Palantir's ICE Infrastructure Is Now Screening Federal Grants for 'Gender Ideology

What HHS Built in Secret

A federal agency has been quietly running grant applications through an ideological filter built by a defense contractor and a two-year-old startup. The program was not announced. It was discovered in a required inventory of the agency's own AI use cases.

Since March 2025, the Department of Health and Human Services has deployed AI tools from Palantir Technologies and Credal AI inside its Administration for Children and Families to screen grants, grant applications, and job descriptions for language the White House has labeled "DEI" or "gender ideology." The disclosure appears in HHS's 2025 AI use-case inventory, published without fanfare. Neither Palantir nor HHS has publicly acknowledged the purpose of this work. HHS, ACF, Palantir, and Credal AI did not respond to WIRED's requests for comment.

The inventory describes an "AI-based" grant review process that "reviews application submission files and generates initial flags and priorities for discussion." Flagged material is routed to the ACF Program Office for what the document calls "final review." ACF staffers ultimately review any job descriptions, grants, or grant applications the system flags. The inventory marks these use cases as "deployed" — actively in use.

Palantir holds the sole contract to produce a list of "position descriptions that may need to be adjusted for alignment with recent executive orders." Credal AI, founded by two Palantir alumni, was brought in to help audit "existing grants and new grant applications." Federal spending records show ACF paid Credal roughly $750,000 last year for its "Tech Enterprise Generative Artificial Intelligence (GenAI) Platform." The payment descriptions in the Federal Register do not mention DEI or gender ideology. In August 2025, Credal was awarded a five-year definitive contract (140D0425C0056) worth up to $6.2 million for the HHS ACF Tech Enterprise GENAI Platform. A second sole-source award (7571TE26C00003) followed in August 2026, valued at up to $5.3 million over four years.

During the first year of Trump's second term, Palantir collected more than $35 million in payments and obligations from HHS alone. None of the transaction descriptions reference ideological screening.

The legal framework arrives from two executive orders signed on inauguration day. Executive Order 14151, "Ending Radical and Wasteful Government DEI Programs and Preferencing," demands termination of any policy, program, contract, or grant that mentions or concerns DEIA, DEI, "equity," or "environmental justice." Executive Order 14168, "Defending Women From Gender Ideology Extremism and Restoring Biological Truth to the Federal Government," defines "sex" as an "immutable biological classification" and recognizes only "male" and "female." It declares "gender ideology" and "gender identity" to be "false" and "disconnected from biological reality," and bars federal funds from promoting gender ideology. "Each agency shall assess grant conditions and grantee preferences and ensure grant funds do not promote gender ideology," the order reads.

The inventory does not list the specific terms or patterns the models flag. It does not describe the model architecture, training data, or false-positive rate. It does not explain how a system designed for enterprise data governance came to enforce political compliance on scientific grantmaking. The next section examines the technical mechanism.

How the Technology Works: Credal's Governance Layer

Credal's platform sits between an organization's data and the large language models that process it. The company describes itself as a "centralized trust layer for governing and scaling AI agents and MCP servers, connected to any system, operating within the security and permission boundaries your organization requires." In practice, that means Credal does not serve models itself. It brokers the connection: a user or automated workflow sends a prompt, Credal checks permissions, redacts sensitive data, routes the request to an approved model — Anthropic's Claude, OpenAI's GPT-4, or an enterprise-hosted variant on Microsoft Azure Foundry or AWS Bedrock — then returns the response, logging every step.

The architecture centers on three primitives: Agents, MCP Servers, and Actions. Agents are specialized, permission-aware systems that "reason, read your skills, and take action to complete goals across your enterprise data and systems." Each agent is authored by a subject-matter expert for a narrow task: researching a prospect, answering an HR question, auditing a contract. Agents can call other agents, forming composable chains while respecting the original invoker's permissions. This inter-agent authorization is enforced at every hop; an agent never sees data the user could not already access in the source system.

MCP Servers (Model Context Protocol) are the deployment layer. They expose governed collections of tools, data sources, and agents as a single endpoint that any compatible client — Claude Desktop, Slack, Microsoft Teams, ChatGPT, a custom workflow — can call. "Whether your users are in Claude, Slack, Microsoft Teams, ChatGPT, or a custom agentic workflow, Credal publishes a centrally managed context layer to any of these surfaces with minimal setup," the documentation states. Critically, governance rules travel with the server: access controls, human-in-the-loop approval requirements, and usage policies are identical whether the tool is invoked from a chat window or an automated pipeline. Admins can restrict access by user or team, require approval for sensitive actions, and monitor usage across every connected surface from a single dashboard.

