# Cisco biometric AI features and their governance Canonical: https://warmtransfer.net/knowledge/cisco-biometric-ai-features Last verified: 2026-09-05 Cisco's AI transparency technical note treats head detection and face recognition as 2 features[^11]. Head detection underpins cinematic meeting capabilities, whereas face recognition is invoked only for name labels[^11]. ## Feature design and technical operation The note's document address names facial recognition, while its title names head detection and face recognition[^13]. The face recognition model's stated input is a section of a video frame determined to contain a face, and its output is a vector compared against reference vectors[^5]. Head detection runs during camera modes and cross-view, which are available both in Webex meetings and in third-party meeting platforms[^12]. The privacy statement in the note states that no user data is retained, but this applies strictly to data at inference on the device beyond a call or a managed diagnostic session, leaving reference vectors unaddressed[^8]. Neither the word stored nor delete appears in the note, and it states no storage location, retention period, deletion path, or un-enrolment path for a face representation[^10]. Furthermore, the words biometric and consent do not appear anywhere in the head detection and face recognition note[^14][^15]. ## Governance and usage restrictions The head detection and face recognition note restricts usage using the words should not, stating that the features and models should not be used for security or any form of active surveillance or identity recognition beyond the scope of the meeting[^9]. It names no party bound by this restriction and specifies no consequence for disregarding it[^7]. In contrast, the video gesture recognition note twice uses the word cannot where the face note uses should not[^4]. The gesture note states that the model cannot be used outside Webex applications or Cisco devices, and that the feature cannot be used to translate sign language[^4]. The only appearance of the word consent in the gesture recognition note concerns training data donors—stating that real datasets are derived from internal Cisco employees with explicit consent—and does not address the consent of a person appearing on camera[^3]. ## Model evaluation and fairness claims The gesture recognition note claims that Cisco expects the feature to perform similarly regardless of the demographic of any data subject in the frame[^2]. The face recognition note makes no claim about output parity across populations; its fairness discussion addresses only the composition of training data and closes on a hedge[^1]. Neither note publishes an accuracy figure, and neither publishes an accuracy figure disaggregated by population[^6]. ## See also See also [The Cisco AI transparency note set beyond the contact centre](https://warmtransfer.net/knowledge/cisco-ai-transparency-census). See also [Enrolment storage and deletion for Cisco face recognition](https://warmtransfer.net/knowledge/bio-face-enrolment-lifecycle). See also [The Webex name labels feature that face recognition serves](https://warmtransfer.net/knowledge/webex-calling-name-labels). ## Applicability Applies to: Cisco Collaboration Devices. Sources checked 2026-09-05. ## What remains uncertain Whether consent is obtained for face enrolment and from whom is not covered by the sources below. Where a face reference vector is stored and for how long is not covered by the sources below. How a person is un-enrolled and what is deleted is not covered by the sources below. Whether an unenrolled participant's face is processed is not covered by the sources below. What becomes of a vector that matches no reference is not covered by the sources below. The accuracy of face recognition across populations is not covered by the sources below. Whether Cisco ships biometric features beyond these two is not covered by the sources below. What is processed and retained when the feature runs in a third-party meeting is not covered by the sources below. ## Sources [^1]: The face note makes no claim about output parity across populations; its fairness material concerns the composition of training data and closes on a hedge. Source: [Head Detection and Face Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-facial-recognition-technical-note.pdf), Fairness. Checked 2026-09-05. [^2]: The gesture recognition note states that Cisco expects the feature to perform similarly regardless of the demographic of any data subject in the frame. Source: [Video Gesture Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-gesture-recognition-technical-note.pdf), Fairness. Checked 2026-09-05. [^3]: The single use of the word consent in the gesture recognition note concerns training data donors, stating that real datasets are derived from internal Cisco employees with explicit consent, and does not concern the consent of a person appearing on camera. Source: [Video Gesture Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-gesture-recognition-technical-note.pdf), Data Sources for Training and Evaluation. Checked 2026-09-05. [^4]: The video gesture recognition note twice uses the word cannot where the face note uses should not, stating that the model cannot be used outside Webex applications or Cisco devices and that the feature cannot be used to translate sign language. Source: [Video Gesture Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-gesture-recognition-technical-note.pdf), Usage Guidelines. Checked 2026-09-05. [^5]: The face recognition model's stated input is a section of a video frame determined to contain a face, and its output is a vector compared against reference vectors. Source: [Head Detection and Face Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-facial-recognition-technical-note.pdf), Model Inputs and Outputs. Checked 2026-09-05. [^6]: Neither note publishes an accuracy figure, and neither publishes one disaggregated by population. Source: [Head Detection and Face Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-facial-recognition-technical-note.pdf), Model Evaluation in both notes. Checked 2026-09-05. [^7]: The note names no party bound by its restriction and states no consequence for disregarding it. Source: [Head Detection and Face Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-facial-recognition-technical-note.pdf), Usage Guidelines. Checked 2026-09-05. [^8]: The note's privacy statement that no user data is retained applies to data at inference on the device beyond a call or a managed diagnostic session, and does not address reference vectors. Source: [Head Detection and Face Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-facial-recognition-technical-note.pdf), Privacy and Security. Checked 2026-09-05. [^9]: The note restricts use with the words should not, stating the features and models should not be used for security or any form of active surveillance or identity recognition beyond the scope of the meeting. Source: [Head Detection and Face Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-facial-recognition-technical-note.pdf), Usage Guidelines. Checked 2026-09-05. [^10]: The note states no storage location, no retention period and no deletion or un-enrolment path for a face representation; the words stored and delete do not appear in it. Source: [Head Detection and Face Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-facial-recognition-technical-note.pdf), whole document. Checked 2026-09-05. [^11]: Cisco's note treats head detection and face recognition as 2 features: head detection underpins the cinematic meeting capabilities and face recognition is invoked only for name labels. Source: [Head Detection and Face Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-facial-recognition-technical-note.pdf), Feature Overview. Checked 2026-09-05. [^12]: Camera modes and cross-view run head detection and are available in Webex meetings and in third-party meeting platforms. Source: [Head Detection and Face Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-facial-recognition-technical-note.pdf), Feature Overview. Checked 2026-09-05. [^13]: The note's address names facial recognition while its title names head detection and face recognition. Source: [Head Detection and Face Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-facial-recognition-technical-note.pdf), document address compared with the title. Checked 2026-09-05. [^14]: The word biometric does not appear in the head detection and face recognition note. Source: [Head Detection and Face Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-facial-recognition-technical-note.pdf), whole document. Checked 2026-09-05. [^15]: The word consent does not appear in the head detection and face recognition note. Source: [Head Detection and Face Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-facial-recognition-technical-note.pdf), whole document. Checked 2026-09-05. [^16]: Where the gesture recognition note says its training data came from Cisco employees with explicit consent, the face recognition note says only that it uses internally collected Cisco volunteer videos, with no consent permission or licence language. Source: [Head Detection and Face Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-facial-recognition-technical-note.pdf), data sources section compared with the gesture note. Checked 2026-09-05. [^17]: Cisco's face recognition transparency note states that when the feature is enabled by an organisation, end users individually must opt in to use it. Source: [Head Detection and Face Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-facial-recognition-technical-note.pdf), feature description. Checked 2026-09-05. [^18]: Neither the transparency note nor any of the three product articles states where a face representation is stored or for how long. Source: [Head Detection and Face Recognition - AI Transparency Technical Note](https://www.cisco.com/c/dam/r/ctp/docs/white-paper/collaboration/ai-transparency-facial-recognition-technical-note.pdf), whole document set. Checked 2026-09-05.