Is Palantir trustworthy?

Clear the folklore first

The Palantir debate runs on claims that fail verification, in both directions. The full fact-check grades each one against primary sources. The short version:

Anti-Palantir claims that fail

• FALSE “Palantir owns and sells your data.” Customers own their data, contractually and per SEC filings. The real capability it sells is integration of data its customers already hold.
• UNSUPPORTED “Palantir built Lavender and Gospel,” the Israeli targeting systems. The original reporting attributes them to Israeli military intelligence, and Palantir denied both by name in its letter to the UN.
• FALSE “Palantir built the Muslim registry.” It never existed; the 2016 protests were about capability, and the protest became remembered as the event.
• EXAGGERATED “A master database on all Americans.” Deployments are per-agency instances and the government holds the data. The real, documented story is 2025 cross-agency data-sharing and a $1B DHS agreement.

Pro-Palantir claims that fail

• UNSUPPORTED “Palantir found bin Laden.” A documented laundering chain the company never corrected, and CIA insiders called unearned.
• UNSUPPORTED “Palantir has stopped hundreds of terror attacks.” Zero publicly documented attributed cases; the most-quoted example is a fictional character from Palantir’s own product demos.
• EXAGGERATED “Palantir only serves democracies.” The Gulf business exists, and the CEO concedes those clients have “their own code of ethics.” What survives: a real, SEC-disclosed refusal of China and Russia.

Some of what matters most is classified: any specific counterterror outcome, the technical role inside Israeli systems, error rates of the immigration tools. Classification protects the company’s heroic claims (unfalsifiable, so they persist) while starving accountability claims (unprovable, so they die). The fog is asymmetric, and honest analysis says so.

Four questions, not one

“Is Palantir good?” collapses four separable questions. Answered separately (full synthesis), the fog clears:

• Should the West build this capability? Largely yes. The diagnosis holds, and the demand is real and mainstream.
• Is the tooling better than the alternative? Yes. Audit logs and access controls beat the unlogged baseline, per the government’s own privacy assessments.
• Have the deployments been good? Mixed, honestly. Ukraine and vaccines on one side; the ICE record, the watchlists, and the voided German deployment on the other.
• Is the institution’s word a good predictor of its behavior? No. One clean ethics-versus-incentives test on record, and ethics lost it; checkable candor is impeachable; the structure prevents correction.

The pushback that changed the analysis

The conclusions were argued against before they were published. The full dialogue is part of the evidence; the exchange that forced a concession:

Aren’t they doing this because they think Western values should win? They could make way more money working for Russia or China.

The money claim is backwards. Selling to either would forfeit the clearances and the entire US government book; refusing adversaries protects their largest revenue stream. When ethics and self-interest command the same action, the action tells you nothing about which is driving. And the “values” line sits exactly on the US alliance map: Gulf authoritarians yes, Washington’s adversaries no. The concession: their pro-West conviction is probably sincere. But exactly once on record did ethics and incentives point opposite ways, the 2020 refusal versus the 2025 acceptance, and ethics lost. A value that never wins a conflict with incentive is real as psychology and useless for prediction.

Aren’t “pro-Western state” and “pro-Western values” basically the same?

They look the same from outside the country and come apart inside it. Western values, at bottom, ARE constraints on the state: consent, due process, accountable power. The West is the West because the state is strong AND leashed. A company can strengthen the muscle while eroding the leash, and the domestic record is that transaction, repeated.

The predictive model

Its one dependable commitment, never arm Washington’s adversaries, is dependable precisely because conviction and self-interest fully reinforce each other there. Inside that floor, incentives predict behavior and stated principles do not; when a position flips, the principles get restated to make the flip look like consistency.

The company’s operative ethic is loyalty to Western state power. Western values are the leash on that power, and the leash is what the record shows being traded away: health data reaching deportation tools, programs hidden from city councils, every audit resisted, every oversight proposal dead under permanent founder control.

The mission passes. The instrument passes. The institution fails on verification, and it fails by choice. If these tools matter as much as the company says, they deserve more independent scrutiny than a photo-sharing app. Wanting the capability AND wanting its vendor auditable, candid, and governable is the most genuinely pro-Western-defense position available.

Sources and method

This site is the readable version of an open research project in the Substrate repository: four independent AI research efforts across different model vendors, a neutral factual record read from SEC filings and court documents, a claim-by-claim fact-check, an adversarial review pass, and roughly 150 verified sources in the source inventory. The methodology, including its limitations, is published with it.

The research was executed by AI systems with scaffolding built to emulate research rigor, and it should be read as a rigorous starting point rather than a settled answer. Found an error, or a stronger argument for either side? Open a pull request. Claims with sources beat claims without them.