Actions are the executable primitives: sending an email, updating a Jira ticket, querying a database, calling a private API. Credal ships an open-source library of pre-built Actions and supports custom endpoints with configured authentication. Every Action invocation is logged: what was requested, why, which agent invoked it, who approved it. The audit trail is immutable for the retention period and visible to authorized administrators. Human-in-the-loop gates can be placed on any Action; reviewers see the full decision context before approving or denying.

At ACF, Credal integrates directly with the grant-making infrastructure: SharePoint, Box, Confluence, Jira, GitHub, and GitLab. The HHS Privacy Impact Assessment notes that "users select specific documents or data from these systems to build their workflow assistants. Credal only surfaces content from documents users have access to in the source systems." Some integrations are bidirectional; configured Actions can write back to the source system. The platform ingests grant applications, review notes, and associated metadata into a Data Catalog that administrators can search, audit, and purge. All data remains in ACF's government-owned cloud tenancy; ACF supplies its own LLMs via enterprise subscriptions, ensuring prompts and completions are never used to train foundation models.

The PIA also reveals the scale: between 100,000 and 999,999 individuals' personally identifiable information resides in the system. Biometrics and resumes flow in through the Box integration. Administrators complete privileged Rules of Behavior training; some hold security certifications. Auto-redaction of sensitive fields — once a native Credal feature — is being sunset in favor of the enterprise LLM providers' native guardrails.

How this becomes an ideological filter is straightforward in principle. A workflow assistant — an agent — is configured with a prompt or classification schema that scans incoming grant narratives for terminology associated with DEI or "gender ideology." The agent reads every application the user has permission to see, flags passages, assigns a risk score, and routes the package to the ACF Program Office for final review. The WIRED investigation confirmed the inventory's description: the "AI-based grant review process" does so. Because the agent runs inside Credal's governed framework, every flag, every routing decision, and every human approval is logged and auditable. The same permission mirroring that prevents a program officer from seeing another division's HR files ensures they see every grant application — and only those — when the screening agent runs.

Palantir's role in this stack is less documented. The company's AIP Chatbot Studio "is built on the same rigorous security model that governs the rest of the Palantir platform," granting an LLM "access only to what is necessary to complete a task." Palantir has held HHS contracts for years; its Foundry platform already hosts ACF data pipelines. Credal's founders, Jack Fischer and Ravin Thambapillai, are Palantir alumni. The technical integration likely runs through shared identity providers, common data catalogs, or Palantir's own agent framework calling Credal-hosted MCP servers, but neither company has publicly described the interface. What is clear is that Credal provides the governance layer — permissions, audit, human-in-the-loop, model routing — while Palantir contributes the incumbent data infrastructure and the contract vehicle.

The result is a system that can be retargeted overnight. Change the classification prompt, redeploy the agent to the same MCP server, and the new policy propagates to every interface — Slack, Teams, custom dashboard — without reconfiguring permissions or rewriting access controls. That operational agility is what makes the platform attractive to an enterprise. It is also what makes it a potent instrument for political enforcement.

The Political Driver: Trump's DEI Rollback and Executive Orders

When Donald Trump signed Executive Order 13950 on September 22, 2020, he launched what would become one of the most contentious policy battles of his presidency. The order, titled "Combating Race and Sex Stereotyping," didn't just tweak federal diversity programs. It aimed to dismantle them. It prohibited nine specific "divisive concepts" from any training for federal employees, military personnel, contractors, and grant recipients, and it came with teeth: contractors found in violation could have their contracts canceled, terminated, or suspended, and face debarment from future government work. The administration even set up a hotline for employees to report violations by their employers. The order's real power lay in its strategic ambiguity—terms like "divisive concepts" and "psychological discomfort" lacked clear legal definitions, creating what lawyers call a "chilling effect." Rather than risk their federal contracts, many organizations simply canceled all diversity-related activities. The Department of Justice suspended all diversity training for its staff. The Office of Personnel Management paused such training for the entire Executive Branch. Military service members walked out of a screening of Spike Lee's "Malcolm X," fearing it could violate the new policy. Major defense contractors like Lockheed Martin and Boeing immediately began reviewing their training programs.

