FOIA suits against USCIS in 2026: Pangea, Mukherji v. Miller, and AI disclosure in adjudication

FOIA suits Pangea v. USCIS Mukherji v. Miller Loper Bright 2026

FOIA suits against USCIS 2026 are lawsuits trying to force the agency to disclose exactly how AI is used in adjudicating petitions. Two of them — Pangea Legal Services v. USCIS and Refugees International v. USCIS — demand documents: prompts, training data, vendor contracts. The third case, Mukherji v. Miller, stands apart: it is not a records request but a direct challenge to a specific EB-1A denial — and in January 2026 the plaintiff won. In this article I go through all three in order and explain what they actually give petitioners right now.

This is one of four articles in the USCIS and AI cluster. This one covers courts only. A general overview and verdict are in the main article, and specific USCIS AI systems are analyzed separately.

Contents

Why bring FOIA suits against USCIS at all

?

Which lawsuits compel USCIS to disclose how exactly AI is used in decisions about my petition?

FOIA is the Freedom of Information Act, 5 U.S.C. 552. Under it any person or organization can request records from a federal agency, and if the agency refuses or drags its feet — sue. With AI at USCIS the logic is: an organization asks “produce all records about the use of AI in adjudication,” USCIS responds with refusals or heavily redacted records, and then the case goes to federal court, which decides what and in what form must be released.

Why this matters to a petitioner is best explained by a former USCIS asylum officer. The problem is not that AI exists, but that its workings are invisible to the applicant.

Joshua Perez Garcia, former USCIS Asylum Officer (6 years), ILW.com, May 11, 2026
“The officer has seen what the system surfaced; the applicant and attorney have not. The NOID does not disclose it. The denial letter does not disclose it.”

The officer saw what the system highlighted; the applicant and attorney did not. The NOID (notice of intent to deny) does not disclose it. The denial letter neither. It is this invisibility that the FOIA suits seek to uncover. Source: ILW.com.

Below I analyze each of the three cases according to one scheme: who is the plaintiff, what they demand, current status, and what has already been disclosed. I’ll start with the largest.

How a FOIA lawsuit works, in brief

This is not a criminal or ordinary civil action but a dispute over access to documents. First the agency searches for responsive records and produces them in batches (rolling production), withholding parts via redactions. Then the parties submit Joint Status Reports to the court every 60–90 days about what has been produced and what remains. The key document is the Vaughn Index: a formal list of everything the agency refused to release, with justification for each withholding. Until the Vaughn Index is issued, the case is far from a merits decision. And FOIA litigation typically takes years.

Pangea Legal Services v. USCIS

The main and most advanced case. It was filed by a coalition of civil-rights organizations, and USCIS has already produced some documents in this one.

Pangea Legal Services v. United States Citizenship and Immigration Services
1:24-cv-02809-ACR D.D.C. Judge Ana C. Reyes filed October 3, 2024
Plaintiffs are Pangea Legal Services, Mijente Support Committee, and Just Futures Law (the latter appears both as plaintiff and counsel). The suit demands that USCIS disclose 16 categories of records about the use of AI and ML in adjudication: lists of AI tools, training data, vendor contracts, PIAs and Algorithmic Impact Assessments, accuracy metrics, bias monitoring, notice and opt‑out policies.
Case number: 1:24-cv-02809-ACR (D.D.C.), Judge Ana C. Reyes. Plaintiffs’ filings: Just Futures Law - DHS AI FOIA.

Pangea Legal Services is a California non-profit defending immigrants. In summer 2024 it, together with Mijente and Just Futures Law, filed FOIA requests to USCIS, DHS, and ICE — asking the agencies to disclose how they use AI. They received no substantive response, and on October 3, 2024 they went to court. The case was assigned to Judge Ana C. Reyes — an important detail, because she also presides over the second suit (Refugees International). Her ruling in one case will almost certainly affect the other.

To be clear why this suit matters: the plaintiffs are not asking for abstract commentary but concrete materials. The complaint lists AI in DHS not as an abstraction but as contracts with the same companies whose products we use every day.

