Can Immigration Lawyers Use ChatGPT? Evidence Letters and Redaction

Immigration lawyers can use ChatGPT for drafting support and evidence letters if the firm uses a business plan with no-training defaults, redacts every identifying detail before pasting, and has a lawyer verify and sign the result. Names, dates of birth, passport and case numbers, addresses and sensitive history stay out of the chat.

Those rules belong to the firm owner or managing partner, not to whoever happens to be typing, because the risk is firm-wide: one paralegal pasting a full asylum interview into a personal account is a confidentiality incident for the whole practice. If you want prompts for each stage, best AI for immigration lawyers, stage by stage covers the workflow; the decisions here are the plan, the redaction list, who may do what and how drafts are reviewed. This is guidance, not legal advice: check your own professional rules.

Follow me on Instagram@sagnikteaches

Decide the plan before you write a single rule

Rules are only as strong as the product underneath them. Immigration files hold identity documents, family histories, sometimes persecution or medical details, so the plan decision comes first.

Connect on LinkedInSagnik Bhattacharya

ChatGPT Business lists at $20 per user a month billed annually or $25 monthly on OpenAI's business pricing page, for teams of 2 to 200, so a solo practitioner pays for two seats or stays on a personal plan with strict placeholders. Check that page before buying, because seat types and prices change. A third point matters more than price: no plan removes your duty of confidentiality. A business plan reduces the chance that content is used for training and gives you admin controls. It does not make pasting a client's full file harmless.

Subscribe on YouTube@sagnikteaches

Two operational facts to plan around. Members cannot export their own workspace data (they have to ask the workspace owner), and by default owners and admins cannot see members' private chats; removing a member ends their access at once. So do not let ChatGPT become the place where a case record lives. Treat it as a scratchpad and keep the record in your case management system. For the free-plan question in more depth, read free ChatGPT for lawyers: where the free plan stops being safe.

The redaction list: what never goes into the chat

Write the list as a card people keep on the desk. It must be specific enough that a new paralegal can apply it without asking.

• Names: client, spouse, children, parents, sponsors, witnesses, employers, referees, and the firm's own client contact.
• Identifiers: passport, visa and travel-document numbers, national ID numbers, case and receipt numbers, registration and tax numbers, driving licence numbers.
• Dates tied to a person: date of birth, entry and exit dates, marriage dates, hearing dates. Use "[DATE 1]" or a relative description such as "three years before filing".
• Contact and location details: home and work addresses, phone numbers, email addresses, workplace names, school names, a village or small town that narrows the person down.
• Sensitive history: medical conditions, criminal or arrest records, details of persecution, sexual orientation, religion or political activity, when the case turns on them. These stay out even when names are removed, because the story itself can identify someone.
• Financial details: account numbers, salary slips, tax returns, any figure that traces to one employer.
• Images and scans: never upload passports, birth certificates or letters. Text only, after redaction.
• Metadata: file names such as "Passport_JaneDoe.pdf", document properties and email headers.

Beyond the list, add a rule for judgement: if a stranger could work out who this is from the paragraph alone, it is not redacted enough. A woman from a small town who fled a named employer's regional office can be identified without a name. Generalise the detail: "a former employee of a regional manufacturer" gives the model what it needs for tone and structure.

Redaction by placeholder, not by blacking out

Use consistent placeholders so the final letter can be restored: [APPLICANT], [SPOUSE], [EMPLOYER], [SPONSOR], [DATE 1], [TOWN]. Keep a mapping table in the case management system, never in the chat. After the draft comes back, restore the details in your own document editor, then run the lawyer's review. For an overview of how to redact files before sharing them with any AI tool, see how to redact personal data from documents with AI, and note the warning there: asking an AI to redact is not a redaction process unless a person checks the result.

Support letters: what to draft with AI and what to leave alone

Support and evidence letters come in three kinds, and the rules differ for each.

Letters from the firm (cover letters, evidence indexes, legal arguments, requests for documents). These are the safest use. The text is the firm's own and the facts come from the file. A redacted prompt can ask for a structure and a first draft, and a lawyer finishes it.

