ChatGPT vs. a Phonetically Validated Generator

Many SLPs now draft therapy materials with ChatGPT, and for some tasks it works well. But an articulation word list is not ordinary text. Every word on it is a claim: this word contains the target phoneme, in the target position, and is worth a child's practice trials. The real comparison is not ChatGPT versus another brand. It is an unverified draft versus a draft that passes a deterministic phonetic check.

What ChatGPT actually does with a word list

Large language models process text as subword tokens, not as phonemes. They never hear a word; they infer pronunciation indirectly from spelling patterns in training data. That works surprisingly often in English, and fails exactly where English spelling diverges from sound: silent letters, digraphs with multiple pronunciations, and clusters that are spelled one way and said another.

This is a measured limitation, not speculation. The PhonologyBench benchmark tested leading models on grapheme-to-phoneme conversion, syllable counting, and rhyme generation, and found gaps of 17–45% behind humans on phonological tasks, with no model reliably strong across all three.

A documented failure, not a hypothetical one

SLP Lori Price described asking ChatGPT for words with the “sk” cluster and receiving “mosquito.” The /sk/ sound is there when you say the word, but the spelling uses “qu” — a poor teaching word for a student who is also attending to letters. Her conclusion was balanced: text-based material generation is a good use of the technology, as long as the clinician stays mindful of its limits.

A 2025 peer-reviewed evaluation by Birol and colleagues had fifteen practicing SLPs rate ChatGPT-4 across six clinical task types. Creating therapy stimuli for speech sound disorders scored the lowest of all six tasks (1.60 on a 3-point accuracy scale), while tasks like clinical decision support scored near the top. The weakest link was exactly the task word lists depend on.

The errors are hard to catch precisely because the output looks polished. A list of twelve tidy “initial R” words with one silent-letter trap in the middle reads as done. Catching it requires saying every word aloud and knowing what to listen for — which is the preparation time the tool was supposed to save.

What a phonetically validated generator does differently

A phonetically validated generator separates drafting from verification. The draft can come from an AI model, but every candidate word is then checked against a fixed pronunciation dictionary — not against spelling. Ga-loo uses the CMU Pronouncing Dictionary, an open North American English dictionary with over 134,000 entries in ARPAbet phonemic notation, the same resource used in speech recognition and text-to-speech research for decades.

For each word, the validator answers concrete questions:

Does the word's phoneme transcription actually contain the target sound?
Is the target in the requested position — initial, medial, or final — based on phonemes, not letters?
Does the word contain competing sounds that could confuse the contrast being taught?
Is the word in the dictionary at all? Unknown words are excluded rather than guessed.

The result is a validation report attached to the worksheet: target sound, occurrence count, position matches, competing sounds, and any excluded words. Because the check is deterministic, the same list always produces the same report, and a failure is visible instead of silent.

An honest head-to-head

Speed of a first draft: tie. Both produce a list in seconds.
Flexibility: ChatGPT wins. It will write anything — stories, jokes, IEP-adjacent prose — with no guardrails.
Phonetic accuracy: the validated generator wins. Accuracy is checked against a pronunciation dictionary, not assumed.
Transparency: the validated generator wins. A report you can read beats confidence you have to trust.
Repeatability: the validated generator wins. Deterministic checks give the same verdict every time; a chatbot may not.
Cost: ChatGPT wins at the free tier — if you price your own review time at zero.

The pattern is simple: ChatGPT is a general-purpose drafting tool, and it is genuinely useful as one. It is not a verification tool, and articulation word lists are a verification problem.

A practical workflow for SLPs

If you continue using ChatGPT for word lists, treat the output as unverified:

Constrain the prompt: ask for words that contain the sound and are spelled the expected way, if orthography matters for your student.
Say every word aloud before printing. Check the target sound and its position by ear, not by eye.
Watch for silent letters, soft and hard consonant spellings, and cluster substitutions.
Drop vocabulary the child is unlikely to know — models routinely overshoot age level.

Or move the verification into the tool. Browse our verified word lists by sound and position to see what deterministic checking looks like on real lists, or generate a validated worksheet and read its phonetic report yourself.

Where Ga-loo fits

Ga-loo uses AI where it is strong — drafting themed, engaging material — and a deterministic validator where AI is weak. Every generated worksheet is checked word-by-word against the pronunciation dictionary before you see it, and the validation report stays attached. The clinician still makes the clinical call: whether the vocabulary, theme, and difficulty fit the learner is a professional judgment no validator can make.

Bottom line

ChatGPT and a phonetically validated generator are not competitors on the same axis. One predicts plausible text; the other proves a phonetic property of every word it outputs. For casual drafting, use whichever you like. For materials a child will practice from dozens of times, use a process where accuracy is checked — by a validator, by you, or ideally both. For the broader safety picture, see our guide on what's safe and what's not when using AI in speech therapy and the companion piece on why phonetic validation matters clinically.

Sources and further reading

This article is educational and is not legal, compliance, or clinical advice.