Ambient clinical documentation can confuse an encounter's authorized speakers with speech overheard from elsewhere. The study first inserted incidental small talk into 576 simulated patient-clinician dialogues: frontier models repeated such content in 35% of notes and clinically misattributed it in 3.7%. That experiment tests context filtering during note…
Inspect the ClaimsCase · DiggingBeagle record
Ambient clinical AI can misattribute another patient's speech
Controlled clinical-note experiments show that unrelated speech can enter transcripts and sometimes the wrong patient note, without any adversarial instruction. The study separates 576 synthetic dialogues from 57 mock audio consultations. Real patient disclosure and clinical harm were not established.
Scope
Controlled finding or reported disclosure as of 2026-10-07; Synthetic/mock encounters; no confirmed clinical privacy breach.
Assessment profile
Assessment method- Setting
- simulation
- Exploitation
- not established
- Evidence
- Grade B
- Remediation
- not applicable
- AI role
- AGAINST AI
Basis and provenance
Desk review of scoped canonical Claims, the pinned SRC-2026-CP3-AMBIENT-CLINICAL-CONTAMINATION Source, and the publicly accessible primary disclosure. The classification applies to the documented setting and prerequisites only. Impact remains unbanded because real victim consequences are not established, and no independent reproduction is claimed. Details: Real-world incidence and effects on actual clinical decisions were not measured.
Assessed Oct 13, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-CP3-AMBIENT-CLINICAL-CONTAMINATION-1 · CLM-CP3-AMBIENT-CLINICAL-CONTAMINATION-2 · CLM-CP3-AMBIENT-CLINICAL-CONTAMINATION-3 · Ambient clinical AI notes contaminated by unrelated background conversations
At a glance
The relevant trust boundary is the patient encounter, not whether the background words contain instructions. Validation must track speaker and encounter provenance from audio to transcript and final note; transcript errors and note contamination are different metrics.
Read the implicationsReal-world incidence and effects on actual clinical decisions were not measured. How performance changes with microphones, room layouts, speaker overlap and clinician review remains untested.
Limits and uncertaintyFull account
Ambient clinical documentation can confuse an encounter's authorized speakers with speech overheard from elsewhere. The study first inserted incidental small talk into 576 simulated patient-clinician dialogues: frontier models repeated such content in 35% of notes and clinically misattributed it in 3.7%. That experiment tests context filtering during note writing rather than malicious prompt injection.
A separate experiment placed audio from another mock patient consultation in the background of 57 recordings. At minus 10 decibels, 48.2% of transcripts contained the outside speech, whereas 5.3% of downstream notes were contaminated. This distinction matters: transcription leakage does not necessarily propagate into the final note, but the note-generation step is not a dependable barrier. No real patient encounter, clinical disclosure or resulting health injury was shown.
Mechanism and trust boundary
- 01
Audio intake
A recording includes the intended clinical speakers plus speech from a distinct simulated encounter.
Boundary: A recording includes the intended clinical speakers plus speech from a distinct simulated encounter.
- 02
Attribution error
Speech separation or transcription fails to bind some utterances to the correct encounter.
Boundary: Speech separation or transcription fails to bind some utterances to the correct encounter.
- 03
Note generation
The clinical note may copy or clinically use the unrelated speech as if it belonged to the active patient.
Boundary: The clinical note may copy or clinically use the unrelated speech as if it belonged to the active patient.
Timeline
- Oct 7, 2026disclosure
Public disclosure
Claims & evidence
CLM-CP3-AMBIENT-CLINICAL-CONTAMINATION-1The study reports small-talk inclusion in 35% of notes from 576 simulated patient-clinician dialogues and clinical misattribution in 3.7% of notes.supported
Basis: reported finding
- supportsAmbient clinical AI notes contaminated by unrelated background conversationsprimary disclosure
Abstract, dialogue-note evaluations
CLM-CP3-AMBIENT-CLINICAL-CONTAMINATION-2In 57 mock audio visits at -10 dB background speech, 48.2% of transcripts contained leakage and 5.3% of resulting notes were contaminated.supported
Basis: reported finding
- supportsAmbient clinical AI notes contaminated by unrelated background conversationsprimary disclosure
Abstract, mock recorded consultations
CLM-CP3-AMBIENT-CLINICAL-CONTAMINATION-3The study distinguishes transcript exposure from final-note contamination: 48.2% of mock audio transcripts contained background leakage, but 5.3% of generated notes retained it at the tested sound level.supported
Basis: reported finding
- supportsAmbient clinical AI notes contaminated by unrelated background conversationsprimary disclosure
Abstract, 57 mock consultations
Implications
The relevant trust boundary is the patient encounter, not whether the background words contain instructions. Validation must track speaker and encounter provenance from audio to transcript and final note; transcript errors and note contamination are different metrics.
Controls and mitigations
- Bind transcript spans and generated note claims to the active patient and authenticated encounter.
- Flag ambiguous or external speakers for clinician review rather than silently promoting them into the note.
- Test noisy multi-speaker settings while separately measuring transcription leakage and clinically consequential note misattribution.
Unknowns and contradictions
- Real-world incidence and effects on actual clinical decisions were not measured.
- How performance changes with microphones, room layouts, speaker overlap and clinician review remains untested.
Sources and citation
Material revision history
- Oct 13, 2026 · Published version · first publication · revision 137