Practice Areas
AI for Immigration Lawyers: Forms, Evidence Bundles and the Hallucination Risk
AI immigration lawyers use models for forms, evidence bundles, translation drafts and multilingual client updates. Country research and case law still need verification.
On this page
- The practice area that adopted fastest and governed least
- UK v SSHD and what "public domain" means for a client letter
- Four jobs AI does well in an immigration practice
- Country-condition research: the trap
- A safe immigration workflow
- Translation, interpreters and consent
- Tribunal rules and the other side's AI
- Where to start this month
Immigration lawyers use AI more than any other practice group and govern it less: 82% of practitioners use it and 40% use it daily, while only 8% of firms have a written AI policy. The work it does well is extraction, bundle indexing, chronology, translation drafts and client updates in many languages. The work it does badly is country-condition research and case law, and an Upper Tribunal has already said that a client letter pasted into a public AI tool is in the public domain.
The practice area that adopted fastest and governed least
The numbers come from DocketWise's 2026 survey data (21 August 2026): 82% of immigration practitioners use AI against 69% across all practice areas, 40% use it every day, and among the 110 respondents whose firms use AI for immigration casework, 58% use it for document drafting, the most common application. Only 8% of firms have a written AI policy. Around 40% of immigration lawyers cite unauthorised-practice concerns about AI tools, which is a reasonable worry in a field where notarios and visa consultants already operate at the edge of the law.
That combination, heavy use and thin governance, is what produces disciplinary cases. The Australian federal circuit court referred an immigration advocate to the regulator in Valu v Minister for Immigration (No 2) [2025] FedCFamC2G 95 (31 January 2025) after fake AI-generated authorities appeared in submissions. The pattern is the same as in every other practice area, but immigration adds volume, deadlines, vulnerable clients and languages, all of which push lawyers towards whatever tool is open in the next browser tab. Before any of the workflows below, a firm needs the one-page policy described in how to write a law firm AI policy.
UK v SSHD and what "public domain" means for a client letter
The Solicitors Regulation Authority's warning notice on the misuse of AI (17 August 2026) records 42 AI-misuse reports between July 2025 and July 2026 and cites three cases, one of them from the immigration tribunal: UK v SSHD [2026] UKUT 81 (IAC), in which the Upper Tribunal warned that pasting client letters into public AI tools puts them "in the public domain". The notice tells solicitors they must understand a tool "well enough to review its output critically", and the Law Society's press release accompanied it with the same message. The detail is in the SRA's 2026 warning notice.
The tribunal's phrase is not rhetoric. Consumer ChatGPT (Free, Plus and Pro) trains on conversations unless the user opts out; consumer Gemini keeps activity for 18 months by default and lets human reviewers read a subset of chats, with Google's own advice being not to enter confidential information; consumer Claude asks users to choose, with five-year retention if training is on. For an asylum seeker's witness statement, a trafficking victim's medical report or a sponsor's financial records, none of those is acceptable. Business, Team, Enterprise and API tiers of the same vendors do not train by default, and that is where immigration client material belongs, as set out in the confidentiality checklist. The US view of the same problem, that a client's own chatbot use is not privileged, comes from United States v. Heppner (S.D.N.Y., February 2026) and is discussed in privilege and AI.
Four jobs AI does well in an immigration practice
| Job | What the model does | Tool type | What the lawyer checks |
|---|---|---|---|
| Form population and data extraction | Reads passports, prior applications, payslips, tenancy agreements and letters; returns every field with a source page | Business-tier model with file upload; OCR pass first for scans | Every field against the document; the current form version |
| Evidence bundle index and chronology | Indexes 300 pages of evidence, dates each document, builds a chronology with page references, lists gaps | Long-context model (1M tokens holds a full bundle) or Gemini Notebook for a closed universe | Dates, names, gaps the model missed |
| Translation first drafts | Translates supporting documents and client correspondence for triage and review | Business-tier model; certified translator for anything filed | Terminology; certification requirements |
| Multilingual client communication | Drafts status updates and requests for documents in the client's language at a set reading level | Business-tier model with a template and tone instructions | Accuracy; nothing that reads as legal advice the lawyer has not given |
The extraction task rewards a dedicated OCR pass before upload. Claude and ChatGPT read scans by vision, but for passport numbers, dates and reference numbers a text layer is safer, and Claude is the model most likely to say "illegible" rather than invent a digit; the method is in extracting data from PDFs and scans. Translation for filing is a different question from translation for reading, and the certification rules of each tribunal are unchanged; see AI legal translation.