The Biden administration moved immediately to reverse Trump-era policies. On his first day in office, Biden revoked Executive Order 13950. He followed with orders launching a "whole-of-government equity agenda" that mandated every major federal agency create detailed Equity Action Plans and appoint Chief Diversity Officers. Pursuant to Executive Order 13985 and follow-on orders, nearly every federal agency and entity submitted "Equity Action Plans" to detail the ways that they had furthered DEI's infiltration of the Federal Government. A Government Accountability Office review covering 2017-2022 found that federal agencies had "mixed success" in increasing diversity. While some agencies saw improvement, others stagnated or even lost ground. A 2024 GAO report on the Internal Revenue Service found that while the agency's workforce was more diverse than the national civilian labor force overall, this diversity was heavily concentrated in lower-paying positions with limited advancement potential. After accounting for education, experience, and job classification, employees from underrepresented groups were 9 to 34 percent less likely to be promoted across most pay grades compared to their white counterparts.

Then came January 20, 2025. President Donald Trump issued Executive Order 14151, which seeks to end diversity, equity, and inclusion (DEI) programs—sometimes referred to as diversity, equity, inclusion, and accessibility (DEIA) programs—throughout the federal government. The order states that the Biden Administration "forced illegal and immoral discrimination programs, going by the name 'diversity, equity, and inclusion' (DEI), into virtually all aspects of the Federal Government in areas ranging from airline safety to the military." On January 21, 2025, President Trump issued an additional E.O. titled "Ending Illegal Discrimination and Restoring Merit-Based Opportunity," which revokes numerous DEI-related executive orders issued by previous administrations, including Executive Order 11246 (which had mandated affirmative action for federal contractors since 1965, signed by President Johnson on September 24, 1965, more than two years after the Rev. Martin Luther King Jr. gave his "I Have A Dream" speech). Six GOP presidents (including Trump during his first term) had kept the Johnson executive order in place, while others expanded it through amendments.

E.O. 14151 requires the director of the Office of Management and Budget (OMB), the Attorney General, and the director of OPM to coordinate to terminate all DEI- and DEIA-related mandates, policies, programs, preferences, and activities across the federal government "under whatever name they appear." Agency heads are required to complete a list of actions within 60 days, including terminating all DEI, DEIA, and "environmental justice" offices and positions; all "equity action plans," equity actions, initiatives, or programs, equity-related grants or contracts; and all DEI or DEIA performance requirements for employees, contractors, or grantees. They must provide the OMB director with a list of DEI, DEIA, or "environmental justice" positions, committees, programs, services, activities, budgets, and expenditures; federal contractors who have provided DEI training or materials; and recipients of federal grants related to DEI, DEIA, or "environmental justice" programs, services, or activities. The executive orders direct the Attorney General, in consultation with the heads of relevant federal agencies, to identify private sector companies with "egregious and discriminatory" DEI programs, signaling potential investigations of publicly traded corporations and large non-profits as well as litigation and potential "regulatory action and sub-regulatory guidance" impacting the private sector. Going forward, federal contracts and grants must include terms requiring the contractual counterparty or grant recipient to agree that its compliance with all applicable Federal anti-discrimination laws is material to the government's payment decisions, and requiring such counterparty or recipient to certify that it does not operate any programs promoting DEI that violate any applicable Federal anti-discrimination laws.

In a memorandum issued shortly after the January 20 DEI Order, the Office of Personnel Management directed the heads of all federal agencies to close any federal agency DEIA offices, place the employees of those agencies on paid leave, take down any outward facing media of any such DEIA offices, and "terminate any DEIA-related contractors" by January 22, 2025. Additionally, federal agencies are directed to submit written plans "for executing a reduction-in-force action regarding the employees who work in a DEIA office" by January 31, 2025. An OPM spokesperson explicitly linked this action to the administration's executive orders aimed at shuttering DEIA efforts government-wide. In spring 2025, the Trump administration's Office of Personnel Management removed public-facing race and ethnicity data filters from FedScope, the government's primary online portal for federal workforce statistics. The removal made it significantly more difficult for the public, researchers, and policymakers to track demographic trends and assess workforce policy impacts.

