How Scared Should We Be of Palantir?

A former U.S. secretary of labor says Palantir is a grave threat to Americans. Thirteen of the company's own former employees say its leaders have abandoned the values it was built on. The company now supplies software used in the most aggressive deportation campaign in modern American history.

This is the companion piece to the second episode of my Palantir investigation. The video does what the video does best: it moves, it shows you the systems, it puts the documents on screen and the map in front of you.

Both tell the same story, and it's a story built on a single discipline: separating what the record can prove from what you might imagine.

▶ Watch the episode:

The Accusation Nobody Can Quite Prove

Robert Reich drew four circles. Artificial intelligence. The military. The federal government's collection of personal data. And the part of Silicon Valley he believes is trying to turn a democracy into something else. Where the circles overlap, he wrote, sits a single company and a single man: Palantir, and Peter Thiel.

The specific charge is that Palantir is building an AI-powered super-database on every American. Palantir says that's false. It says it does not own its customers' data, does not combine it into one master file, and was not hired to build such a thing.

Two witnesses. One says grave threat. One says false. Most people pick the version that matches what they already believed.

But before we ask what a national database might do to us, we have to establish whether anyone has shown it exists.

What the Record Shows, and What It Only Suggests

The all-encompassing national database does not exist in the evidence. Not the way Reich says it does.

Here is what is verifiable. Palantir works inside multiple federal agencies that hold sensitive information. Its software is built to connect scattered data and make it usable. The Trump administration has said, in writing, that it wants government information easier to share. Put those facts together and a super-database becomes easy to imagine.

Imaginable is not the same as built. No disclosed contract shows Palantir assembling one file containing everything the government knows about everyone. One agency using Palantir's software does not hand another agency the keys. Compatible tools are not shared permissions, and none of it is proof that the sharing has occurred.

The way Reich handles a particular number tells you how he perception-engineers the larger claim. He writes that Palantir disclosed Karp was paid $6.8 billion in 2024. The filing he cites uses a required accounting label, "compensation actually paid," then states in plain language that the figure is not compensation actually paid, earned, or received. Easy to misunderstand that SEC terminology that mostly reflects the rising value of stock. Reich read the number in exactly the way his own source warned him not to.

That is his perception-engineering method in miniature. Take real ingredients. Combine them into the most alarming shape possible. Present the possibility as fact.

One rule governs everything that follows, and it cuts both ways. What we don't know is not evidence. Secrecy can't prove the hidden system a critic imagines, and it can't verify the assurances a company offers.

The master database has not been proved, even it the government's effort to make its information easier to connect has been proved completely.

The Thing Underneath That Turned Out to Be Real

Strip away the myth and something solid remains.

In March 2025, an executive order instructed federal agencies to increase access to, sharing of, and consolidation of unclassified records, to the maximum extent the law allows. This is not a rumor about a secret program. It is a signed instruction. Then the Treasury Department hired Palantir to help build a common interface, a way for separate government systems to request information from one another.

That interface is the pivot the whole story turns on. Separate systems can be made to exchange selected information without physically merging into one giant database. Reich pictured a single vault. The reality is more like a set of doors installed between rooms that used to have none.

For the person on the wrong end of a government decision, the distinction may not comfort. It matters enormously for law and oversight whether the data sits in one place or many. But if an authorized user can reach across the rooms, connect what they find, and act on it inside one workflow, the walls between the databases start to feel decorative.

So the fear was aimed at the wrong target. It was never really about centralization. It was about access. Which raises the next question, the one the accusers seemed positioned to answer better than anyone. If outsiders got the architecture wrong, surely the company's own people can tell us what is really happening inside.

The Insiders Who Could Not See Inside

Thirteen former employees signed a document titled "The Scouring of the Shire." It says Palantir's leadership abandoned commitments to democracy, privacy, accountability, and the rule of law. It's a serious document from people with real history at the company, and it establishes something worth having: the alarm about Palantir is not only an invention of outside activists. It grew from the inside.

Ari Gesher was an early engineer with roughly a decade at Palantir, but he left years before the programs now in dispute. Linda Xia was there until 2024 and can speak to the recent company, but she has said her work was commercial, not the government systems at issue. Their standing is cultural and architectural. It's not operational access to the thing being alleged. The letter is ethical testimony, not a leak.

And yet the signers point, without needing to reveal a single secret, at the strongest piece of evidence in this entire story. Because the most damning witness against Palantir is Palantir.

In 2020, the company said it had deliberately declined to work with ICE's Enforcement and Removal Operations, the branch that arrests and deports, citing the risk of serious human-rights violations. In 2025, it accepted exactly that kind of work: enforcement prioritization, departure tracking, removal logistics. No critic is required to establish the reversal.