Letters from an employer or organisation (employment verification, job description, business need). The structure is standard, but the content must be the employer's. Collect facts from the employer with a short questionnaire, and let the model turn the answers into a draft that the employer's officer reads, edits and signs. The model must never fill gaps.

Letters and declarations from individuals (family, friends, community members, the applicant's own statement). These are the riskiest, because the signer attests to personal knowledge. The only acceptable method is an interview: the signer tells you what they know, you write down their facts, the model helps arrange them clearly, and the signer reads, corrects and confirms in their own voice. A polished letter full of details the signer never gave you will not hold up if anyone asks about it.

A redacted drafting prompt, with a sample output

You are helping a law firm draft an employer support letter. Use only the
facts below. Do not invent dates, figures, titles or duties. Where a fact
is missing, write [MISSING: description] instead of guessing.

Facts from the employer's questionnaire:
- Applicant: [APPLICANT], role [ROLE], employed since [DATE 1]
- Duties: manages supplier contracts, leads a team of five
- Why this role needs the applicant: runs the migration of the
  company's inventory system, due to finish next year
- Signatory: [SIGNATORY], [SIGNATORY TITLE]

Write a one-page letter in first person from the signatory. Plain tone.

An illustrative reply opens with "I am writing to confirm that [APPLICANT] has been employed by [EMPLOYER] as [ROLE] since [DATE 1]" and describes the duties as given, then includes "[MISSING: annual salary]" because the questionnaire did not provide it. That placeholder is the model behaving properly. What you would fix: the draft may add a sentence such as "[APPLICANT] has consistently exceeded expectations", which the employer never said. Delete every claim that the facts do not support, and send the employer a version with the unsupported lines removed.

Roles and permissions: who may do what

Put the following in a half-page rule sheet and have each person acknowledge it.

1. Everyone: uses only the firm workspace login, never a personal account, for anything connected with a matter.
2. Paralegals and assistants: may redact and draft, and must show the redacted prompt to the supervising lawyer for the first twenty uses.
3. Lawyers: own the review. No letter is filed, sent or signed without a lawyer reading the restored final text against the file.
4. Owner or practice manager: sets the plan, controls who has seats, reviews the rule sheet every quarter and records any incident.

Add a "stop and ask" trigger. If a document includes a minor, a medical record, a criminal history or a protection claim, the team member asks the supervising lawyer before any AI use, and the default answer is no.

Review steps before anything leaves the firm

AI drafts fail in predictable ways. Build the review around them, not around a vague "check it".

• Fact check. Each factual statement is ticked against the source document or the signer's interview note.
• Placeholder sweep. Search for square brackets so no [MISSING] or [APPLICANT] remains in the final.
• Consistency check. Dates, spellings, titles and job descriptions match the application forms and other letters.
• Tone check. Is it in the signer's voice? Several letters from different people that share sentence structure look generated. Rewrite each in the signer's own words.
• Legal check. The lawyer confirms the letter addresses the legal test the application depends on, without overclaiming.
• Record. Save the final and the interview note in the case system, and delete working chats you no longer need.

Three short cases for the rule sheet

A paralegal pastes a full witness statement into a personal account. It contains a name, address, a medical diagnosis and a hearing date. This breaks three rules at once: personal account, no redaction, sensitive history. The response is to stop, record the incident, delete the chat if possible, and tell the supervising lawyer, who decides whether the client needs to be informed under the firm's obligations. Prevent it by not issuing personal logins in the first place.

A letter where every signer sounds identical. A firm drafts five community letters with the same prompt. Reading them side by side, the same opening and closing appear in each. The fix is to interview each signer, write their points in their own order and edit from there. A reviewer can spot the repetition in a minute if you ask them to read the letters together.

A fully redacted evidence index. The firm pastes a list of exhibits as "Exhibit 1: [DOCUMENT TYPE], [DATE 1]" and asks the model to produce a consistent cover letter table. No identifying content leaves the firm, the output is useful, and a lawyer restores the details. This is the green case you want people to learn from.