Complaint in Pangea v. USCIS, October 3, 2024 (note 1)
“DHS is contracting with OpenAI, Anthropic, Meta, Microsoft, Google and Amazon to build and host its AI technologies.”

DHS contracts with OpenAI, Anthropic, Meta, Microsoft, Google and Amazon to build and host its AI technologies. So the same Anthropic that makes Claude is officially one of DHS’s contractors. This directly connects to the Claude PDF Intake discussed in the article about AI systems. Source: complaint text, Just Futures Law.

And the plaintiffs demand disclosure not of “all AI” but of specific systems — they name them. That matters because it shows what civil-rights groups consider most risky.

  • ATA (Automated Text Analytics) — a system that scans application texts for similarity.
  • GenAI — USCIS’s generative AI tools (the same class that includes Claude PDF Intake).
  • FDNS-DS NextGen — an updated fraud-detection and national-security system.
  • "Predicted to Naturalize" — a tool that tries to predict who will obtain citizenship.

How do we even know about these tools if USCIS doesn’t disclose them properly? The plaintiffs rely on their own investigation — the report “Automating Deportation” (Just Futures Law and Mijente, June 2024). It is the most detailed public account of how DHS and USCIS deploy AI; if you want the primary source in full — it’s here.

Now let’s see what has actually been produced so far — three facts from the most recent Joint Status Report (the report parties file with the court about disclosure progress).

1
USCIS produced its part

USCIS production was completed June 30, 2025. Just Futures Law has already received and is reviewing part of the documents. But prompts and system instructions do not yet appear among what was produced.

2
DHS is dragging

DHS identified ~3,500 pages for the narrowed request, but as of March 31, 2026 produced only about 159 of them. The rest were sent for consultation with other agencies. The Vaughn Index has not been issued.

3
The shutdown froze everything

Since February 14, 2026 — a lapse in appropriations. DHS FOIA staff are on unpaid leave and production stopped. Plaintiffs challenged the legality of stopping productions during the shutdown.

The third point needs separate explanation because it explains why everything stalled now. On February 14, 2026 the US experienced a government shutdown — specifically, a lapse in appropriations for DHS (the department that includes USCIS). This is real: DHS funding lapsed and employees, including those handling FOIA, were placed on unpaid leave. In plain terms — the people who were supposed to produce documents for the suit were temporarily sent home without pay, and production halted. That is what the parties told the court in the Joint Status Report of March 31, 2026. The DHS shutdown is publicly confirmed (Government Executive), and the docket can be tracked on PACER (PacerMonitor).

Why Judge Ana C. Reyes matters

Reyes is a federal judge, sworn in February 2023, known for detailed opinions in FOIA cases. If she finds that AI prompts and model outputs are not covered by the deliberative‑process privilege, that will set a precedent for all FOIA suits against DHS. But to date she has not ruled on the merits: no Vaughn Index, no substantive ruling. It’s realistic to expect prompts to be disclosed no earlier than 2027–2028.

Refugees International v. USCIS

The second case — almost a mirror of the first, but about another system and before the same judge. If Pangea concerns adjudication generally, this one focuses on a specific AI tool for asylum.

Refugees International v. United States Citizenship and Immigration Services
1:24-cv-03559 D.D.C. Judge Ana C. Reyes filed December 20, 2024
Plaintiff is Refugees International, a Washington non‑profit. Counsel are the Harvard Immigration & Refugee Clinical Program and Jenner & Block LLP. The suit targets a specific tool: ATA (Asylum Text Analytics) runs on the AI engine Pangaea Text (described in PIA‑085) and searches asylum application texts (Form I‑589) for "plagiarism‑fraud" — i.e., matching text fragments across cases. The suit demands disclosure of guidelines, training materials, a sample ATA report, and 20 statistical data points: how many cases were reviewed and flagged, broken down by nationality and by office.
Case number: 1:24-cv-03559 (D.D.C.), Judge Ana C. Reyes. Plaintiff statement: Refugees International.

The legal question here is the same as in Pangea: are AI prompts and outputs records that must be disclosed under FOIA. That both suits proceed in parallel before the same judge is no accident but a tactic: one precedential opinion by Reyes would work for both cases.