The intellectual framework for this movement came from conservative organizations like the Heritage Foundation and Manhattan Institute. Christopher Rufo, a senior fellow at the Manhattan Institute, became the public face of this movement. A documentary filmmaker turned policy activist, Rufo had spent years investigating what he saw as left-wing ideological capture of American institutions. He was remarkably candid about his tactics: "We will eventually turn [critical race theory] toxic, as we put all of the various cultural insanities under that brand category. The goal is to have the public read something crazy in the newspaper and immediately think 'critical race theory.'" His strategy worked with devastating effectiveness. Rufo's reporting on concepts like "white privilege" and "systemic racism" being taught to federal employees directly influenced President Trump and served as the primary catalyst for the executive order.

This political environment (specifically the mandate to terminate "equity-related grants or contracts" and provide lists of grant recipients related to DEI programs) created the direct operational requirement that HHS needed to fulfill. The administration framed diversity, equity, and inclusion initiatives as "anti-American propaganda" that promoted division rather than unity. The Trump administration's war on DEI wasn't fought with armies or sanctions. It was fought with definitions, legal language, and the power of the federal purse. By March 2025, HHS had turned to Palantir Technologies and the startup Credal AI to automate the compliance screening that these executive orders demanded.

Reactions: Backlash and Defense

The WIRED investigation that revealed the HHS program triggered pushback from researchers, civil society groups, and former federal employees. Neither Palantir nor HHS has publicly announced that the company's software was being used for ideological screening, and HHS, ACF, Palantir, and Credal AI did not return WIRED's requests for comment. That silence has become its own story: a $6.2 million, five-year contract operating without public procurement debate or congressional oversight.

Inside Palantir, employees have questioned on Slack whether the company can put pressure on ICE and pushed for more transparency on its work with enforcement agencies, according to WIRED's reporting. Palantir's leadership has consistently defended its work for ICE, framing it as mission-critical support for law enforcement. But the pivot from border enforcement to grant-language policing marks a new category of domestic political deployment.

The National Science Foundation's parallel flagging of terms like "female," "inclusion," "systemic," and "underrepresented" suggests the filtering logic has already metastasized beyond HHS.

The administration's justification rests on two pillars: legal compliance and fiscal discipline. Executive Order 14151 demands an end to any policies, programs, contracts, or grants that mention or concern DEIA, DEI, "equity," or "environmental justice." Executive Order 14168 defines "sex" as an "immutable biological classification" and declares "gender ideology" and "gender identity" to be "false" and "disconnected from biological reality." It explicitly states that no federal funds can be used "to promote gender ideology" and that "each agency shall assess grant conditions and grantee preferences and ensure grant funds do not promote gender ideology."

Supporters argue this refocuses resources on proven health outcomes. A 2025 HHS report on gender dysphoria found no solid evidence supporting sex-denying medical interventions for minors, citing potential long-term harms like infertility. By weeding out these elements, the administration claims to save taxpayers from funding unverified experiments, redirecting money to tangible needs like family support and child safety. Automated auditing, they say, has streamlined compliance, allowing human reviewers to focus on final decisions.

Critics counter that the automation itself is the problem. When an LLM flags "female" or "inclusion" as noncompliant, the chilling effect is immediate and self-reinforcing: over 1,000 nonprofit organizations rewrote their mission statements last year to remove any language believed to be adjacent to DEI, seemingly to avoid losing federal funding. The National Center for Missing and Exploited Children and RAINN removed transgender references despite data showing higher victimization rates among trans individuals. CDC retracted studies involving "LGBT" or "transgender" terms. SAMHSA eliminated services for LGBTQ+ youth in July 2025. These aren't hypothetical harms — they're documented behavioral changes across the research and service-delivery ecosystem.

AI ethicists tracking the deployment say the Credal-Palantir stack represents a governance layer repurposed for political enforcement. Credal's platform was built to enforce permissions, guardrails, and instructions between AI assistants and company data, essentially an "Okta for AI," as founder Jack Fischer described it. In the HHS context, those guardrails have been coded with the administration's ideological definitions. The precedent matters: a startup's enterprise security tool, adopted non-competitively by a federal agency, is now automating the interpretation of executive orders that redefine legal categories like "sex" and "gender ideology" without regulatory process or judicial review.

The through-line from Palantir's ICE work (FALCON Search & Analysis, the ELITE targeting app, $30 million added in April for "near real-time visibility" on self-deportations) to this HHS grant screening is architectural. The same data-integration infrastructure that builds dossiers on immigrants now reads grant abstracts for forbidden words. Palantir earned more than $35 million from HHS alone in the first year of Trump's second term, and over $1 billion in net federal payments across all agencies. Credal, founded by Palantir alumni Jack Fischer and Ravin Thambapillai, secured about $750,000 for its generative AI platform from ACF. The vendor relationship is tight; the oversight is absent.

