Sophie Zhang Found Governments Faking Public Opinion. Nobody Wanted to Know.

In September 2020, on her last day at Facebook, a data scientist named Sophie Zhang posted a memo of nearly eight thousand words to the company’s internal forum. It described something she said she had been chasing for most of two years: networks of fake accounts and fake pages being used, in country after country, to manufacture the appearance of public support for national political figures. BuzzFeed News obtained it and published extracts within days. The line most people remember is her own: she wrote that she had blood on her hands. Then the memo came down, and for the most part, so did the story.

I keep returning to her case because it is the cleanest test I know of a question that sits underneath everything I write about how technology gets captured. Not “does the platform cause harm?” — we have argued that one to exhaustion. The sharper question is: when a platform does act against harm, what is it actually responding to? Her account suggests an uncomfortable answer, and the discomfort is not really about one company.

Who she is, and what she was actually hired to do

Zhang is an American data scientist. She joined Facebook in 2018 and worked on the Site Integrity organisation’s fake engagement team — the group concerned with inauthentic likes, comments, shares and reactions. That is worth sitting with, because it is a fairly narrow brief. Her job, as assigned, was closer to spam and gaming of metrics than to the integrity of elections. The political manipulation work that made her name was, by her account and as widely reported, something she took on herself, largely outside her formal remit and on top of what she was actually paid to do.

That detail changes the shape of the story. She was not a senior executive with a mandate and a budget. She was a junior-to-mid-level analyst who happened to have query access to the right data and who, having stumbled into something, could not persuade herself to look away. Much of what followed — the improvised investigations, the escalations that went nowhere, the sense that she was the only person on a problem that plainly needed a team — follows from that starting position.

What she says she found

The pattern she described was not sophisticated in a technical sense, which is part of why it unsettles me. According to her account and the reporting built on it, an operator could create large numbers of Facebook pages — not user profiles, which were subject to real-name rules and duplicate detection, but pages, which were not — give them plausible names and photographs, and use them to like, share and comment as though they were ordinary people. The result was a crowd that did not exist, cheering at scale. One of the journalists covering it called it the digital equivalent of bussing in a fake audience for a speech, and that is about right.

The clearest case in the reporting is Honduras. Zhang said she found a very large network of fake pages inflating engagement on posts by the then president, Juan Orlando Hernández, whose 2017 re-election was already widely contested. She said she traced a substantial share of the fake activity to an operation connected to his administration. In Azerbaijan, she said she found the ruling party running a comparable apparatus, in that case aimed outward — fake entities used to harass opposition figures and independent voices in comment sections at volume.

Beyond those two, she has said she saw coordinated campaigns of varying scale and confidence in a long list of other countries — India, Ukraine, Spain, Brazil, Bolivia, Ecuador and others among them. I would treat that list the way the reporting does: as her account of what she encountered, at very different levels of certainty, not as a verified ledger of state-run operations. Some of those networks were removed. Some, she says, were not.

The manipulation was not technically clever. It exploited an ordinary product feature that nobody had been assigned to watch.

Her structural complaint was sharper than any individual case. There was, she said, no team whose job it was to find and remove abusive pages of this kind — an enforcement gap, in her phrase. A policy existed, or came to exist; the staffing to act on it did not follow. Anyone who has worked inside a large organisation will recognise the shape of that failure. It is not a conspiracy. It is a line item that nobody owns.

The part that actually matters: what triggered a response

The accusation people expect from a whistleblower is that the company approved of the bad thing. That is not her claim, and flattening it into that does her argument a disservice. Nobody in her telling defended fake crowds. As she recalled the internal reaction to MIT Technology Review, everyone agreed it was terrible; what nobody could agree on was who should be responsible, or what should actually be done.

Her claim is narrower and, I think, worse. It is that the thing which reliably converted agreement into action was not the scale of the harm but the scale of the risk to the company — press attention, regulatory heat, the prospect of a hearing. Where that risk was high, resources appeared. Where it was low, the case waited. By her account, the Honduras network she raised took close to a year to be acted on, while cases in markets with active newsrooms and interested legislatures moved considerably faster.

If that is right, then enforcement was not neutral. It was a function of coverage. And coverage is unevenly distributed across the world in ways that have nothing to do with where people are being harmed. A fabricated crowd in a country that Western newsrooms staff thinly is not a smaller harm to the people living in it. It is simply a smaller problem for the company, because fewer people with power are going to hear about it. That is the mechanism I find genuinely disturbing — not malice, but a triage system that mistakes attention for severity, in a company whose engagement algorithms are already tuned to whatever generates the loudest response.

Leaving, and what the company said

Zhang was dismissed in 2020. By her account she turned down a severance package reported at around sixty-four thousand dollars, because it came attached to a non-disparagement agreement, and she wanted to keep the right to speak. She posted the memo on her way out, anticipated that it would be removed, and put a copy where colleagues could still reach it; that copy, she says, was taken offline too.

In April 2021 she went fully public in a Guardian series built on her documentation, with her name and face attached. In October 2021 she gave evidence by video link to the UK Parliament’s joint committee scrutinising the draft Online Safety Bill, where she argued that the company knowingly tolerates manipulation by authoritarian governments when acting would be inconvenient. She has said she made material available to US authorities and was willing to testify before Congress.

The company disputed her characterisation, and that belongs in the record. A Facebook spokesperson said the company fundamentally disagreed with her characterisation of its priorities and its efforts to root out abuse, said that combating coordinated inauthentic behaviour was a priority, and pointed to the many networks it had taken down and to a dedicated team continuing the work. Notably, the substance of what she found was not really contested; the argument was about whether it represented neglect or ordinary prioritisation under constraint. I cannot settle that from outside, and I am not going to pretend otherwise. What I can say is that “we had to prioritise” is an answer that invites the obvious follow-up — prioritise by what?

Why you have probably not heard of her

Two weeks before Zhang appeared in front of British lawmakers, Frances Haugen testified before a US Senate subcommittee, and the world stopped to watch. Haugen went first in the public imagination, arrived with a large trove of internal documents, had a coordinated media and legal strategy behind her, and — this matters more than anyone likes to admit — told a story about harm to American teenagers and American democracy. Zhang went public months earlier, alone, with a story about Honduras and Azerbaijan.

She was arguing that harms in less-covered countries wait longer for attention — and the reception her own disclosure got demonstrated the point rather neatly.

I do not say this to diminish Haugen, whose disclosures were substantial and whose choice to put her name to them was costly. I say it because the comparison is itself the evidence. Zhang’s central claim was that the world’s attention allocates itself unevenly, and that platforms quietly track that unevenness when deciding what to fix. Then she made her disclosure, and it was received unevenly, along exactly the lines she had described. Her thesis was proved on her own body of work.

That is the part worth carrying away. Every list of history’s whistleblowers is also, silently, a list of which revelations happened to land in front of an audience that could be bothered. The ones whose findings concerned people far from the newsrooms tend to be remembered as footnotes, if at all — not because they were wrong, but because being right is not sufficient. Someone has to decide the harm counts.

Sophie Zhang did the work that a large integrity organisation had not assigned to anyone, gave up money to keep her voice, said what she found, and then largely dropped out of the conversation. The networks she described did not all disappear. If you want one lesson from her, it is this: when a platform tells you it is prioritising, ask what the ranking criterion is. And when the answer turns out to be the risk of being written about, understand what that means for everyone the press has never gotten around to writing about.