Before and after: redacting a real-looking paragraph

A paragraph from an interview note, invented for illustration, and the version that may go into the chat:

BEFORE (never pasted)
Jane Doe, born 4 March 1991, has worked at Example Freight Ltd, 14 Sample
Lane, since June 2019 as logistics manager. Her passport number is
X1234567. She has a heart condition and her sister Mary lives with her.

AFTER (may be pasted)
[APPLICANT] has worked at [EMPLOYER] since [DATE 1] as logistics manager.
Her duties include supplier contracts and a team of five. She is the
primary support for a family member [RELATIVE 1].

Notice what was dropped as well as replaced. The passport number is gone entirely, because it is not needed for drafting. The heart condition is removed because it is not relevant to an employer letter. The relative is kept only as a role in the story. Do the redaction in your own system with find-and-replace against the mapping table, then read the result slowly once; automated redaction tools miss names in odd spellings and nicknames, so a human read is part of the process.

Documents in other languages and translations

Many immigration files include foreign-language documents. Do not paste an original certificate into a chat for translation. Certified translations are normally required for filings, and the translator's certification is part of the evidence, so an AI translation is a working aid for your lawyer to understand a document, not a filing document. If you use a model for a working translation, redact the text first, and mark the file "working translation, not for filing". Check with the person who certifies translations in your firm before any AI-assisted step touches them.

Wording for the engagement letter and the client notice

Telling clients is both honest and practical. Add one short paragraph to your engagement terms. Illustrative wording:

We may use secure AI tools to help draft documents such as cover
letters and indexes. We remove identifying details before using them,
we do not allow them to use your information to train their models, and
a lawyer reviews and takes responsibility for everything we send on your
behalf. Tell us if you would rather we did not use these tools on your
matter.

Only promise what you do. If anyone on your team sometimes uses a personal account, this paragraph is untrue, and untrue assurances are worse than silence. Review the wording whenever you change the plan or the tools, and keep a copy of each version so you know what each client was told.

If something goes wrong: a five-step incident routine

1. Stop and preserve. Note who, when, which account and what was pasted. Do not keep using the same chat.
2. Delete what you can. Remove the chat from the account, and check whether the account was personal or business.
3. Tell the supervising lawyer the same day. They decide whether the incident is reportable to the client or any authority under the rules that apply to your firm.
4. Find the cause. Personal login, an unclear rule, a rushed deadline? Fix that, not the person alone.
5. Update the rule sheet. Add the missing line, brief the team in five minutes and log the incident for the quarterly review.

A firm that logs and fixes small incidents is in a far better position than one that discovers a pattern during a dispute.

Where the model is weakest in this practice

Models write fluent legal-sounding prose, and fluency is exactly why errors slip through. They invent citations, so never accept a case, statute or policy reference that you have not looked up yourself. They tend to overstate: "extraordinary", "exceptional" and "unquestionably" appear in drafts where the facts are ordinary, and an overclaiming letter damages credibility. They flatten differences between visa categories, so a sentence true for one route is wrong for another. Give the model the category and the legal test in the prompt, and have the lawyer confirm the framing independently.

Finally, ask whether AI is the right tool at all for the document in front of you. An applicant's personal statement in their own words, even with grammar problems, can be more credible than a polished version. For that kind of document, use the tool, if at all, for light editing suggestions on a redacted copy and keep the voice theirs.

Retention and clean-up: a monthly five-minute job

Chats pile up. Once a month, the owner or practice manager asks every user to delete working chats from closed matters, and confirms that nothing in the workspace is the only copy of a letter or a note. Because members cannot export their own workspace data, anything worth keeping must already be saved in the case system. Put the clean-up on the same calendar entry as your regular file audit, so it does not rely on memory, and record the date in the incident log.

Hand over to the team in one hour

A workable briefing runs for an hour. Spend ten minutes on why the rules exist, twenty on the redaction list with the before-and-after paragraph above, twenty on a live exercise where each person redacts a made-up interview note and swaps with a colleague to find what was missed, and ten on the review checklist. Collect a signed acknowledgement of the rule sheet. Repeat the exercise for new joiners in their first week, with a different invented note each time so people do not memorise one answer.

