What Does Palantir’s AI Do, and What Risks Do Critics Raise?

Palantir sells data-integration and analytics software, now including AI tools, to government and commercial customers. Its four main platforms are Gotham, Foundry, Apollo and the Artificial Intelligence Platform (AIP). Critics, including Amnesty International, the Electronic Frontier Foundation (EFF) and Germany’s Society for Civil Rights, say that in government use, especially immigration enforcement and police analytics, the software carries serious human rights and privacy risks. Palantir has disputed parts of this. This article separates what Palantir says about its products from what named critics allege.

What Palantir says its products do

Palantir’s annual report on Form 10-K for fiscal 2024 describes the four platforms. The descriptions below are Palantir’s own, from its SEC filing.

• Gotham, according to the filing, helps users identify patterns in datasets, including signals intelligence sources and reports from confidential informants, and helps hand off work between analysts and operational users.
• Foundry “transforms the ways organizations operate by creating a central operating system for their data. Individual users can integrate and analyze the data they need in one place.”
• Apollo “enables the rapid, secure delivery of our software and updates across our business, and also enables our customers to securely deploy their own software in virtually any environment.”
• AIP “enables responsible AI-advantage across the enterprise and government by using primary, core components built to effectively activate LLMs and other AI within any organization.”

In plain terms: Gotham is the analyst and operations tool used in intelligence and security settings, Foundry is the data-integration layer for organizations, Apollo is the deployment system, and AIP connects large language models and other AI to an organization’s data.

Risks Palantir itself discloses

The same filing lists risk factors. They include that Palantir “may fail to receive and maintain government contracts or there may be changes in the contracting or fiscal policies of the public sector,” that “unfavorable news or social media coverage may harm our reputation and business,” and that “issues raised by the use of AI (including machine learning and large language models) in our platforms and business may result in reputational harm or liability.” The filing also says Palantir “may not enter into relationships with potential customers if we consider their activities to be inconsistent with our organizational mission or values.”

What critics claim

Amnesty International: immigration enforcement due diligence (2020)

On 28 September 2020, Amnesty International published a briefing on Palantir’s work for US Immigration and Customs Enforcement (ICE). Amnesty’s assessment was that Palantir’s contracts with ICE create a risk that the company contributes to serious human rights violations against migrants and asylum-seekers; it also described how the software lets ICE combine data about people. This is Amnesty’s position, not a court finding that the company committed any violation. Michael Kleinman, director of Amnesty’s Silicon Valley Initiative, said in an Amnesty USA press release that the company’s ethical commitments were “deeply ironic, given the company’s willingness stateside to work directly with ICE.”

EFF: ICE tools and medical data (2026)

In an April 2026 post, the Electronic Frontier Foundation argues that Palantir’s stated human rights commitments conflict with its work for ICE. EFF says a Palantir tool called ELITE lets ICE use Medicaid data, and argues that using sensitive medical information for enforcement sweeps is hard to reconcile with the company’s stated commitments to privacy and equity. EFF writes that “human rights are an outcome, not just a process.” This is EFF’s position, not a court ruling.

For context on ICE enforcement patterns, a Cato Institute analysis updated through 28 July 2025 reported that about 19% of ICE arrests in July 2025 were street arrests of people with no criminal charges, convictions or removal orders, a group the author describes as roughly one in five and as Latino. That figure covers one period and does not show what role, if any, Palantir’s software played.

Germany: police data analysis and the constitutional court (2023)

On 16 February 2023, Germany’s Federal Constitutional Court ruled against provisions in Hesse and Hamburg that authorized automated data analysis by police. I could read only the court’s official press release, not the full judgment, so this article does not quote the ruling or say whether it names any software vendor. The Society for Civil Rights (GFF), which campaigns on the issue, says on its Palantir page that the court “set narrow limits on automated data analysis by the police in a landmark ruling in 2023,” and about Palantir’s software that “the software secretly analyzes citizens’ data. They are unaware of this, just as they are unaware of any surveillance measures that may follow.” GFF also says the Bavarian legislature “has not complied with these standards.” These are GFF’s views. The court’s ruling concerned state laws, not Palantir’s conduct.

How Palantir has responded

• In a letter to Amnesty dated 18 September 2020, Palantir distinguished between ICE’s Homeland Security Investigations (HSI) and its Enforcement and Removal Operations (ERO). Its statement that its technology is not used in detention or deportation was limited to ERO’s civil immigration enforcement. This is the company’s own account.
• Palantir published a response to an EFF report from January 2026. In it, the company disputes the description of its work as a single, huge unified database, and describes how ELITE works, including limits on who can access data and auditing. That is Palantir’s account of its own product; it does not settle the disagreement with EFF.

How to weigh these claims

• Palantir’s descriptions are marketing-adjacent but appear in a legally regulated SEC filing, so they are reliable as statements of what the company claims to sell.
• The critics’ claims are mostly about how customers use the tools, particularly ICE, rather than about the AI technology itself. Whether a vendor bears responsibility for how government customers use its tools, and whether those uses and safeguards comply with the law, is the core disagreement.
• The German ruling shows a separate type of risk: the legal limits on police data analysis, regardless of vendor.
• Amnesty, EFF and GFF are advocacy groups with stated positions, and Palantir’s responses are its own. None of these statements is an independent finding of wrongdoing.