How SLPs Can Use ChatGPT & Claude for Therapy Materials (Prompts, Limits & Privacy)

Most speech-language pathologists we hear from aren't asking whether to use ChatGPT anymore. They're asking how to get something usable out of it without spending twenty minutes fixing a word list, & whether they're about to break a privacy rule by typing in a student's goal. Both questions have concrete answers, & that's what this guide is: prompts you can copy, the specific ways AI gets speech therapy content wrong, & the privacy lines you shouldn't cross with ChatGPT, Claude, or anything else with a chat box.

Here's the TLDR: general AI assistants are genuinely good at the text-heavy parts of an SLP's week, & genuinely unreliable at the parts that depend on sounds rather than spelling. Ranked by how much time they save with the least risk:

1. Parent letters & home-practice instructions - best use. Plain-language rewriting is what these models do best.
2. Short stories, scripts & social scenarios - great first drafts at a set reading level, as long as you check every target word.
3. WH-question, category & vocabulary sets - fast & mostly accurate.
4. Articulation word lists & minimal pairs - useful, but this is where the errors live. Always ask for a phonetic transcription so you can catch them.
5. Goal-bank drafting - fine as a starting template with no student details. You write the real goal.
6. Anything about a real child - don't. No names, no birthdates, no school, no rare diagnoses, no audio.

Prices checked October 2026. If you want a broader look at which tools are worth paying for, our companion guide to the best AI tools for speech-language pathologists ranks the options side by side.

The Privacy Rules First (Because They Decide Everything Else)

ASHA has published direct guidance on this, & it's blunter than most vendor marketing. From ASHA's Prioritizing Security page: "Sensitive, proprietary, and personally identifiable information (PII) should not be entered into generative AI tools." It also tells clinicians to "assume that all generative AI tools are not HIPAA compliant, unless vetted by your facility's privacy, legal, or compliance team," & adds that "even if your version of that tool is a paid account, do not share clients' or staff members' PII." Paying $20 a month doesn't change the rule.

Two federal frameworks sit behind that advice, depending on where you work:

• School SLPs: FERPA. The Department of Education's student privacy glossary defines personally identifiable information broadly: direct identifiers like a name or ID number, indirect identifiers like a date of birth, & other information that's "linkable" to a specific student. The FERPA regulations (34 CFR 99.3) also list voiceprints under biometric records. That matters for us: a recording of a child's speech is not an anonymous file.
• Clinic & private-practice SLPs: HIPAA. HHS's de-identification guidance describes the Safe Harbor method, which removes 18 types of identifiers (names, geographic detail smaller than a state, most dates, full-face photos, biometric identifiers including voiceprints, & more). If a tool's vendor won't sign a Business Associate Agreement with you, treat it as off-limits for protected health information.

Here's where the big assistants stand on that, from their own help pages:

• ChatGPT's personal plans (Free, Plus & up) may use your chats to train models by default. You can turn that off under Settings → Data Controls → "Improve the model for everyone." OpenAI says it only offers a BAA on sales-managed Enterprise & Edu accounts, & its help center states it doesn't offer one for ChatGPT Business.
• Claude (Free, Pro, Max) asks consumer users to choose whether their chats can be used for training, & that choice lives under Settings → Privacy. Anthropic only offers a BAA on HIPAA-enabled Enterprise plans.
• The free teacher editions - ChatGPT for Teachers & Claude for Teachers - are free for verified US K-12 educators (Anthropic explicitly includes specialists & interventionists) & don't train on your conversations. Both are described as FERPA-aligned. But Anthropic's own data page says that without district authorization you still shouldn't enter identifiable student information. "FERPA-aligned" is not the same thing as "your district approved it."

The practical rule we'd give any SLP: write prompts about a profile, never about a person. "A 7-year-old working on initial /s/ who loves dinosaurs" is fine. A first name, a school, a diagnosis rare enough that one child in your district has it (ASHA calls this one out specifically), or an uploaded evaluation report is not. ASHA's informed consent guidance also says to tell clients when generative AI is used in their care, which is worth building into your intake paperwork if AI touches anything client-specific.

8 Prompts That Actually Work for Therapy Materials

Each of these works in ChatGPT, Claude or Gemini. The pattern behind all of them is the same: give the model the clinical constraints up front (sound, position, syllables, age, reading level), ask for a format you can check quickly, & ask it to flag its own uncertainty. Replace anything in [brackets].

1. An articulation word list you can actually check

Make a list of 30 picturable nouns for a [6]-year-old with the /[s]/ sound in [initial] position.
Rules:
- The target must be the SOUND /[s]/, not just the letter (no "sugar", no "shoe").
- Singleton only: no consonant clusters at the start (no "star", "snake").
- [1-2] syllables, common words a [first grader] knows.
- Show each word with an IPA transcription in a table: word | IPA | syllables.
- After the table, list any words you are unsure about.

The IPA column is the whole trick. It forces the model to commit to how the word is pronounced, & it makes errors visible at a glance instead of buried in a list of 30 English words.