Five questions for the owner after month one

At the end of the first month, ask five questions. Who has a seat and who actually used it? Which letter types did the model handle well, and which needed heavy rewriting? How many drafts went to a lawyer with leftover brackets or unsupported claims? Did anyone hit the "stop and ask" trigger, and what was decided? Which line of the rule sheet was misread? Put the answers on one page. If a letter type needed more than half its text rewritten, take it off the approved list until the prompt improves. Adjust the list based on evidence from your own files rather than on enthusiasm, and write down the reason for every change so the next review does not start from scratch.

A small worked example: costing the pilot

An invented four-lawyer practice buys ChatGPT Business for five seats (four lawyers and one paralegal) at $20 a seat billed annually: $100 a month. The owner estimates each user will save about three hours a month on cover letters and evidence indexes, so fifteen hours in total. Setup takes the owner half a day: choosing settings, writing the rule sheet, running a one-hour briefing. Review time is the real cost, because each AI-assisted letter still needs a lawyer's read-through of perhaps twenty minutes. If the firm values the time saved conservatively and counts the review time, the pilot still pays back when it saves more than a couple of hours a month per user. The numbers are illustrative, and the real test is the quarterly review: are drafts shorter to finish, and are errors fewer?

Compare the alternative: case software with a built-in writing assistant. Docketwise or ChatGPT for forms sets out that choice, and the tutorial on answering a request for evidence with AI shows where the review burden rises when a deadline is tight.

How to check that the policy is working

Every quarter, pull ten matters that used ChatGPT and check four things: a lawyer reviewed the final; no identifiers appear in the workspace chats you can see; the placeholders were restored correctly; and the engagement letters match what you now do. If the answer to any is no, tighten the rule that failed rather than adding a new one. Short rule sheets get followed; long ones do not.

Sorting your document types into green, amber and red is also where my AI implementation consultation starts, if you want a second pair of eyes on the rule sheet before the team sees it.

Questions a firm owner should settle before rollout

Is the free ChatGPT plan acceptable for client work?

Not for anything that identifies a client. Consumer plans let a user switch off use of chats for model training, but that is a personal setting that each person can forget or change. Business plans do not train on your content by default and give the firm admin controls, which is the standard a confidentiality policy needs.

Does redaction mean the AI cannot help with the letter?

No. Redaction swaps identifying details for placeholders such as [APPLICANT] or [EMPLOYER], and the facts that matter stay. The model drafts structure and wording, then a person puts the real names back into the final document in your own system, not in the chat.

Can the AI write a support letter from a third party?

It can draft a structure, but the writer must supply the facts and sign only what is true. An AI-drafted letter containing details the signer never gave you is a serious problem. Use interview notes from the signer, let the draft follow them, and have the signer confirm every statement.

What should happen to chats when a staff member leaves?

Plan for it before you roll out. On ChatGPT Business, removing a member ends their access at once, owners and admins cannot see members' private chats by default, and members cannot export their own workspace data. So staff should never treat chats as a record. Keep the case file in your practice system and use ChatGPT only for working drafts with placeholders.

Do clients need to be told we use AI?

Many firms add a short line to the engagement letter, and professional-ethics guidance on generative AI ties the use of confidential client information to the duties of confidentiality and client communication, in some cases to informed consent. Check the rules that apply to you, and if you redact fully so nothing identifying is sent, say so.

Further reads

• Docketwise Pricing and the AI Writing Assistant: What You Get — The writing assistant inside case software.
• Immigration Agent AI: Client Document Checklists and Case Updates — Client checklists and case updates.
• Claude v ChatGPT for Lawyers: A Same-Task Test on Your Own Documents — Test both assistants on your own documents.
• Can Personal Injury Lawyers Put Client Records Into ChatGPT? — The same rules written for injury practice.
• Immigration Case Software With Automation: A Buyer's Checklist — Buying case software with automation built in.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI help articles on data controls and ChatGPT Business (members, export, privacy), OpenAI business pricing page, Claude pricing page, professional-ethics guidance on generative AI in legal practice.