Sometime this weekend, perhaps late at night, you may have typed a question into the assistant that now lives inside Facebook, Instagram, and WhatsApp. It might have been about a loan you cannot cover, or a parent whose health is slipping. Since December 16, 2025, Meta has used those exchanges to choose the posts and ads you see. The company’s own example was gentle: chat with its assistant about hiking, and offers for hiking boots may follow. Reporting at the time found no opt-out short of staying away from the assistant altogether.
The same notice carried a promise. Conversations about politics, health, religion or sexual orientation, Meta said, would not be used to show anyone ads. That promise arrives with a track record attached. On October 21, New Mexico District Judge Francis Mathew ordered Meta to pay $25 billion after a state jury had found the company liable for more than forty-three million violations of New Mexico’s consumer-protection law and judged twenty-six of twenty-nine statements by Meta and its leadership on privacy and content moderation to be misleading. From the bench, Mathew summarized the core finding in two short sentences: Facebook told users they controlled their data, and that was untrue.
A second thing has probably already happened to you, and you had no way to notice it. If you voted in a battleground state in 2016, Donald Trump campaign’s database may have sorted you into one of eight algorithmic audiences, each fed its own messages. Channel 4 News, which obtained that database, reported that the campaign ran almost six million versions of its Facebook ads, many of them dark posts that vanished from feeds once the campaign stopped paying. Facebook had no public ad library at the time, so no complete record survives. Your neighbor may have seen an entirely different campaign from the one you saw, and neither of you could have compared notes.
The $25 billion is the largest penalty Meta has ever faced, five times the $5 billion the Federal Trade Commission imposed on Facebook in 2019. New Mexico Attorney General Raúl Torrez said all of it will go into a fund for the state’s education system. The suit, filed in 2021, grew partly out of the Cambridge Analytica affair. Meta will appeal. Its lawyer Matt Nicholson argued that the jury never found any New Mexican had seen the statements or been harmed by them. Chief legal officer CJ Mahoney went further, calling the lawsuit an attempt to impose “a tax on speech.”
Anyone who analyzes institutional language for a living notices the reclassification. A penalty for broken privacy promises has been relabeled—by the company that broke them—as a levy on expression. Organizations tend to describe their products in the categories they actually run them in. On Meta’s books, the gathering of what you reveal and the messages built from it appear to have fused into one operation, so a blow against the first registers as an attack on the second. That fusion is the real subject of this verdict, and it cuts in two directions.
The first edge is collection. In 2014, a Cambridge University academic named Aleksandr Kogan built a personality quiz app. According to the United Kingdom Information Commissioner’s final account to Parliament, some three hundred thousand Facebook users took it, and because the app could also read data belonging to their friends, Facebook estimated that information on up to eight-seven million people was swept up. The overwhelming majority of them never saw the quiz. Consent, in that design, was given by someone else on your behalf.
The Commissioner’s investigation also surfaced the part of the story that rarely survives the retelling. Cambridge Analytica’s parent firm was buying commercially available personal data at enormous scale, by one estimate over 130 billion data points, mainly on American voters, from sources that included Acxiom and Experian. The Facebook material was matched against that purchased file by name, birth date, and location, then used to enrich it. The scandal concerned one leaky app. The infrastructure underneath it was an open market, and that market never closed.
Today the quiz is the chat window. Meta says more than a billion people use its assistant every month, and since December their exchanges with it feed the same recommendation machinery as their likes and reels. For a behavior analyst, the difference in signal quality is enormous. A like is a public gesture, made with an audience in mind. A question typed to an assistant at midnight is closer to a diary entry, specific and unguarded and stamped with the mood of the hour. The psychologists who built this field saw it coming. In their 2017 paper on psychological targeting, Sandra Matz and her colleagues noted that mood can already be inferred from written language, and that moments of good mood, when people lean on mental shortcuts, may be the best windows for persuasion. Meta’s ad-ranking model, the company reported in January, now trains on longer sequences of user behavior with double the computing power it had before.
The second edge is application, and it rests on two findings. The first concerns prediction. In 2015, researchers at Cambridge and Stanford showed that a model reading Facebook likes judged personality better than a work colleague after ten likes, better than a friend after seventy, better than a family member after 150 and better than a spouse after three hundred.
The second concerns matching. Across field experiments that reached more than 3.5 million people, Matz and her co-authors found that ads tuned to a person’s level of extraversion or openness drew up to 40% more clicks and up to 50% more purchases than mismatched or generic versions. Their targeting relied on a single like per person. One tailored line, written for an introverted audience of mobile gamers, read: “Phew! Hard day? How about a puzzle to wind down with?” The standard ad it outperformed promised intense action.
The mechanism is ordinary, which is why it works. Psychologists call it self-schema matching, and the Matz paper builds on that literature. A message framed in the vocabulary a person already uses about herself gets processed as recognition, and recognition lowers the guard that a sales pitch raises. The introvert at the end of a long day reads the puzzle ad and feels understood. Nothing in the experience signals that a machine sorted her into a bin.