2. A sound-loaded story at a set reading level

Write a 150-word story for a [2nd-grade] reading level about [a dog at the beach].
Use at least 12 words with /[k]/ in [final] position. Bold every target word.
Keep sentences under 12 words. Avoid other words with /[g]/ so the contrast is clean.
End with 3 comprehension questions (one who, one what, one why).

3. Minimal pairs for a phonological pattern

Give me 15 minimal pairs that target velar fronting (/k/ vs /t/), word-initial.
Both words in every pair must be real, picturable nouns a preschooler knows
(e.g. key / tea). Show IPA for both words. Mark any pair where one word is
hard to picture.

4. WH-questions around a theme

Write 20 WH-questions about [a trip to the grocery store] for a [5]-year-old
working on answering who, what, where, when and why. Group them by question type.
For each one, give a correct answer and one plausible wrong answer I can use as a
choice. Keep vocabulary concrete.

5. Social scenario cards

Write 8 short social scenarios (2-3 sentences each) for kids aged [8-10]
practicing perspective taking at [school]. For each: what happened, a question
about how the other person might feel, and two possible responses, one more
helpful than the other, without labelling one as "bad". Use neutral names.
Avoid scenarios that teach kids to hide stimming or force eye contact.

That last line matters. Left alone, models tend to reproduce older compliance-style social skills material. If your practice is neurodiversity-affirming, say so in the prompt.

6. A parent home-practice letter

Write a friendly half-page note to parents explaining how to practice the /[r]/
sound at home for 5 minutes a day. Reading level: grade 6. No jargon; if you use a
term like "articulation", explain it in the same sentence. Include 3 everyday
activities (car rides, dinner, bedtime) and one sentence on praising effort.
Leave [CHILD NAME] as a placeholder.

If you then ask for a Spanish (or any other language) version, have a fluent speaker read it before it goes home. Translation is good, not perfect, & a parent letter is exactly where a stiff or wrong phrase undermines trust.

7. A goal-bank starting point (no student details)

Draft 5 example measurable goals for a generic school-age student working on
[producing /l/ in conversation]. Each goal needs a condition, a behavior, a
criterion and a measurement method. Show a progression from word level to
conversation. These are templates; do not invent baseline data.

Use the output to speed up wording, then write the real goal from your own evaluation data inside your district's IEP system. ASHA's Documentation in Schools practice portal page is the better reference for what a defensible goal needs.

8. A mixed-group session plan

Plan a 30-minute speech group for 3 second graders: one working on /s/
blends, one on past-tense -ed, one on answering why-questions. Use one shared
[pirate] theme so they play the same game but each gets their own targets. List
materials, a 5-minute warm-up, a 20-minute main activity and a 5-minute wrap-up.

Where AI Gets Speech Therapy Wrong

Language models are trained on text, so they think in spelling. Speech therapy happens in sounds. That mismatch produces a predictable set of mistakes, & once you know them you can catch them in seconds:

• Letter, not sound. "Knife" in a /k/ list, "sugar" in an /s/ list, "giraffe" in a /g/ list, "phone" left out of an /f/ list.
• Voiced vs voiceless "th". Models often mix "this" (/ð/) & "thumb" (/θ/) in the same list.
• Clusters sneaking in. Ask for singleton /s/ & you'll still get "spoon" & "star" unless you say otherwise.
• Vocalic /r/ treated as one target. "Car," "bird" & "deer" are different /r/ contexts; a model will happily mix them unless you name the vowel context you want.
• Dialect assumptions. Output defaults to General American English. For a child who speaks another dialect, a "correct" target in the list might not be the child's target at all. ASHA's Cultural Responsiveness page is the right frame here, not the chatbot's.
• Confident norms & fake citations. Ask "at what age should a child produce /r/?" & you'll get a crisp number, sometimes with a citation that doesn't exist. Check developmental expectations against ASHA's Speech Sound Disorders practice portal & the research it cites. We've written about why AI hallucinations happen if you want the longer version.
• It can't hear anyone. A general chatbot can't judge a child's production. Don't paste in "she said 'wabbit' for 'rabbit', what's the disorder?" & treat the answer as anything more than a brainstorm.

ASHA's generative AI hub for clinicians frames it the same way: AI doesn't replace clinical judgment, & the clinician stays responsible for anything it produces. In practice that means AI drafts, you check, & nothing goes in front of a child unread.

ChatGPT vs Claude vs the Alternatives - Which Should You Use?

