. A summary that satisfies the reader and starves the writer is not a service to anyone but the platform that keeps the audience captive inside its own walls.
And captive is the operative word, because the captivity is engineered into the smallest details. We have written before about the walled garden, about an advertising apparatus tuned to keep you inside the app rather than send you anywhere useful, about links that open in Meta’s own in-app browser instead of your real one, about ad performance that underwhelms while the dashboards above it glow with the very kind of inflated figures the courts have been litigating for years. None of this is incidental. It is the consistent expression of a single worldview, the one the surveillance memo made explicit: your attention and your data exist to be extracted, the open exit is a bug to be closed, and the public-facing language about connection and community is the wrapping, not the contents.
And then there is the matter of the button. In the early years of the last decade, Facebook handed websites a small piece of code, the Like button, and the web took it up everywhere, on news sites and shopping carts and personal blogs, because a little blue thumb promised engagement and reach. What the sites installing it did not advertise, and what their visitors had no way to grasp, was that the button was a beacon. Your browser, loading any page that carried it, quietly reported back to Facebook that you had been there, who you were, and when — whether or not you ever clicked the thing, whether or not you were logged in, and, most chillingly, whether or not you had ever opened a Facebook account in your life. This was before browsers blocked trackers by default, before the phrase third-party cookie meant anything to anyone outside the industry. For years, across a vast stretch of the open web, Facebook was silently assembling the browsing history of much of humanity, including the people who had expressly refused to join it.
This is not a dark theory; it is a matter of record. When the Belgian data-protection authority commissioned researchers from the University of Leuven and the Vrije Universiteit Brussels to examine the practice, their report concluded that Facebook occupied a position no company in history had held, able to link a person’s browsing behavior to their real-world identity, their social connections, their offline purchases, and what the experts listed in plain words as their medical information, their religion, and their sexual and political preferences. A Brussels court ruled the tracking of non-users illegal in 2015 and ordered it stopped on pain of a quarter of a million euros a day; after the predictable jurisdictional appeals, a court ruled against the company again in 2018, ordering it to halt and to destroy the data it had gathered without consent. And lest this be waved away as ancient history, the architecture merely changed its name. The Like button became the Meta Pixel, and the pixel is still out there on millions of sites, caught by investigators funneling the patient data of a third of America’s top hospitals back to Meta, funneling the financial details people typed into online tax-filing services back to Meta, until a federal magistrate judge ordered Mark Zuckerberg himself to sit for a deposition about it in 2025. This was never a bug. It is, and always was, the load-bearing wall.
Now hold still and consider what an archive like that actually contains, because this is the part that should put a cold weight in your chest, more than any layoff or outage. It is not merely what you chose to post. It is where you went when you believed no one was watching: the symptom searched at two in the morning, the lawyer looked up, the diagnosis, the debt, the thing about yourself you have told no one. Multiply one person’s file by a decade and two billion people, and then ask the only question that matters. Who is in that dataset? The senator is in it. The judge is in it. The regulator who is meant to police him, the reporter who is meant to expose him, the rival across the negotiating table — and you, reading this, are in it too. I want to be exact here, because the exact version is the frightening one. I am not telling you that Mark Zuckerberg has reached into that vault and used what he found there against anyone; I have no evidence that he has, and neither does anyone else, and that absence is precisely the problem. Leverage does not have to be used to be real. It only has to exist, to be held by one man, and to be impossible for the rest of us to inspect. You cannot see your own page in that file. You cannot learn what it says. You cannot compel him to delete it, and you cannot know who else — by subpoena, by breach, or by some quiet arrangement — has been permitted to read it. A man sitting on that much of everyone’s hidden life, answerable to no court that can move him and no regulator he has not already outlasted, is the one now asking you for the benefit of the doubt. Sit with who is actually asking, and with what he is holding while he asks.
The fact-check that can never happen
Which brings us back to where we started, and to the cleanest irony of the entire week. The instinct, watching a powerful man issue a self-serving statement, is to want it fact-checked, to want some independent authority to weigh it against the record and stamp it for what it is. But here is the thing about Mark Zuckerberg’s memo that makes it different from an ordinary corporate apology. There is no independent fact-checker left to rate it, because he fired them.
He is the man who built the most elaborate fact-checking partnership in the history of social media, and he is the man who, eighteen months ago, tore it out by the roots and rebranded its abolition as a victory for free expression. The successor he installed, a community-notes system modeled on the one over at X, will never be allowed near an internal memo to staff in any case; it adjudicates posts, not the founder’s pronouncements. So the most powerful arbiter of truth the internet has ever known has arranged things such that his own most consequential claims float beyond the reach of any check at all. He decided what was true for billions of people for the better part of a decade. Then, when the machinery of truth-rating became inconvenient to his politics and his costs, he abolished it and called it liberation. And now he issues a statement that, by his own design, no one is positioned to verify.
That leaves only one honest audit available, and it is the one I have been conducting for this entire piece: the conduct. You do not need a fact-checker’s verdict when you have the outage, the gulag, the keystroke surveillance and the sixteen hundred signatures, the seven-billion-dollar metrics case and the unsealed emails, the starved publishers and the walled garden. Measured against that record, “we’ve made mistakes and will almost certainly make more” is not an admission of humility. Parse the second clause. He is telling you, in plain language, that he intends to keep doing it, that more mistakes are coming, that this is simply the nature of the world he insists is out of his control. And he is asking you to receive that promise of repetition as if it were contrition.
The world is not doing this to Meta
The tell, the single phrase to refuse, is “out of our control.” The world is not doing this to Meta. The layoffs were a decision. The conscription of seven thousand engineers into a unit they call a gulag was a decision. The keystroke surveillance with no opt-out was a decision, and so was the grudging partial retreat once the workforce organized against it. The dismantling of fact-checking, the suppression of news and its convenient reversal, the metrics the company defended in court for the better part of a decade rather than correct, the architecture built to trap attention and intercept the click — all of it, every line, was chosen. A man does not get to make ten thousand deliberate choices and then describe the sum of them as weather.
So no, I will not extend the benefit of the doubt, and I would suggest you withhold it too. The remedy for an executive who keeps making the same mistake, who has built a career on the rhythm of confess-and-continue, is not another beautifully timed memo about how much he has learned. It is for him to stop. Until he does, treat the apology for what the record proves it to be — not the moment he changed, but the cost of doing business, paid in words, so that everything underneath can go on exactly as before.