What happens next depends on whether Congress, courts, or internal whistleblowers force disclosure of the prompt logic, the training data, and the human-in-the-loop procedures — if any exist. Until then, researchers submit applications into a black box that may reject them for using the word "systemic."

Implications for AI Governance and Ethics

The HHS deployment marks the first documented case of a U.S. federal agency using a commercial AI governance platform to enforce political ideology on grant applicants. That precedent matters because the technical architecture (Credal's governance layer sitting atop Palantir's Foundry infrastructure) was designed for enterprise data security, not ideological screening. Repurposing it for content-based filtering of scientific proposals converts a tool built to prevent data exfiltration into an automated political gatekeeper. As the Stanford Law School analysis of AI in criminal justice warned, "Without governance, AI in criminal justice will be perilous. Good guidelines and standards are necessary, but they are not sufficient. A governance entity would bridge the gap between general principles and messy reality." The same logic applies here: the HHS program operates without a governance entity, public audit trail, or appeal mechanism for flagged applications.

The mechanism of that conversion is instructive. Credal's platform "builds a layer between AI assistants and company data, enforcing permissions, guardrails and instructions." In enterprise settings, those guardrails prevent employees from pasting PII into chat prompts or querying unauthorized databases. At HHS, the guardrails have been rewritten as ideological tripwires: terms associated with DEI frameworks, gender-affirming care, or environmental justice trigger automatic flags. The WIRED investigation confirmed that since March 2025, HHS has used "tools from Palantir and the startup Credal AI to weed out perceived alignment with 'DEI' or 'gender ideology.'" Palantir's AIP Chatbot Studio, which "is built on the same rigorous security model that governs the rest of the Palantir platform" and "grant[s] an LLM access only to what is necessary to complete a task," provides the underlying inference engine. The result is a hybrid system where a startup's governance middleware directs a defense contractor's LLM infrastructure to perform political content classification at scale.

This architecture inherits the accountability deficits that researchers have documented across public-sector AI. The arXiv study of algorithmic governance across U.S. federal, state, and municipal levels found that "algorithmic systems shift the balance of power by transferring decision-making authority from administrative officials to technical tools that function based on their own logic of data analysis and evaluation. This shift leads to distinct accountability structures. One core challenge in algorithmic accountability is the so-called 'black box' problem. Complicating this further is the heavy reliance on private vendors for developing and operating public-sector AI systems. This vendor involvement increases information gaps and limits transparency, independent auditing, and public accountability." At HHS, two private vendors (one a Palantir alumni-founded startup, the other Palantir itself) control the model weights, the prompt engineering, the flagging thresholds, and the training data for the ideological classifier. No public documentation describes the taxonomy of prohibited concepts, the false-positive rate, or the human review process for edge cases.

The CSET analysis of frontier AI governance proposals underscores why that opacity is dangerous. Across five major governance frameworks (from the White House executive order to SB-1047), "proposals lack consensus regarding the techniques that are most effective at mitigating AI risks and harms." The only assumption shared across four of five proposals was "identifying and tracking frontier AI risks before they become harms." The HHS program does the opposite: it deploys an unvalidated ideological classifier into a high-stakes decision pipeline without public risk assessment, incident reporting, or third-party audit. The CSET authors warned that "if these underlying assumptions are not true, then implementation of most proposals will require provisions to upskill key actors, attract and retain talent, or clarify high-level guidance." HHS has done none of those things; the program was awarded as a non-competitive action to Credal, a startup with roughly $7M ARR, expanded to a $6.2M five-year contract.

The political context amplifies the governance failure. The Brookings assessment of federal AI adoption noted that "the growing politicization of large language models (LLMs) represents another challenge that could further fuel public skepticism. A recent executive order threatens federal procurement for models with built-in 'ideological biases or social agendas.' Recognition of political bias in LLMs is not without merit, but it is difficult (if not impossible) to build systems that are truly neutral." The HHS program operationalizes that executive order by outsourcing the definition of "ideological bias" to a private vendor's prompt engineering. Researchers at Chatham House observed that "private companies largely shape the development and deployment of frontier AI systems. Outside of China's state-directed AI labs, the most consequential decisions on AI development happen inside private companies, in opaque and largely unregulated environments. Private companies have commercial incentives to push back against costly governance provisions like transparency requirements, capability limitations or verification processes, particularly when competition rewards secrecy and speed as richly as it does in AI." Credal's incentive is contract retention; Palantir's is platform stickiness. Neither has a fiduciary duty to scientific integrity or civil rights.