Country-condition research: the trap
Country evidence is where immigration lawyers are most tempted and most exposed. A model asked about conditions in a given country will produce a fluent summary with citations to reports that may not exist, may be years out of date, or may say the opposite of what is claimed. The AI Law Librarians review of the science (February 2026) reports a 2025 study in which frontier models hallucinated on 45% of Los Angeles questions, 55% of London questions and 61% of Sydney questions. Search does not cure it: a University of Pennsylvania study (3 April 2026) found search-augmented models and deep-research agents fabricating 3–13% of the URLs they cited, with agents hallucinating more because they cite more.
The safe method has three parts. First, the lawyer assembles the sources: the official country-information notes, UNHCR and NGO reports, expert reports already on file. Second, the model works only from those uploads, with an instruction that it must quote the passage and give the page for every proposition. Third, every quoted passage is opened and read before it is relied on. A general model is never the source of a country fact, just as it is never the source of a case; the citation-checking protocol applies to country evidence with the same force. For open-web research, deep research tools explains which products show their sources and which do not.
A safe immigration workflow
This sequence is built for a firm running dozens of matters in parallel. It assumes a business-tier model or a legal platform, never a consumer chatbot.
- Intake in the client's language. A structured intake form, or an intake agent, collects the facts and documents; anything the client types is stored in the matter, not in a chat history. The design is in AI client intake.
- OCR and extraction. Scanned identity documents and supporting evidence go through OCR; the model returns a field table with source pages; a caseworker checks each field.
- Bundle index and chronology. The model indexes the evidence and produces a dated chronology with page references and a list of gaps and inconsistencies (the prompt below).
- Country evidence from uploaded sources only. The lawyer selects the reports; the model maps propositions to passages; the lawyer reads every passage.
- Drafting from the record. Cover letters, representations and witness-statement skeletons are drafted from the chronology and the client's own account, with placeholders for anything the model cannot source.
- Translation and certification. First drafts for review; certified translations for filing where the tribunal requires them.
- Verification and sign-off. Every authority and every country citation opened by the signing lawyer; the review recorded on the file.
- Client update. A status note in the client's language, from a template, reviewed before sending.
You are assisting an immigration lawyer with an evidence bundle.
Work only from the uploaded documents. Do not add facts, dates or
country information from outside the bundle.
Produce:
1. An index: one row per document with document type, author or
issuer, date, language, page range, and a one-line description.
2. A chronology of events described in the documents, in date order,
each entry citing the document and page it comes from. Where two
documents give different dates for the same event, list both and
flag the conflict.
3. A gap list: periods with no documentary evidence, documents referred
to but not present, and undated or illegible documents.
4. A question list for the client, in plain language, covering every
gap and conflict above.
Quote exactly. If a passage is illegible or a date is unclear, say so.
Do not assess credibility or the merits of the claim.
Module 3 of the AI Lab for Lawyers course works through multimodal file handling of exactly this kind, scanned identity documents, mixed-language bundles and long PDFs, using browser tools only.
Translation, interpreters and consent
Two habits protect a high-volume practice. The first is a written rule on which translations are for reading and which are for filing. A model's translation of a birth certificate lets a caseworker check a date; it does not replace the certified translation the tribunal requires, and a model's rendering of a legal term of art in a foreign judgment needs a human translator's eye before it is relied on.
The second is consent for recording and transcription. Immigration interviews involve interpreters, family members and, often, distressed clients. The New York City Bar's Formal Opinion 2025-6 (22 December 2025) on AI recording of client meetings, and Formal Opinion 2026-2 (5 August 2026) on non-client calls, require the consent of all parties and treat not recording as the default. Those are New York opinions, but the practice they describe, consent first and transcription off by default, is the right one everywhere. The setup is in AI meeting notes for lawyers.
Tribunal rules and the other side's AI
Rules differ by forum. In the United States, more than 30 of 94 federal district courts had AI disclosure or certification requirements as of April 2026, and, counted separately, more than 300 individual federal judges have AI standing orders; Illinois, by contrast, requires no disclosure and holds the filer accountable. The map is in court standing orders on AI. In England and Wales the judiciary's AI guidance (updated 31 October 2025) tells judges not to enter private information into public AI tools and makes them responsible for everything in their name, and it notes that AI tools "may be the only source of advice or assistance some litigants receive".
