A specific claim entered workplace discourse in 2026: that heavy use of AI writing tools is eroding people's ability to speak without preparation. It is an unusually testable claim, and it is worth separating what has been measured from what has been asserted.
What the survey evidence says
The most direct evidence is a survey conducted by Preply with Dataframe in May 2026, covering 1,142 US respondents of whom 1,002 were employed.[1] Its findings describe a split between written and spoken confidence:
| Finding | Share |
|---|---|
| Say using AI has made spontaneous conversation feel more difficult | 51% |
| Same, among Gen Z workers | 66% |
| Say they sometimes freeze in person because they cannot review or edit first | 44% |
| Use AI to write work communications several times a day | 33% |
| Report AI increased their confidence in their communication skills (daily users) | 71% |
| Same, among workers overall | 53% |
The internal contrast is the interesting part. Daily AI users report higher communication confidence than workers overall, 71% against 53%, while a majority of the same population reports that unscripted conversation has become harder. Confidence in the assisted channel and confidence in the unassisted one are moving in opposite directions.
A larger 2026 survey by SHRM, covering 5,875 US workers, documents the surrounding context: AI adoption is changing which skills the job requires.[2]
What this evidence cannot establish
Four limits apply, and none of them are minor.
It is self-report about a counterfactual. Respondents are being asked whether something has become harder than it used to be, which requires them to compare present experience against a remembered baseline. People are poor at this in general, and the question invites a narrative that is already culturally available.
Selection and attribution are unresolved. Workers who adopt AI writing tools daily may differ systematically from those who do not, including in their prior comfort with live conversation. Someone who dislikes unscripted speech has a reason to reach for a tool that removes it.
One vendor's survey is one vendor's survey. Preply sells language tutoring. The finding is plausible and the methodology is disclosed, which is more than most such surveys offer, but it has not been independently replicated.
No behavioural measurement exists. Every figure above is a perception. None of it counts a word, times a pause, or measures a disfluency.
What behavioural measurement would require
The obvious test is a corpus: take recordings of unscripted speech across a period of rising AI adoption and measure whether disfluency rates, answer length, or response latency shifted.
We attempted exactly this on a corpus of 5,438 AI-led mock interview sessions and could not answer it. Monthly medians of measured filler rate ranged from 0.49 to 3.72 per 100 words, but the step changes aligned with changes in the automatic transcription pipeline rather than with any plausible behavioural pattern. Automatic speech recognisers differ in whether they preserve filled pauses, and a recogniser change moves the measurement without anyone's speech changing at all.
This is a general obstacle rather than a local one. Most large corpora of contemporary spontaneous speech are assembled from automatic transcripts, and most are assembled by products whose transcription stack evolves. A valid longitudinal disfluency comparison needs the recognition pipeline held constant and documented across the whole window, or re-transcription of the full archive under a single model. Any study reporting a disfluency trend across 2024 to 2026 without addressing this should be read with that in mind.
Within a fixed pipeline, cross-sectional description remains valid. Our corpus study of 16,928 interview answers reports the current distribution of answer length and filler rate.[6]
Why disfluency is the wrong outcome measure anyway
Even with a clean longitudinal corpus, filler frequency would be a weak proxy for the thing people are worried about.
Filled pauses are not noise. Clark and Fox Tree analysed uh and um as words that signal an upcoming delay, with um projecting a longer one.[3] Listeners use them, recalling speech better and anticipating the speaker's difficulty.[3] Disfluency rates vary systematically with age, relationship to the listener, topic, and conversational role,[4] and shift between public and private settings.[5] A rate is a property of a situation as much as of a speaker.
If AI use is degrading unscripted speaking, the damage people describe is not "more ums." It is structural: losing the thread, failing to organise a point in real time, freezing when the reviewable draft is unavailable. Those are measurable, but they require scoring of structure and coherence rather than token counting.
What would settle it
- A longitudinal corpus with a documented, frozen transcription pipeline, or a full archive re-transcribed under one model.
- Outcome measures covering structure and coherence, not filler frequency alone.
- A within-subject design comparing the same speakers before and after a change in AI writing usage, which removes the selection problem that the cross-sectional surveys cannot.
- Independent replication of the Preply finding by a party that does not sell communication training. That includes us.
Current state
The claim that AI use makes spontaneous speech harder is, as of September 2026, widely self-reported and not behaviourally demonstrated. The self-report is consistent across a plausible mechanism and a coherent internal contrast, which is a reason to take it seriously rather than a reason to treat it as established.
The honest description is that a measurable question is being answered with surveys because the measurement is harder than it looks.