But this case has a detail important for anyone filing in English as a non‑native speaker. Plaintiffs directly ask: what the system treats as a “common language pattern” (i.e., a sign of copied, template text) may actually be an attempt by someone with poor English to fill out the form as best they can. Their main question to USCIS is whether such a person is given a chance to explain before their application is flagged as suspicious for fraud.

Yael Schacher, Refugees International, statement when filing the suit, December 20, 2024
“What looks like a common language pattern may be an attempt by a pro se asylum seeker to fill out an application to the best of their ability in a foreign language. Do asylum officers give them a chance to explain this when ATA flags their application as fraudulent?”

What appears as a common language pattern may be an attempt by a person without counsel to fill out an application as best they can in a foreign language. Do officers give them a chance to explain this when ATA flags their application as fraud? Source: Refugees International.

A detail that shows how DHS treats its own inventory

ATA was added to the DHS AI Inventory in December 2024, removed after 11 days, then re‑added and marked inactive. This shows the inventory’s classifications are mutable and the agency can rewrite them under pressure. Another reason plaintiffs need primary documents rather than just the inventory.

Pressure is not only through courts. In September 2024 a coalition of 142 organizations (including EFF, EPIC, AILA, Mijente) sent DHS Secretary Mayorkas a letter with a direct demand.

Coalition letter from 142 organizations to DHS Secretary Mayorkas, September 4, 2024
“Cancel DHS Use of AI Technologies for Immigration Enforcement and Adjudication by December 1, 2024.”

Cancel DHS use of AI technologies for immigration enforcement and adjudication by December 1, 2024. The demand was to suspend AI tools that do not comply with OMB Memorandum M‑24‑10. Confirmation: Just Futures Law.

Mukherji v. Miller — the case the plaintiff won

This is the most important of the three and the only one with a result favorable to the petitioner. Mukherji is not a FOIA suit about documents but a direct APA challenge: the plaintiff contested a specific denial of her petition, alleging USCIS applied an unlawful standard. The court agreed.

Anahita Mukherji v. Loren K. Miller
4:24-cv-03170 D. Nebraska Senior Judge Joseph F. Bataillon decision January 28, 2026
Plaintiff Anahita Mukherji, an Indian journalist, filed an EB‑1A petition. She met 5 of the 10 regulatory EB‑1A criteria (minimum is 3), but USCIS still denied at the final merits determination stage. The court found the agency’s application of that standard arbitrary and capricious, relying on Loper Bright, and ordered USCIS to approve the petition outright. DOJ has appealed to the Eighth Circuit (No. 26‑1578).
Case number: 4:24-cv-03170 (D. Neb.), Judge Joseph F. Bataillon. Case analysis: Cyrus Mehta blog.

The rarest aspect of this decision is not the vacatur itself but that the court ordered direct approval rather than remanding to USCIS for reconsideration. Reddy Neumann Brown highlighted this in their case analysis.

Reddy Neumann Brown PC, analysis of Mukherji v. Miller, January 2026
“the court ordered USCIS to approve the applicant's EB-1A petition outright, rather than send the case back”

The court ordered USCIS to approve the petitioner’s EB‑1A petition outright, rather than send the case back for reconsideration. This is extremely rare: courts usually remand with instructions. “Outright approval” means the judge found the denial so arbitrary that she did not trust USCIS to correct it.

Why Mukherji matters more than it seems

FOIA suits decide only disclosure. In Mukherji the court — for the first time after Loper Bright — directly told USCIS: your EB‑1A evaluation standard itself is arbitrary and capricious. That gives petitioners who receive opaque denials a concrete legal basis for an APA challenge — without waiting for Pangea to disclose prompts. The case is currently at the district level, not circuit, and DOJ is appealing — but the precedent is set.

Loper Bright and the end of Chevron deference

To understand why Mukherji became possible, you need one Supreme Court decision. It removed the legal shield USCIS had relied on for decades.

