Skip to main content

AI Scribe Audio Privacy: Browser Memory, Uploads, Retries and Deletion

“Audio stays in browser memory” describes one possible location, not the complete privacy story. Map capture, upload, processing, retry copies, logs and deletion separately. Confirm what happens when recording stops, a request fails, a tab closes or the patient changes their choice. Keep the audio policy distinct from retention of the final clinical note.

By Bractis · Updated · 9 min read

Indian clinic owners, doctors and technical reviewers evaluating AI dictation or ambient-recording tools before using patient conversations.

Turn a storage slogan into questions about actual copies

A vendor may say that audio is not stored while meaning that it is not retained in the main application database. The same workflow may still hold an upload temporarily, retain a retry copy or send content to a processing service. Ask which location and which period the statement covers.

Distinguish audio from its derivatives. A transcript, generated note, debugging attachment and final signed record can each contain sensitive information even after the original sound file is removed. A policy covering only raw audio leaves the rest of the information journey unexplained.

Use ordinary questions the clinic can answer: what is captured, where does it go, who can access it, why is each copy needed and how is it removed? The technical team can supply evidence for browser, server and provider behaviour. The doctor should not have to infer a retention policy from a microphone icon.

Do not claim that browser-memory handling guarantees the absence of all traces or makes the whole service compliant. It can be one deliberate control within a wider design. Its value depends on the actual implementation, device environment, transmitted data and downstream handling.

Draw the lifecycle from capture to the final clinical record

Start with the microphone and follow the information through each transition. A short clinician dictation and a full ambient conversation may use the same button but capture very different material. Record the intended scope and the alternative available when recording is unsuitable.

For every stage, distinguish normal completion from failure. A request that times out can leave the clinic uncertain whether the provider received the audio. A retry feature may preserve data that would otherwise disappear. These are design choices that need explanation rather than assumptions.

The table is an original evidence map. Its entries are possible locations to investigate, not a claim that every scribe uses all of them.

On a narrow screen, scroll the table horizontally to read all columns.

Possible audio and text copies across a scribe workflow
StageInformation that may existEvidence to request
CaptureActive stream and buffered audio on the deviceStart/stop behaviour, session scope and what the interface reports
Upload and processingTransferred audio and provider-side handlingDestination, access, contractual retention and failure behaviour
Retry or recoveryA temporary copy or request referenceLocation, encryption where applicable, expiry and authorised recovery
TroubleshootingLogs, traces or support attachmentsWhether content is excluded or specially protected
Clinical recordReviewed text and relevant provenanceSeparate record retention, correction and access rules
DeletionRemoval from specified storage locationsWhich copies are covered, timing, exceptions and verification limits

Stopping recording is different from discarding data

Browser APIs make this distinction concrete. MDN documents that stopping MediaRecorder ends recording and makes recorded data available; stopping a MediaStreamTrack ends that track’s use of its source. Neither action establishes that an application has deleted audio already buffered or uploaded.

The clinic-facing control should therefore say what it does. “Stop recording” may end new capture while preserving the clip for review. “Discard” may remove an unsent clip through a different process. “Cancel request” may stop waiting locally without proving that remote processing stopped.

Test each control with fictional audio. Observe the recording indicator, whether another clip can be started and what happens after navigating away. Ask the technical reviewer to confirm the underlying state rather than accepting an animation as evidence that capture ceased.

Also examine multiple tabs or devices. Ending one track does not establish that another active capture session ended. The team needs a clear way to identify the session being controlled and to handle unexpected recording through its approved incident process.

Source: MDN: MediaRecorder stop() method

Source: MDN: MediaStreamTrack stop() method

Decide what temporary recovery is worth keeping

Keeping no recoverable audio can reduce the number of stored copies but may mean that a failed request cannot be resumed. Retaining a short-lived recovery copy can avoid repeating a dictation but creates additional custody and deletion responsibilities. State the trade-off explicitly.

Define the allowed purpose, location, access and expiry for any retained copy. “Encrypted” answers only part of that question. Ask who can decrypt it, which encounter it belongs to, whether a different user can recover it and what happens when the expiry process fails.

Do not use patient recordings as an informal quality-testing library. Any evaluation use needs its own appropriate authority and safeguards. For routine procurement testing, use fictional scripts or suitably approved material so the clinic can inspect failure states without exposing real consultations.

Keep the final clinical record policy separate. Deleting a temporary audio clip does not necessarily mean deleting a signed consultation note, and preserving a required note does not automatically justify keeping the original recording indefinitely. Obtain appropriate advice for the clinic’s actual obligations and purposes.

Test cancellation, timeout and shared-device recovery

Use a fictional clip to interrupt the workflow before upload, during processing and after a draft returns. For each point, record whether the system knows the request outcome, what remains recoverable and which action the doctor can take next.

A lost response is especially important. The provider may have processed the request even though the browser did not receive the answer. Repeatedly pressing retry can create additional requests or copies unless the system manages that state deliberately. Ask the vendor to demonstrate the actual behaviour.