The more consequential mechanism operates at the level of the crowd. Mass persuasion in the broadcast era carried its own brake. A newspaper advertisement or a television spot reached everyone in the same words, so rivals could answer it and reporters could quote it. Persuasion happened in public, and public text invites counter-argument. Individual tailoring removes the shared text. Each person meets a version written for her alone, and the people best placed to rebut that version never see it. For the target, the experience is a private agreement with an argument that seems to come from nowhere in particular. For the persuader, the cost of a false or inflammatory variant drops close to zero, since there is no audience in front of which to be embarrassed.
Honesty about the record matters here, because the legend has outrun the evidence. The Information Commissioner found that Cambridge Analytica’s models were built mostly from off-the-shelf tools, that their real-world accuracy was likely much lower than advertised, and that the company’s own staff were uneasy about their leadership’s boasts. The Matz experiments drew a published critique arguing they failed to show that personality targeting beats ordinary advertising, and the authors replied that their effects were robust. The evidence does not support the story that one consultancy put a president in the White House.
The trouble is that the weakest link of 2016, the guessed personality profile, is the part the newest machinery barely needs. In December 2025, Nature published experiments by David Rand, Gordon Pennycook, and colleagues in which voters held short text conversations with AI models advocating for a candidate, a median of about six minutes in the American trial. Among Trump supporters, the pro-Kamala Harris model moved candidate preference by 3.9% on a 100-point scale, against effects of roughly one point for traditional political ads in earlier studies. In Canada and Poland the effect was almost three times the American one, and about a third of the United States shift was still there more than a month later. When the researchers told the models not to personalize, persuasion held up. When they barred facts and evidence, it fell by more than half. The persuasive engine was a responsive interlocutor answering each person’s own stated concerns.
The researchers also gave the American model a second goal: to lower the likelihood of voting among people who favored the other candidate. On that count it failed. The goal itself was familiar. Channel 4 reported that the 2016 Trump campaign placed 3.5 million black Americans in a category it named Deterrence, people it hoped would stay home on election day. The campaign’s digital director had said he was nearly certain it ran no campaign targeting African Americans.
For an analyst, a chatbot conversation is the purest form of the individual message. Every sentence is composed for one listener in reply to her own words, and the transcript exists for no one else. Cambridge Analytica had to infer who you were from a quiz your friend took. The assistant simply asks, and you answer.
That conversation sits inside a system with two kinds of customers. Commercial and political advertisers are the first. Meta’s video generation tools for advertisers reached a $10 billion annual revenue run-rate by the last quarter of 2025, and the company describes its goal as uniquely personal experiences for users alongside better performance for the businesses paying to reach them. The ad you see is increasingly written by a machine for a profile of you.
Governments are the second. A panel advising the Director of National Intelligence concluded in January 2022 that American intelligence agencies were buying commercially available data in significant volume, and that such data “can reveal sensitive and intimate information about individuals.” The report stayed classified until June 2023. The Department of Homeland Security takes the position that the Supreme Court’s warrant requirement for cellphone location records does not apply when the same records are bought from a data broker. In 2025, Immigration and Customs Enforcement (ICE) signed contracts for a social media monitoring platform that claims to analyze more than eight billion posts a day, and for Penlink tools that combine social media activity with location data to build profiles of a target’s daily movements. Agency leaders have said they intend to turn these capabilities on anti-ICE protesters and those who fund them.
The purchases span administrations of both parties. The state has no need to seize what the market has already gathered. It buys the file with the taxpayer’s money, and a profile assembled to sell hiking boots can end up placing its owner at a protest.
Which brings the argument back to Meta’s defense. A company that calls a fraud penalty a tax on speech is inviting libertarians to see it as one more victim of state overreach, and some will accept the invitation. Libertarian philosopher Murray Rothbard’s theory of contract offers a cleaner answer. In The Ethics of Liberty he argued that a contract should be enforceable when the failure to fulfill it is “an implicit theft of property.” Users handed over their data on stated terms, chief among them that they would control who saw it. The court found those terms false. Data held under conditions its owner never agreed to is held without consent, and Rothbard’s framework files that under theft. It leaves no room to call the arrangement expression.
The opposing case deserves its full weight. Nicholson’s objection stands as a legal argument: the jury found statements misleading without finding that any particular New Mexican relied on them or suffered for it. Some libertarians reject the premise that behavioral data is property at all. And the verdict reached Meta’s statements about content moderation, which is where a state penalty can shade into pressure on what a platform publishes. That portion of the judgment merits close attention on appeal.
Then there is the money. Torrez said every dollar will go to New Mexico’s education fund. The people whose data was taken on false terms are not in line for any of it. The penalty flows to a government, the same kind of institution that, at the federal level, buys harvested data on the open market. The defrauded user loses once to the company and again to the treasury that collects in her name.
In September 2020, a Facebook spokesperson told Channel 4 that what happened with Cambridge Analytica “couldn’t happen today.” Five years later, Meta announced that conversations with its assistant would become one more signal for its feeds and ads. The friends whose data left through a quiz in 2014 have been replaced by you, typing freely at midnight to a machine that answers in a voice composed for one reader. Whatever the appeals court does with the $25 billion, that conversation is still being recorded, and the account it feeds is still open.