That last point matters in immigration more than anywhere. Of the 2,041 decisions in Damien Charlotin's hallucination database (14 September 2026), 1,174 involve litigants in person. Unrepresented appellants now file AI-drafted grounds, and on 18 September 2026 the Law Gazette reported a litigant filing a 300-page skeleton argument "generated entirely by ChatGPT". An immigration lawyer opposing such a filing checks the other side's authorities as carefully as their own; in Noland v. Land of the Free (Cal. Ct. App., 12 September 2025) the court criticised respondent's counsel for failing to notice the appellant's fabricated citations.
Where to start this month
- Write the policy: which tiers are permitted, which client categories are excluded from any AI use (protection claims, minors, trafficking), who signs off on filings. Eight per cent of firms have one; be in that group.
- Move every practitioner to a business-tier account under a data-processing agreement and switch off consumer tools on firm devices.
- Adopt the bundle prompt above as a team template, with the "open every passage" rule for country evidence written into the file-review checklist.
- Agree the translation rule: model drafts for reading, certified translations for filing.
- Log the training, since Article 4 of the EU AI Act asks deployers to support AI literacy and national authorities have supervised it since 2 August 2026; see AI literacy for law firms.
The efficiency gains in immigration practice are real and the case for using the tools is strong. The cases that end in the disciplinary record are not about efficiency; they are about the step between the model's output and the lawyer's signature.
Frequently asked questions
Can AI fill in immigration forms?
It can extract the data a form needs from passports, prior applications, payslips and letters, and populate a draft, but it cannot be trusted to know the current version of a form or the current rules. Run scanned documents through OCR first, have the model return every field with its source page, and check each field against the document before anything is submitted.
Is it safe to use ChatGPT for immigration client work?
Only in a business or enterprise tier that does not train on inputs and under a data-processing agreement, and never for asylum, protection or vulnerable-client material in a consumer tier. In UK v SSHD [2026] UKUT 81 (IAC) the Upper Tribunal said that pasting client letters into public AI tools puts them in the public domain, and the SRA cited that case in its August 2026 warning notice.
Can AI do country-condition research for an asylum claim?
Not from its own memory. Models invent reports, misdate them and misattribute findings, and even search-enabled research agents fabricate a share of the URLs they cite. Upload the official country-information reports and expert evidence yourself, instruct the model to work only from those files, and open every passage it relies on before it goes into a witness statement or skeleton.
Do I have to tell the tribunal that I used AI?
It depends on the forum. More than 30 US federal district courts and 300 or more individual federal judges have AI disclosure or certification requirements, while the UK tribunals have no general disclosure rule but hold the representative responsible for everything filed. Check the standing orders and practice directions for your tribunal and treat verification, not disclosure, as the constant.
Put this into practice
AI Lab for Lawyers: a live, hands-on course for practising lawyers
4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.
Related reading
Regulation & Ethics
The SRA's 2026 Warning Notice on AI Misuse: What UK Solicitors Must Know
The SRA AI warning notice of 17 August 2026 followed 42 misuse reports in a year. Solicitors must understand a tool well enough to review its output critically.
8 min read
Use Cases
AI Legal Translation: Accuracy, Confidentiality and Workflow
AI legal translation gives reliable working drafts of contracts and pleadings when a glossary, system prompt and bilingual review surround it. German-English example.
10 min read
Playbooks
Extract Data from PDFs and Scans with AI: Contracts, Invoices and Court Files
How lawyers extract data from PDFs and scans with AI: OCR first, a fixed schema, a full lease-extraction prompt, batch runs in Cowork or n8n, and sampling.
11 min read
Regulation & Ethics
Attorney–Client Privilege and AI: Heppner, Warner, the NYT Logs and Public Chatbots
Consumer chatbot memos are not privileged. US v. Heppner, Warner v. Gilbarco and the 20-million-log OpenAI order, plus rules that keep privilege intact.
10 min read
Use Cases
AI Client Intake: Agents That Qualify, Conflict-Check and Onboard
AI client intake agents qualify leads, run conflict pre-checks, collect KYC documents and draft engagement letters. The flow, the fields and the guardrails.
8 min read
Regulation & Ethics
Court Standing Orders on AI: Where You Must Disclose or Certify
Court AI disclosure rules in 2026: over 30 of 94 US federal districts and 300+ judges require certification, two Ohio judges ban it, Illinois requires none.
10 min read