An articulation word list looks like the simplest material in speech therapy. It is also one of the easiest to get quietly wrong. A word list is a clinical instrument: each word is a promise that practicing it produces the target phoneme, in the target position, one more time. Phonetic validation is the step that checks that promise before a child spends effort on it.
The dosage argument: trials are the budget
Speech sound intervention is a numbers game. Tutorials on motor learning in speech treatment emphasize high amounts of practice, and intervention research repeatedly points at trial counts as a driver of outcomes. Williams (2012) recommends a minimum of about 50 production trials per session for phonological intervention, and more for severe speech sound disorders. Edeal and Gildersleeve-Neumann (2011) compared moderate and high production frequency directly and found greater learning and maintenance for targets treated with more productions.
Now put an off-target word on the list. Every repetition of it spends part of that session's trial budget on the wrong pattern — or at best on nothing. The studies were not written about AI materials, but the implication is hard to avoid: if trials drive progress, the accuracy of what is being drilled is not a cosmetic detail.
Spelling is not sound
The reason lists go wrong is that English orthography is an unreliable witness. The letter is there while the sound is not, or the sound is there while the letter is not:
Silent letters: “island” has an s in the spelling and no /s/ in the word.
Digraphs that shift: “ch” is /tʃ/ in “chair” but /k/ in “chorus” and /ʃ/ in “chef.”
Letters that change sound: “s” is /ʃ/ in “sugar,” /z/ in “dogs.”
Sounds spelled unexpectedly: “mosquito” contains /sk/ but spells it “qu” — a real example an SLP documented after asking ChatGPT for “sk” words.
Humans fall for these traps when skimming. Language models fall for them systematically, because they process text as spelling-based tokens and infer pronunciation from orthographic patterns. Benchmarks such as PhonologyBench show leading models trailing humans by wide margins on phoneme-level tasks, and a 2025 clinician-rated evaluation found that generating therapy stimuli was ChatGPT's weakest clinical task. The model is not careless; it simply never checks sound, because it has no sound to check.
What phonetic validation actually is
Phonetic validation means checking every word against its phoneme transcription — how the word is said — rather than its spelling. In practice, a deterministic validator does four things:
Transcribes: looks each word up in a pronunciation dictionary (Ga-loo uses the CMU Pronouncing Dictionary, 134,000+ North American English entries in ARPAbet notation).
Locates: confirms the target phoneme is present and classifies each occurrence as initial, medial, or final — in the phoneme sequence, not the letter sequence.
Screens: flags competing sounds that could blur the contrast being taught, which matters for minimal pairs and phonological goals.
Refuses to guess: words missing from the dictionary are excluded and reported, not silently kept.
The output is a report, not a reassurance: target sound, occurrence counts, position matches, competing sounds, excluded words. Because the check is rule-based, it gives the same answer every time and cannot be talked into confidence. That is the practical difference from asking a chatbot “are you sure?” — which only ever produces more prose.
How to validate any word list yourself
You do not need software to apply the principle. Before using any list — from a chatbot, a marketplace PDF, or a colleague:
Say every word aloud. Confirm the target sound by ear, in the position you need.
Check the traps: silent letters, digraphs, soft and hard consonant spellings, clusters spelled unexpectedly.
For phonological goals, listen for competing sounds inside the same word.
Cross-check doubtful words against a pronunciation dictionary rather than your reading of the spelling.
Drop words the child is unlikely to know, whatever their phonetics.
This takes a few minutes per list — which is exactly the cost that automated validation removes. Our word lists by sound and position are built this way: every list is generated from the pronunciation dictionary with the validation applied up front, from initial R to vocalic R variants.
Where Ga-loo fits
Ga-loo was built around this single idea: let AI draft the engaging part — themes, layouts, illustrations — and let a deterministic validator hold the clinical part to a fixed standard. Every worksheet ships with its phonetic report, so you can see what was checked instead of taking it on faith. Generate a worksheet and read its validation report — the report is the product as much as the worksheet is.
Validation covers phonetics, not appropriateness. Whether a word fits a particular child's vocabulary, culture, and goals remains a clinical judgment, and the SLP stays in charge of it.
Bottom line
Phonetic validation matters because therapy time is finite and trials are the currency. Spelling-based generation — human or AI — leaks errors precisely where English spelling lies about sound, and those errors spend a child's practice on the wrong target. Check the sound, not the letters: with your own ears, or with a validator that does it deterministically. For the comparison in depth, see ChatGPT vs. a phonetically validated generator and our framework for deciding whether to trust AI-generated therapy materials.
Sources and further reading
Williams, A. L. (2012). Intensity in phonological intervention: Is there a prescribed amount?
Maas, E., et al. (2008). Principles of motor learning in treatment of motor speech disorders
Birol et al. (2025). Is there any room for ChatGPT AI bot in speech-language pathology?
This article is educational and is not legal, compliance, or clinical advice.