Loper Bright Enterprises v. Raimondo
No. 22-451 SCOTUS 2024 Chief Justice Roberts
“Courts must exercise their independent judgment in deciding whether an agency has acted within its statutory authority... courts need not and under the APA may not defer to an agency interpretation of the law simply because a statute is ambiguous.”
Source: SCOTUS opinion No. 22‑451 (June 2024).

Before Loper Bright the Chevron doctrine (1984) applied: if a statute was ambiguous, courts were required to defer to the agency’s interpretation. After Loper Bright, courts decide for themselves. Applied to AI this changes the balance: previously USCIS could say “AI only assists the officer, that’s our interpretation of the INA” — and courts would have to accept it. Now a court can say: “your interpretation is wrong; AI in practice makes decisions, which violates the APA.” Mukherji is the first case where this worked in practice.

That this is not an isolated problem but systemic is noted not only by individual attorneys but by a professional association — quoted by former Immigration Judge Robert Ratliff.

AILA, cited in BMD client alert (Robert Ratliff), February 10, 2026
“The American Immigration Lawyers Association (AILA) has documented patterns of inconsistent adjudication where reasoning in the record does not always align with submitted evidence.”

AILA has documented patterns of inconsistent adjudication where the reasoning in the record does not always align with submitted evidence. This is now the official position of the association of immigration lawyers, not just one or two attorneys. Source: BMD Law.

What these cases give a petitioner

There is no direct ruling yet that by itself will protect your petition. But the three litigations already provide three tools of varying maturity — from “works right now” to “wait until 2027.”

1

Loper Bright — works already

No new decision required. Any attorney brief challenging USCIS AI policy now carries different weight than in 2023: courts are no longer bound to defer to the agency’s interpretation. This applies already at the RFE stage and in motions to reopen.

When: already in effect.
2

Mukherji — a precedent overturning an EB‑1A denial

If the Eighth Circuit upholds Judge Bataillon’s decision, petitioners will gain a direct argument: an opaque final merits determination can be attacked under the APA. You can already cite this decision in a motion to reopen after a denial.

When: appeal in the Eighth Circuit; opinion expected 2026–2027.
3

Pangea and Refugees International — disclosure of prompts

If Judge Reyes orders USCIS to produce prompts and training data, attorneys will see what exactly AI systems are trained on and what they look for. This is the most powerful tool — but also the furthest away: the shutdown froze production and the Vaughn Index has not been issued.

When: realistically 2027–2028.

A practical step available right now

You don’t have to wait for systemic disclosure. Any petitioner can file a personal FOIA request for their A‑file phrased "including all AI‑generated summaries, classifications, flags, and alerts in my case." USCIS typically produces A‑files in 30–90 days — this will get your specific documents faster than Pangea will yield systemwide materials.

Conclusions

1
Pangea and Refugees International — FOIA for documents

Both are before Judge Ana C. Reyes. The goal is to force USCIS to disclose how AI is used in adjudication. USCIS produced its part, but prompts are not yet included, and the government shutdown since February 14, 2026 froze further production.

2
Mukherji v. Miller — APA challenge, plaintiff prevailed

The federal court in Nebraska (Judge Bataillon, January 28, 2026) found the final merits determination on EB‑1A arbitrary and capricious and ordered approval of the petition outright, relying on Loper Bright. DOJ is appealing to the Eighth Circuit.

3
Loper Bright changed the balance already

The end of Chevron deference means courts now consider arguments against USCIS AI policy independently, without mandatory deference to the agency. This is already effective at the RFE stage.

4
Systemic disclosure of prompts is still far off

All FOIA litigations are at an early stage: the Vaughn Index hasn’t been issued, there’s no ruling on the merits, and the shutdown froze production. But a personal FOIA request for your A‑file is available to petitioners now.

Related articles in this cluster

Disclaimer. I am not a licensed immigration attorney or a litigation specialist. Case statuses are stated based on the latest publicly available docket entries and publications as of writing. Before using this in your own petition strategy check the current status on PACER or CourtListener. If any link stops opening — write and I’ll fix it.

Author: Egor Akimov, eliteskillset.com. Published 2026-06-02, updated based on the original analysis May 25, 2026.