The human rights implications are concrete. The Brookings study on overlooked AI harms documented that "algorithmic discrimination is persistent and a well-documented harm, impacting people's quality of life, including in housing, credit, policing, and hiring decisions—well before the rise of AI-generated deepfakes of celebrities online. Vulnerable populations have been and continue to be systemically and disproportionately impacted given the bias of the data fed into these systems. In these less visible use cases, the process to prove algorithmic discrimination is more difficult. It can take a long time and be arduous to enforce, requiring a longitudinal compilation of evidence of discrimination and more time to persuade regulators and lawmakers to act against the harm. Amid the current rollbacks of protections against discrimination and the disparate impact standard, this process has only become more difficult and opaque." Grant applicants flagged by the HHS system face exactly that dynamic: no statutory right to explanation, no transparent appeal, and a chilling effect that discourages research on health disparities, LGBTQ+ health outcomes, or environmental justice, topics that legitimately require the vocabulary the classifier treats as verboten.

The case also illustrates the regulatory capacity gap that Chatham House identified: "Regulatory capacity gaps mean that, even if they wanted to impose constraints on the private companies developing AI, governments lack the computational resources, technical expertise and legal authority to independently evaluate proprietary models or compel disclosure. The scale of private investment far exceeds regulatory capacity. Industry estimates put 2026 hyperscaler capital spending at $527 billion globally, while OpenAI's CEO projected that future frontier models could require $100 billion in capital per training run. By contrast, the EU's AI Act, finalized in 2024 after years of negotiation, allocated just €1 billion for its enforcement and implementation." HHS has no independent technical capacity to audit Credal's classifier; it relies on the vendor's own assurances. The ACUS report on AI in regulatory enforcement noted that international models like Canada's Directive on Automated Decision-Making "require a robust risk-based assessment process with a sliding scale of obligations based on the results," a framework entirely absent from the HHS deployment.

The precedent extends beyond HHS. The arXiv study found that "at the federal level, AI is most often institutionalized as a tool for high-stakes control: supporting surveillance, enforcement, and regulatory oversight. A defining feature of federal AI implementation is its heavy reliance on private-sector technology providers. Case studies show that even high-security agencies depend significantly on commercial cloud services and AI solutions from leading tech firms. This creates a hybrid model of algorithmic governance, where public responsibilities are partly carried out within private digital ecosystems. The result is a more complicated landscape for accountability: independent audits become harder, responsibilities can become fragmented, and it may be difficult to assign blame when algorithmic decisions go wrong." The Credal-Palantir-HHS triangle exemplifies that hybrid model: a startup provides the governance layer, a defense contractor provides the compute and model serving, and the agency provides the legal authority, but no single entity bears full accountability for the ideological filtering outcomes.

Experts argue the solution is not better prompts but structural governance. The Stanford Law School report recommended "specialized AI governance entities" with "formal crisis-response mandates, not just advisory or risk-monitoring roles." The Chatham House paper urged "concentrating legitimate technical authority" in "AI safety/security institutes… positioned in advance as credible public authorities" with "real-time technical tools, such as model shutdown protocols, circuit breakers and patch deployment mechanisms." The Case Studies compendium emphasized that "governance is not a blocker — it is an enabler of safe, sustainable innovation" and listed concrete practices: "mandatory human-in-the-loop checkpoints for high-risk actions," "incident response plans for AI-related failures," "schedule regular third-party audits," "maintain immutable logs of model activity," and "include AI systems in internal compliance audits." None of these exist for the HHS ideological filter.

The stakes are not abstract. When an AI system built for enterprise data loss prevention is repurposed to enforce a political litmus test on scientific grant applications, the boundary between security governance and ideological control dissolves. The technical architecture — MCP servers, agent governance, guardrail layers — is neutral; its deployment is not. As the Nature article on global AI governance concluded, "AI entities are deemed to be operating within legal bounds as long as they are subject to human control, they should not be left to the discretion of states and corporations as they may prioritize their interests over other considerations." The HHS program leaves exactly that discretion to a startup and a defense contractor, with no statutory framework, no public oversight, and no off switch.