For the prompts above, the honest answer is that the major assistants are close enough that your workplace's rules matter more than the model. If your district has approved one of them, use that one. If not, here's how they compare on the things an SLP should care about:

Tool	Price (Oct 2026)	Trains on your chats?	Best for
ChatGPT Free / Plus	$0 / $20 a month	Yes by default; you can opt out	General drafting if you're not a K-12 educator
ChatGPT for Teachers	Free for verified US K-12 staff	No	School SLPs who want ChatGPT with education terms
Claude Free / Pro	$0 / $20 a month ($17 billed yearly)	Your choice in privacy settings	Long, careful rewrites & parent-friendly text
Claude for Teachers	Free for verified US K-12 educators	No	School SLPs; Anthropic names specialists as eligible
Gemini (Google Workspace for Education)	Included in Workspace for Education editions	No, per Google's education terms	Districts that already run on Google
MagicSchool	Free; Plus $12.99 a month or $99.96 a year	No (education vendor with FERPA/COPPA certifications)	Teacher-style templates, including an IEP generator
Speech Therapy Magic (Magic Studio)	$9.99 a month or $79.99 a year after a 7-day trial, no card	Deck briefs are generated through OpenAI as a listed subprocessor	Turning a prompt into an interactive deck that students play

A note on the ChatGPT prices: OpenAI's pricing pages block automated checks, so the $20 Plus figure comes from OpenAI's own pricing page as shown in search results rather than a page we could load directly. Confirm on chatgpt.com/pricing before you pay. Claude's prices are from claude.com/pricing, MagicSchool's from its pricing page. For a wider look at the general assistants, see our ChatGPT alternatives roundup.

Where Speech Therapy Magic's Magic Studio Fits

Everything a chatbot gives you is text. You still have to turn the word list into cards, find pictures, & get it in front of a student. That gap is the reason Magic Studio exists. It's a drag-and-drop deck builder inside Speech Therapy Magic with 22 element types (flashcards, picture choices, minimal-pair pickers, recording prompts & so on). The AI part is the "deck brief": you describe the deck you want in a short form, & it builds a draft deck you edit. Students open a published deck with a 6-digit PIN in any browser, with no account.

What it does that ChatGPT doesn't: the recording prompts score the child's pronunciation of the word & each sound in it using Speech Therapy Magic's own engine, ArticScore, & results come back to the child's profile. What it doesn't do: it isn't a diagnostic tool, it doesn't detect distortions, & it hasn't been validated as a measure of disordered child speech, so treat scores as practice feedback, not assessment data. The plan includes 5 deck briefs a month, 3 student profiles (extras $4.99 a month each, up to 10), & costs $9.99 a month or $79.99 a year after a 7-day trial with no card. There's no free plan. The same privacy rule applies: deck briefs are generated through OpenAI, so describe the target, not the child. Full details are on the pricing page.

If you need to score pronunciation specifically, the sibling article on apps that score a child's pronunciation covers what that technology can & can't tell you.

A 10-Minute Workflow That Holds Up

1. Write the prompt about a profile (age, target, interests), not a person.
2. Ask for IPA or a table so errors are visible.
3. Scan the output for the six error types above. Delete, don't argue with it.
4. Ask it to "replace the words I deleted with new ones that follow the same rules."
5. Move the checked list into whatever your students use: a printable, a slide, a Boom deck, or a Magic Studio deck.
6. Save the prompt that worked. A small prompt library is worth more than a fancier model.

How We Evaluated This

We started from the jobs SLPs most often describe handing to AI (materials, parent communication, planning) & wrote prompts around the clinical constraints those jobs need. The error categories come from how language models represent words (by spelling, not by sound) & are easy to reproduce yourself with prompt #1. Privacy statements are quoted from ASHA, the US Department of Education, HHS, OpenAI's help center & Anthropic's support pages, & plan prices were checked on each vendor's own site in October 2026, with the OpenAI caveat above. Speech Therapy Magic facts come from its live pricing & product pages. We didn't run a formal accuracy benchmark of the models, & we're not quoting one.

FAQ

Is ChatGPT HIPAA compliant for speech therapy?

Not on the plans most SLPs use. OpenAI only offers a Business Associate Agreement on sales-managed Enterprise & Edu accounts, & ASHA tells clinicians to assume generative AI tools aren't HIPAA compliant unless their organization has vetted them. Keep protected health information out of it.

Can I put a student's first name in ChatGPT if I leave out the last name?

We wouldn't. FERPA's definition of PII includes information that's linkable to a student, & a first name plus a grade, a goal & a school is easy to link. Use a placeholder like [STUDENT] & fill it in after the text leaves the chatbot.

Is ChatGPT or Claude better for SLPs?

For the materials in this guide, both are capable. Pick the one your employer approves. If you're a US K-12 SLP, both now have free educator editions that don't train on your chats, which beats either consumer plan.

Can AI write IEP goals?

It can draft the wording of a template goal quickly. It can't know the student's baseline, & it shouldn't be given it in a consumer tool. Write the final goal yourself in your district's system from your own data.

Do I have to tell families I used AI?

For generic materials like a word list, there's no special rule. Where AI is used in a client's care, ASHA's guidance is to tell clients & get informed consent. Check your employer's policy too, which may be stricter.

What's the fastest way to turn AI output into something a child can play?

Paste a checked word list into a tool your students already use. For interactive, self-checking decks, our guide to Boom Cards alternatives for SLPs compares the options, including Magic Studio.

If you're also wondering about the bigger classroom debate, our older piece on whether it's ethical to use ChatGPT in the classroom is still a good starting point.

Disclosure: Justin Gluska, who writes for Gold Penguin, also builds Speech Therapy Magic, which appears in this article.