A client hands you a letter. It’s a benefits eligibility decision — three pages of dense type, code citations, a “notice of adverse action,” an appeal deadline buried in paragraph nine. They don’t understand a word of it, and honestly, it takes you ten minutes and two re-reads to be sure you do. Now you have to translate it into something they can actually act on, before the appeal window closes.
This is one of the most useful, lowest-risk things AI can do for a social worker — if you do it the safe way. Here’s how to turn a wall of bureaucratic language into a clear, plain-language explanation your client understands, without ever putting their confidentiality at risk.
Why these letters are so hard to read on purpose (sort of)
It’s not your imagination. Government benefits materials are routinely written well above the reading level most people can comfortably handle. Studies of Medicare communications — the Medicare & You handbook, plan documents — found them written at high reading levels with complex formatting that measurably reduced how well beneficiaries understood them. Jargon, long sentences, dense paragraphs, and administrative phrases nobody uses in real life all pile up.
Meanwhile, the people receiving these letters are often the ones least equipped to decode them: low literacy, limited English, a crisis already in progress. Health-literacy experts recommend writing client-facing material at around a 6th-to-8th-grade reading level. The gap between that and a real eligibility notice is exactly the gap you’re standing in.
AI is genuinely good at closing that gap — turning “you are being assessed an overpayment recovery” into “the office says they paid you too much and want some of it back.” But before any letter goes near ChatGPT, one rule comes first.
The rule that comes before everything: de-identify
Never paste a client’s identifiable information into ChatGPT. Consumer ChatGPT is not a confidential, HIPAA-compliant environment — what you type goes to OpenAI’s servers, it isn’t covered by client privilege, and in the wrong situation it could be subpoenaed. Using your personal account to process real client details (“Shadow IT,” as it gets called) is how social workers end up facing data-protection violations, licensing problems, or worse.
And de-identifying doesn’t just mean deleting the name. Under HIPAA’s Safe Harbor standard there are 18 identifiers that make data identifiable — names, yes, but also Social Security numbers, dates (including birthdates), addresses, phone numbers, email, case and record numbers, even unusually specific amounts. Removing the name alone is not enough.
The safe four-step workflow
Social-work AI guidance calls this the “review and sign” pattern: strip identifiers, let AI draft, read every word, then put the real details back inside your secure system — with your professional accountability on the final product.
Step 1 — De-identify in your own system. In your secure document, replace the client’s name, case number, SSN, DOB, address, and any uniquely identifying dates or dollar figures with placeholders: [client], [date], [amount].
Step 2 — Ask AI to rewrite the de-identified text. A prompt that works well:
“Rewrite this benefits decision letter in plain language at about a 6th-to-8th-grade reading level, for a client whose first language may not be English. Keep all deadlines and appeal rights accurate and clear. Do not add any new facts or guess at anything that isn’t written here.”
That last sentence matters — it tells the AI not to invent.
Step 3 — Review every word against the original. This is the non-negotiable part. Check that every date, deadline, and appeal option matches the real letter exactly. AI is a strong rewriter and a terrible lawyer — never let it interpret an ambiguous rule, only restate a clear one. Fix anything vague, wrong, or that could mislead.
Step 4 — Re-identify inside your secure environment. Paste the human-checked plain-language version into your agency’s letter, script, or case-management system, and add the real names, dates, and amounts back there — never in the chatbot.
The whole thing takes minutes, and the client walks out understanding what happened and what to do next.
What this means for you
If you carry a heavy documentation load: This is the highest-value place to start. Benefits and eligibility letters are dense, repetitive, and emotionally heavy for clients — exactly where a fast plain-language rewrite changes someone’s day.
If you work with limited-English clients: Ask the AI to write at a plain reading level first, then you can have it produce a translated draft — but treat any translation as a draft to be checked by a qualified human, not a finished document.
If your agency hasn’t set AI rules yet: Use only de-identified text, document that you’re doing so, and raise it with your supervisor. Being the person who used AI carefully and flagged it is very different from being the person who got caught pasting case files into a chatbot.
If you’re new to the field: Use this to save time on wording — never to skip learning how the benefits system actually works. The shortcut is the writing, not the understanding.
What it can’t do
- It can’t be trusted with the law. AI can confidently state a wrong deadline or invent an appeal right that doesn’t exist. Every fact gets checked against the original letter. Always.
- It doesn’t know your client. It reproduces patterns from its training data, which can carry bias and miss cultural context entirely. It cannot account for someone’s history, trauma, or situation — that’s your read, not its.
- It can’t be your confidential workspace. Consumer ChatGPT isn’t built for protected client data. The de-identify step isn’t optional politeness — it’s the line between a helpful tool and an ethics violation.
- It can’t do the human part. AI can make a letter readable. It can’t sit with a frightened client, gauge whether they actually understood, or decide what’s really in their interest. That’s the job. AI just clears the paperwork so you have more time for it.
The bottom line
A dense benefits letter is a problem AI is genuinely good at — and one you can solve in minutes without ever compromising a client’s privacy, as long as the de-identify step is sacred. Strip the identifiers, let AI handle the plain-language rewrite, check every fact against the original, and put the real details back where they belong: in your secure system, under your signature.
Getting the confidentiality part right is the whole game. Our ChatGPT Without the Liability course walks through exactly how to use AI on sensitive work without crossing privacy or ethical lines — the de-identification habits, the prompts, and the review steps that keep you and your clients protected. Start there, then put it to work on the next letter that lands on your desk.