The Vendor Landscape: Why HHS Chose Credal and Palantir

Palantir didn't win the HHS account in March 2025. It had owned the infrastructure for years. The National Institutes of Health, the Centers for Disease Control and Prevention, and the Food and Drug Administration were already running Foundry (Palantir's commercial platform) to stitch together pandemic data across silos that were never designed to talk to each other. When the Administration for Children and Families needed an "AI-based" grant review process that could ingest application files, flag language, and route results to program officers, the plumbing was already in place. The $90 million SHARE blanket purchase agreement, announced in 2024, let any HHS agency spin up Foundry task orders without a new competition. ACF simply reached for the tool it already had.

Palantir's core value proposition to the government is connecting databases that were never meant to be connected.

The numbers bear out the entrenchment. Across the full federal portfolio, the company booked over $1 billion in net payments and obligations — up from roughly $808 million the prior year. ICE payments alone jumped from $20.4 million to $81 million. The SHARE BPA, a five-year vehicle with a $90 million ceiling, was awarded with "limited competition" per federal procurement records. Palantir's president of U.S. government, Akash Jain, called it a "long-term, wide-ranging BPA" showing "confidence in our software." Competitors like Snowflake and Databricks sell data platforms; Palantir sells ontology-driven applications that sit on top of classified and unclassified federal data with accreditations that took a decade to accumulate. Replacing it mid-implementation is expensive enough that most agencies simply do not try.

Credal's path was different but no less political. Founded in 2022 by Jack Fischer and Ravin Thambapillai (both Palantir alumni who "bonded over a mutual interest in security and compliance") the startup raised a $4.8 million seed round led by Spark Capital in October 2023. By early 2024, Credal claimed 11 customers, several on six-figure contracts, and said it had handled more than 250,000 LLM calls across roughly 100,000 ingested documents. Its platform sits between users and third-party LLMs (OpenAI, Anthropic), redacting sensitive data, routing prompts to the "most appropriate" model based on cost and policy, and logging every exchange. That governance layer ("enterprise grade security and real time permissions guarantees synced from your source systems out of the box") is what ACF bought.

Credal's award was a "first Prime win" for the startup, a non-competitive action, per OrangeSlices.ai. The Federal Register payment description for the initial $750,000 does not mention DEI or gender ideology. But the WIRED investigation confirmed Credal "helped ACF audit 'existing grants and new grant applications'" using the same AI inventory that flags language for executive-order compliance. The founders' Palantir pedigree is not incidental. Former Palantir employees have moved into key government IT and DOGE roles, then directed contracts back to Palantir. Stephen Miller, architect of the administration's immigration enforcement policy, holds a financial stake in Palantir. The revolving door between the company and the agencies buying its tools is documented, not alleged.

Why not a competitor? Microsoft Purview for Copilot and AWS Bedrock Guardrails offer governance features, but they lack the federal accreditations and the installed base inside HHS. Nightfall AI and Securiti.ai target data loss prevention, not the agent-orchestration layer Credal built. And no other vendor had a pair of Palantir alumni who understood the Foundry ontology, the clearance levels, and the procurement shortcuts. Credal's Data Privacy Framework registration (a U.S.-EU transfer mechanism) checked a compliance box few early-stage startups could.

The result is a two-vendor stack: Palantir provides the data fabric and the contract vehicle; Credal provides the LLM governance layer that turns executive orders into automated flags. Neither contract was openly competed. Both vendors sit inside a procurement ecosystem where sole-source awards are the norm, not the exception. The next administration will inherit the same stack, and the same lack of an audit trail for how the flags were written.

The Human Cost: Chilling Effect on Grant Applicants

The numbers tell part of the story. By April 2025, the National Science Foundation had canceled 1,574 grants — roughly 90 percent flagged as "related to DEI." The National Institutes of Health terminated dozens of active research grants connected to LGBTQ+, gender-identity, or diversity studies. The Department of Education cut over $600 million in teacher-training grants. The USDA quietly removed diversity, equity, and inclusion scoring criteria from 14 grant programs. By year's end, across NSF and NIH alone, nearly $3 billion in grant funds were either frozen or terminated.

