Apple's Transcription System Converted a Routine Dealership Voicemail into an Obscene Attack
What happened
In March 2025, Louise Littlejohn received a voicemail on her iPhone from Lookers Land Rover in Motherwell, Scotland, inviting her to a dealership event. She had previously purchased a vehicle from that garage. What Apple's AI transcription service rendered to her screen bore no resemblance to that invitation. The system produced a profanity-laced text that called her a "piece of s**t" and included questions about her sex life.
Littlejohn, 66, from Dunfermline, initially assumed she was reading a scam message. She recognized the caller's postcode and realized the number belonged to the actual dealership, which forced her to piece together what had happened. The content of the voicemail itself was entirely routine, a scripted promotional call from a sales representative. The transcription system had turned it into something that would cause any recipient distress.
The technical causes were predictable in retrospect. Apple's transcription model struggled with the caller's Scottish accent, compounded by background noise typical of a busy garage floor and the formulaic cadence of a sales script. Those three factors, an unfamiliar accent, ambient sound, and a speech pattern the model was not well calibrated for, combined to produce output that bore no relationship to the source audio. The system had no visible mechanism to signal low confidence or flag the output for review before delivering it.
The deeper problem is that the transcription arrived with the same visual presentation as any other message. Nothing in the interface communicated that the text was an interpretation rather than a record. Littlejohn had no way to know, in the moment she read it, that she was looking at a model's garbled output rather than words someone had actually spoken. The result carried no uncertainty signal. It simply presented itself as fact.
That gap, between what an AI system produces and what actually occurred, is exactly the space that verification infrastructure is meant to fill. A provable record of what a system did, including the source audio, the confidence level of the transcription, and the conditions under which it ran, would let anyone inspect the failure immediately. Without that record, the system's output becomes the only version of events a user ever sees, and a routine dealership call stays misread as an attack until someone thinks to listen to the audio themselves.
Reported impact
- Affected parties
- Not publicly disclosed
- Harm type
- Not publicly disclosed
- Scale
- Not publicly disclosed
- Financial impact
- Not publicly disclosed
- Regulatory action
- Not publicly disclosed
Classification
Relevant governance controls
Governance control mapping is not available for this record.
- No controls mapped
Not publicly disclosed
Control mapping is analytical. It does not state that any control would have prevented the incident.
Sources and evidence
This record was researched and written by the Index. The event is also catalogued in the following database, which is listed for cross-reference.