Why was a line the company drew in public in 2020 gone five years later?

The Defense That Holds, Right Up to the Line It Can't Cross

Palantir's case is not public-relations filler. The company generally does not own or sell its customers' operational data. Separate deployments do not automatically pool. The customer sets the mission, the legal authority, the users, the permissions, the criteria.

The platforms carry granular access controls, approval requirements, audit trails, and human review. Better-integrated records can even surface errors: a mistaken identity, a court-ordered stay that should stop an action cold. And it is ICE, not Palantir, that chooses whom to pursue and carries out the arrest.

Every one of those statements narrows the accusation. And every one answers the same kind of question. Was this user allowed to see this record? What did they do with it? Can an investigator reconstruct it later? Those are questions about authorization, and Palantir's tools answer them well.

But broader, systemic questions remain. Should the government have been authorized to do this at all? Is the policy lawful as applied to this person? Are the criteria accurate, proportionate, just? Should Palantir have taken the contract? An audit log can tell you an action was permitted. It can't tell you the action was right.

Palantir's human-rights policy says legal and privacy constraints limit its ability to monitor everything its customers do. Its proxy statement says customer ownership of data limits how much the company can dictate outcomes. It says confidentiality may prevent outsiders from ever assessing its most sensitive work in detail.

None of that proves misuse. And none of it lets the public confirm the safeguards work. The company can govern who is authorized to act far more effectively than whether the authorized act should have happened.

What does this all look like in practice?

What the Software Actually Did

In April 2025, ICE expanded its Palantir contract by roughly $30 million for a system built to do three things: 1) help decide whom to pursue, 2) track where cases stood, and 3) move selected people through the deportation process more efficiently. That's the contracted purpose, stated plainly.

Then there's ELITE.

Reporting drawn from internal ICE material, procurement records, and sworn testimony describes a Palantir-supported application that displays people ICE has selected, assembles a dossier on an individual, scores confidence in a current address, and shows potential targets on a map, including where more sought people may be found.

Palantir disputes the picture of an indiscriminate raid tool, saying ELITE surfaces likely addresses for specific people who already meet ICE's criteria, and can reveal legal barriers like a court-ordered stay.

ICE sets the criteria. The software integrates data about the people who match, and helps officers find and rank them. The underlying information can be wrong. Humans are supposed to check.

Under oath in Oregon, an ICE officer described ELITE as something like a map, used to find areas with more people carrying an "immigration nexus," and admitted even a high-confidence address could be wrong. The judge criticized ICE's conduct. The evidence doesn't show that ELITE alone caused a particular arrest, much less that a court found Palantir responsible.

The documented Palantir is not the omniscient machine that secretly picks its victims. It's something far less cinematic, though more consequential: infrastructure that helps a government turn scattered information into coordinated action against specific people.

The myth gets in the way. The verified mechanism is disturbing enough. Which leaves the hardest question.

Who is responsible?

Everyone's True Alibi

Ask each participant and each gives an answer that is accurate. Palantir says it didn't write the law, choose the policy, or make the arrest. True. ICE says its officers, not the software, made the final call. True. The officer says the system produced the lead. True. The agency that supplied the data says it acted under its own authority and did not control what came next. Also true.

Every alibi holds. The outcome still required all of them together. Responsibility did not vanish. It was divided so finely that each actor can point, honestly, at another. That's the structure of the arrangement, and it's what makes accountability so slippery.

It does not follow that Palantir carries the government's legal authority. It doesn't. But it carries something meaningful: the customers it accepts, the capabilities it agrees to build, the safeguards it designs, the accuracy its tools imply, and the truthfulness of what it tells the public.

Those are not the government's to answer for. They are Palantir's.

The Honest Answer

So. How scared should you be?

There's no proven Palantir-owned database holding everything about every American. A roster of politically connected billionaires is not evidence of a specific hidden machine. Thirteen former employees didn't expose a secret current system. The language that turns a software vendor into an all-seeing sovereign is doing emotional work, not evidentiary work.

Be more concerned about the federal effort to make sensitive information easy to connect and act on. That's real and on the record. Palantir's role inside consequential government workflows is growing. It reversed a boundary it once proclaimed.

Meanwhile, the public can't see enough of the data, the criteria, the accuracy, or the field use to check the company's assurances. Safeguards built to stop unauthorized access do nothing about authorized policy that should never have been authorized. And responsibility is scattered across enough hands that no single one has to hold it.

The danger was never that Palantir knows everything about us. The evidence doesn't support that. The danger is that Palantir can help powerful institutions act on far more of what they already know. When the institution is a hospital, that helps someone get a bed. When it's ICE, it helps someone get detained. The software chose neither mission. It made both easier to carry out.

That is why the honest answer is not panic. And it's not reassurance.