But the deeper damage is behavioral. More than 1,000 nonprofit organizations did so last year to remove any language believed to be adjacent to DEI, seemingly in an effort to avoid losing federal funding. The share of successful NSF grants containing DEI language has plummeted, falling to 9 percent in 2025 from 34 percent in 2022. Researchers studying health equity (why certain populations experience worse outcomes) report that their work is being reclassified as social science and deemed outside NIH's core mission.

The mechanism is crude. Early last year, NSF started to flag any research that contained terms associated with DEI (including relatively general terms like "female," "inclusion," "systemic," or "underrepresented") and place it under official review. The Centers for Disease Control and Prevention began retracting or pausing research that mentioned terms like "LGBT," "transsexual," or "nonbinary," and stopped processing any data related to transgender people. At the NEH, DOGE staff used ChatGPT to flag grants as DEI based on keyword pattern-matching, terminating 97 percent of the agency's active portfolio. The AI had no definition to work from. Neither, it appears, did the humans directing it.

Now HHS has formalized that approach. Since March 2025, the Administration for Children and Families has been using tools from Palantir Technologies and the startup Credal AI to screen and audit grants, grant applications, and job descriptions for noncompliance with executive orders targeting "gender ideology" and anything related to DEI. Credal's platform performs that review. ACF staffers ultimately review flagged items during a "final review" stage, but the initial triage is automated.

The result is a compliance nightmare. The certification's scope and vagueness create impossible ambiguity. Does a university's office of student services constitute a "DEI initiative"? Does a medical school's program to recruit physicians for underserved rural communities qualify? Does an engineering department's outreach to community colleges in low-income areas count? The executive orders don't say, the proposed certification doesn't clarify, and the consequences of guessing wrong include loss of all federal funding.

So institutions are preemptively sanitizing. Programs that once described their work in terms of "equity," "diversity," or "inclusion" are restructuring their descriptions around universal access, geographic reach, and outcome metrics. A medical school program that recruited physicians from underrepresented minority backgrounds is now described as addressing "physician shortages in medically underserved areas." A STEM outreach program that targeted students from low-income families is now framed as "expanding the STEM talent pipeline from under-resourced school districts." An engineering program that partnered with HBCUs is now a "multi-institutional collaboration leveraging complementary research strengths." Some universities have renamed offices ("Office of Diversity and Inclusion" becomes "Office of Student Success" or "Office of Access and Opportunity") not because the work changes, but because the label creates unnecessary risk.

Instead of tracking demographic representation directly, institutions are measuring applications from underresourced pipelines, retention rates across socioeconomic categories, and research participation by geographic origin. These metrics capture the same information through a lens that emphasizes institutional performance rather than population characteristics. It is a massive, costly exercise in semantic evasion.

The human toll extends beyond paperwork. Children won't receive high-quality education that Head Start previously would have been able to provide in their first language. Survivors of domestic violence face increased barriers in receiving essential shelter and services. Americans' access to HIV treatment will be limited. Gender-affirming care will be harder to receive. The National Center for Missing and Exploited Children and the Rape, Abuse, & Incest National Network removed any mentions of trans people (who are disproportionately more likely to be victims of sexual abuse) from all of their public materials.

Disruptions to essential federal programs worsen inequality and have an outsized impact on households of color, immigrant families, people with disabilities, and LGBTQ+ individuals, who receive a disproportionate amount of the affected services and programs due to the ongoing systemic inequality and discrimination they face. TRIO programs (which support college access for low-income, first-generation students) face ongoing uncertainty despite Congressional action to preserve their funding in the FY2026 budget. As of March 10, neither TRIO competition had opened.

Organizations that rely on a single federal agency for the majority of their research or program budget face existential risk, regardless of how well they comply with certification requirements. The practical advice from research administration professionals is consistent: comply with the letter of the certification while preserving the substance of your work. Document your decision-making processes rigorously. Build relationships with program officers at federal agencies, who in many cases are sympathetic to grantees navigating an ambiguous compliance environment. And participate in the public comment process; the GSA proposal is not yet final, and substantive comments from the research community may influence its scope.

But legal victories are slow, and the certification requirement could be finalized before courts rule on the underlying executive orders. Institutions that wait for legal clarity before adapting their operations may find themselves without federal funding in the interim. The chilling effect is not theoretical. It is already reshaping what questions get asked, what populations get studied, and what language scientists dare to use.

Working in frontier tech? Zero G Talent tracks the openings: see every open Palantir Technologies role, browse frontier tech jobs, the companies hiring, and the people building the field.