Whistleblowers protect our freedom, we protect their truth.
Whistleblowers are essential to democracy. They expose corruption, abuse, and the stories powerful institutions would rather keep hidden.
But whistleblowing systems are increasingly vulnerable to spam, fabricated evidence, and contradictory claims that bury credible reports under noise. As generative AI lowers the cost of producing convincing fake material at scale, that problem will only intensify.
Palantir for the People equips journalists and investigators with the tools to defend against this new wave of information warfare — helping them verify evidence, detect inconsistencies, and surface credible stories before the truth is drowned out.
Palantir for the People surfaces signal in the face of noise.
We classify, grade, and filter large document dumps so reporters can find the documents that matter. The agent processes the full corpus, prioritizes the sub-groups most worth a reporter’s time, and ranks individual documents on how readily their claims can be externally verified. Documents that contradict themselves are filtered out.
A triaged corpus. Documents are grouped by topic, ordered by significance, and graded for verifiability. Self-contradicting material is set aside. We do not write, summarize, or recommend. We grade. The reporter is the editor.
Not a secure submission platform. Several well-considered open source projects are dedicated to that problem.
Not a truth oracle. LLMs are weak at identifying facts outside their training data, and the information in tips is, by definition, outside the public domain. If you could google it, it would not need to be whistleblown.
Not a replacement for good journalism.
We run several rounds of LLM-driven classification. Case analysis processes the entire document dump to identify the high-level topics the documents belong to. Topic analysis processes all the documents within a single topic to characterize the case being made. Per-document analysis grades individual documents to surface red flags and information-rich characteristics.
For the full breakdown, see the methodology page.