On a shared workstation, sign out and use a different authorised test account. That account should not inherit another clinician’s recoverable clip simply because it uses the same browser. Check session boundaries, patient context and local recovery storage with the technical reviewer.

If the patient withdraws from further recording, stop new capture through the clinic’s agreed process and assess the handling of existing material appropriately. Do not promise immediate erasure of every downstream copy unless the system and applicable obligations support that statement.

Keep troubleshooting useful without copying the consultation everywhere

An error report needs enough information to diagnose a failed operation, but usually does not need the full consultation transcript in a general log. OWASP’s logging guidance identifies sensitive health information and secrets as data requiring exclusion or special treatment. Apply that principle to the actual support workflow.

Ask what appears in browser errors, server logs, analytics and support tickets. A friendly error message can still correspond to an internal trace containing the audio request or transcript. Technical access should be scoped and recorded according to the clinic’s approved arrangements.

Use request identifiers and controlled references where practical. They can support an investigation without repeating patient content in every event. Do not treat an identifier as anonymous if it can be resolved to a patient by the people receiving it.

The original outline below helps the clinic record a bounded claim and the evidence behind it.

Source: OWASP: Logging Cheat Sheet

Approve a precise statement the clinic can explain to patients

After the evidence review, write the clinic’s explanation in terms it can substantiate. For example, a statement about temporary handling by named services with specified retention is more useful than an absolute “nothing is ever stored” claim that nobody has tested.

Include the available manual alternative and who can answer a data-handling question. Staff should not improvise different assurances at each visit. The explanation should match the feature and configuration actually enabled in that clinic.

Revisit the map when a provider, recording mode, retry feature or support process changes. A new recovery option can change the number and lifespan of copies even if the patient-facing recording button looks identical.

Record what remains unverified and make a proportionate adoption decision. This article supplies an evaluation method; it does not establish that any vendor, including Bractis, uses a particular retention design or satisfies the clinic’s privacy obligations.

Worked example: a useful retry feature changes the storage claim

A fictional clinic is told that unsent audio exists only in browser memory. A new feature later allows a failed dictation to be recovered after a browser restart. The team asks where the recoverable copy now lives, which account can access it and how expiry is enforced. It updates its evidence map and patient explanation only after those questions are answered. Recovery may be useful, but it cannot be evaluated honestly while repeating the old, narrower storage claim.

Consultation Audio Lifecycle and Evidence Map

Use these prompts in your clinic’s own notes. Record the result, responsible person and next action for each item.

Copy inventory
Trace one fictional clip and its transcript through capture, processing, recovery, support and final records. Identify every known copy.
Control meaning
Test stop, discard and cancel separately. State what each action demonstrably ends or removes.
Recovery boundary
Check restart, account change and expiry with fictional material. Who can recover the clip and into which encounter?
Patient explanation
Write one plain-language statement supported by the observed behaviour and contractual information. Remove claims broader than the evidence.

Questions doctors ask

Does keeping audio in browser memory guarantee privacy?

No. It describes one location in one part of the workflow. Review uploads, processing services, recovery copies, logs and derivatives as well. Verify the actual implementation instead of inferring the full lifecycle from a storage slogan.

Does pressing Stop delete the recording?

Not necessarily. Stopping capture and deleting an existing clip are different operations. Confirm what the application’s controls do, including whether audio was already uploaded and whether any recovery copy remains.

Is an encrypted retry copy automatically acceptable?

Encryption is one control. The clinic still needs a defined purpose, access, duration, session boundary and deletion process, with appropriate authority for the handling. Assess the complete recovery workflow and its limits.

Should the signed note be deleted when temporary audio expires?

Audio retention and clinical-record retention are separate decisions. Follow the appropriate obligations and policies for each. A temporary clip’s expiry should not silently remove an authorised clinical record or erase its correction history.

What evidence should a clinic request before using a scribe with patients?

Request the configured data-flow map, processing and retention terms, role access, deletion scope and observed failure behaviour using fictional material. Record unresolved questions and ensure the patient explanation matches the verified feature.

Sources and scope

Updated on . Technical and operational evaluation guidance. It does not establish a vendor’s actual storage design, a universal deletion deadline or legal sufficiency of a clinic’s recording process. No independent clinical or legal review is claimed.

  • MDN: MediaRecorder stop() method

    Browser API behaviour for ending recording and delivering recorded data. It does not specify an application’s storage, server retention or deletion policy.

  • MDN: MediaStreamTrack stop() method

    Stopping one track ends that track’s use of its source; other tracks can still use the same device. Browser API state is not evidence of deletion of earlier copies.

  • OWASP: Logging Cheat Sheet

    Technical guidance on useful event records and sensitive data to exclude or protect. It is not a clinic retention schedule or a substitute for access controls.

Try the workflow
with Bractis.

Use a fictional recording in a Bractis demo to inspect capture, cancellation and recovery. Ask for the actual data-handling scope and evidence for each copy, including any external processing service.

Request a clinic demo