Research Note CA-2026-01
How often do people realise they are talking to an AI?
A transcript-level census of 102,154 connected human phone conversations. 253 contained an explicit challenge to the agent's humanity — 0.25%, about one in 404.
Published · Measured 19 July – 17 August 2026
0.25%
Measured — published methodOf connected human phone conversations with a CallableAI AI Voice Agent, 0.25% contained an explicit challenge to the agent's humanity — 253 of 102,154, or about one conversation in 404.
Measured 19 July – 17 August 2026
Method
- Every row in the production call_logs table for the 30-day window (Australia/Brisbane), across 11 business tenants running outbound AI voice agents. Read-only; no records modified.
- Inclusion criteria: a transcript present and longer than 50 characters, duration of at least 25 seconds, and at least three caller turns. 132,276 conversations qualified.
- 28,787 of those (21.8%) were then removed as machine answerers rather than humans — voicemail greetings, carrier and Google Call-Assist screening bots, and “this call is being recorded” announcements. Those systems reliably produce false detections: the robot in the transcript is the network's, not the caller's. That left 103,489 human conversations, of which 102,154 had machine-parseable caller turns.
- Agent speech was discarded before matching, so the AI Voice Agent's own wording can never trigger a detection. Only caller turns were scanned, case-insensitively, across five phrasing families: real-person or human challenges, explicit naming of AI, robot or bot, machine or computer, and recording or prerecorded. A conversation counts once however many times the caller raises it.
- 253 conversations contained a detection utterance: 253 of 102,154, or 0.25%. Wilson 95% interval 0.22%–0.28%, reported for completeness — because every qualifying conversation was scanned, this is an observed population value rather than an estimate.
- Cross-check: a deterministic random sample of 500 conversations from the same population returned 1 detection in 493 with parseable caller turns (0.20%), consistent with the census. At this base rate a 500-call sample is a weak instrument — one conversation moves it by 0.2 points — so the census is the figure to cite.
Stable week to week
| Week commencing | Conversations | Detections | Rate |
|---|---|---|---|
| 20 July 2026 | 17,838 | 49 | 0.27% |
| 27 July 2026 | 19,696 | 50 | 0.25% |
| 3 August 2026 | 25,709 | 81 | 0.32% |
| 10 August 2026 | 33,182 | 70 | 0.21% |
| 17 August 2026 (partial) | 5,731 | 3 | 0.05% |
Detection rises with conversation length
| Conversation length | Detections | Rate |
|---|---|---|
| 25–59 seconds | 128 of 75,262 | 0.17% |
| 60–119 seconds | 79 of 21,611 | 0.37% |
| 2–5 minutes | 43 of 5,207 | 0.83% |
| Over 5 minutes | 3 of 76 | 3.95% |
How the question actually gets asked
| Category | Conversations | Share |
|---|---|---|
| Human / real-person challenge | 187 | 73.9% |
| Called a machine or computer | 32 | 12.6% |
| Asked if it was a recording | 22 | 8.7% |
| Named explicitly as AI | 7 | 2.8% |
| Called a robot or bot | 6 | 2.4% |
What callers actually said
- Are you a real person? — 14 conversations
- Are you real? — 6 conversations
- Is this a recording? — 4 conversations
- Alex, is this a prerecorded message, or are you a real person?
- Are you a human or are you a machine?
- Am I talking to a real person or AI?
What this number does not say
- It counts voiced detection only. Anyone who suspected the agent and simply ended the call without saying so is invisible to this method. The true rate of private suspicion is necessarily higher, and this study places no bound on it.
- Conversations under 25 seconds, or with fewer than three caller turns, are excluded by design — a hang-up with no exchange cannot be scored. If silent suspicion concentrates anywhere, it concentrates in that band.
- The rule matches phrasing, not intent. Three quarters of detections are the generic “are you a real person?”, which people also ask human offshore call centres. Only 7 conversations in 102,154 — 0.007% — named the agent as AI unprompted. That makes 0.25% a conservative ceiling on confident detection rather than a floor.
- Detections rest on automatic speech recognition. Spot-checking found occasional garbles that inflate counts and near-misses that deflate them; the errors are small relative to the effect and run in both directions.
- Single vendor, single market: all conversations were placed by CallableAI AI Voice Agents to Australian numbers, largely in solar, finance and services outbound. It should not be read as a property of conversational AI in general.
- Our AI Voice Agents introduce themselves as AI. This measures what happens after disclosure, not whether a synthetic voice passes for human — a different question, and not one this study answers.
Cite this
Research Note CA-2026-01 (PDF, 5 pages)Callable AI (2026). How often do people realise they are talking to an AI? A transcript-level census of 102,154 connected human phone conversations. Research Note CA-2026-01, 17 August 2026.
Cite this note
Callable AI (2026). How often do people realise they are talking to an AI? A transcript-level census of 102,154 connected human phone conversations. Research Note CA-2026-01, 17 August 2026.
Deployed white glove. Backed by an SLA.
CallableAI works with organisations that need quality, volume and compliance to hold at once. Bring us your requirements — we'll scope the deployment, the rollout and the SLA with you.