Shared AI Avatars Can Outlast Employment

On August 10, a learning manager in a Google Workspace domain can open Vids and find a new way to make the next safety video. The scheduled release of personal avatars began five days earlier. Google says the rollout can take up to 15 days.

A safety specialist scans a QR code with her phone, records a selfie and a short sample of her voice, and completes Google’s verification. She can then appear in a generated clip without standing in front of a camera. A prompt changes the background to the loading bay. A script explains a revised lockout procedure. Her face and voice deliver words she never recorded.

That saves a filming session. It also creates several records that can separate later: the selfie and voice recording, the account-linked avatar, the Vids document, the generated clip, and every published copy.

Now move the calendar forward. The specialist transfers to another division, the procedure changes, or she leaves the company. Deleting her avatar data stops new generation through that avatar. It does not necessarily remove the videos already created and published with it. Google’s current account guidance states that those existing items are not deleted with the recordings.

Another vendor is even more explicit about the employment boundary. Synthesia’s current avatar policy guidance says an avatar can remain in a company workspace after its original user leaves. If it was shared, other users can keep creating videos with it until access is removed. Previously generated videos continue to exist even after the avatar is deleted.

Creation consent matters. It is just the first decision.

Workplace avatars turn a person’s likeness into a reusable production asset while leaving the person, the account, and the finished content on different clocks. An administrator can close an account. An employee can withdraw permission. A communications team can still have exported files. A learning system can still serve an old module. A worker watching that module may assume the familiar speaker still holds the role, reviewed the current script, and stands behind the instruction.

The production capacity has arrived ahead of the employment process that has to govern it.

Google turns a selfie into a workplace asset

Google product manager Justin Luk announced personal avatars in Vids on July 16. His product post said users had created millions of Vids during the prior year. The new feature moved the product from preset presenters toward a user’s own identity.

In August 2025, Google had offered 12 preset avatars with built-in voices. At launch, users could make up to 20 avatar generations a week, each no longer than 30 seconds. Those presenters were reusable, but they did not claim to be an employee inside the company.

Personal avatars change that relationship. They look and sound like the account holder. Google restricts each one to that holder’s likeness and ties it to the holder’s Google Account. Gemini Omni can place the avatar into generated scenes, change visual context, and deliver typed speech.

Internal communications teams can localize an executive update without booking another studio session. Product experts can narrate a walkthrough after every release. Plant guidance can place a manager beside equipment that would be costly to film, while benefits teams can refresh enrollment instructions when a deadline changes.

Google also moved the feature into ordinary enterprise administration. Its Workspace release notice says personal avatars are on by default and can be disabled or enabled at the domain level. The rapid-release rollout began July 16. The scheduled-release rollout began August 5.

At launch, the feature is limited to English, to users aged 18 or older, and to eligible regions outside the European Economic Area, Switzerland, and the United Kingdom. Availability includes Business and Enterprise editions, Education Plus, several education and nonprofit plans, and Google AI Pro and Ultra.

Those controls determine who can reach the feature. Script authority, distribution, and the permitted life of an employee’s likeness remain company decisions.

A domain switch is too broad for those decisions. Turning personal avatars off for an entire company would block legitimate uses along with disputed ones. Leaving the feature on says nothing about whether a regional sales lead can be made to deliver a compliance message, whether an executive avatar can speak a translated script, or whether a former employee can still appear in an onboarding course.

A workplace avatar separates into at least five layers:

The source recording contains the person’s face and voice.
The avatar capability lets an authorized account generate new performance.
The project file contains prompts, scenes, scripts, and editing history.
The rendered video fixes one generated performance into a file.
Distribution creates copies in learning systems, shared drives, websites, downloads, and messages.

Each layer has its own access path. A creator may lose the ability to generate new clips while a learning administrator still has an exported MP4. A source recording can be deleted while a published course remains live. A Vids file can be removed while a copy sits in another content system.

Google gives each generated clip an invisible SynthID watermark. That supplies a provenance signal for detecting AI-generated media. Current ownership, the employee’s approval date, and script accuracy need separate metadata.

Viewers see the face. The source recording, permissions, owners, and copies stay behind the player.

Avatar vendors have built a strong gate at creation. HeyGen’s April consent instructions require a short consent video for every video-based Digital Twin. If one employee creates an avatar for another, the depicted person must record the consent personally. A QR code can send the recording step to that person’s phone.

Synthesia also requires the person represented by a personal avatar to record consent. Its current creation guide asks for a live consent recording and biometric consent. Google uses secure verification tied to the person’s account.

These checks reduce a serious risk: one colleague cannot quietly upload another person’s footage and create a working avatar without the subject appearing in the enrollment process.

After enrollment, reuse expands quickly. HeyGen lets one identity hold as many as 500 avatar looks. Its product guidance says the person records consent once for the first look. Later looks can change outfits, backgrounds, stance, camera angle, and setting without another consent recording, as long as they match the same identity.

The product logic is straightforward: verify who the avatar represents, then let a creator reuse that identity. The employer still has to define the business uses covered by the identity consent.

Consider a product leader who records an avatar for internal launch training. Six months later, another team wants the same avatar to deliver customer pricing, a workforce restructuring message, or a statement about product safety. The face still matches. The authority may not.

Consent has at least four scopes:

identity scope: permission to create a model from the person’s face and voice;
production scope: permission for named people or teams to generate videos;
message scope: permission for defined subjects, claims, languages, and roles;
distribution scope: permission for internal, customer, public, regional, or paid use over a defined period.

A live consent clip can prove the first scope. It may include parts of the others if the script is specific. It cannot prove that every future sentence, audience, and role was accepted unless the agreement actually covers them.

Employment makes the distinction sharper. A worker may feel free to decline an optional avatar for a team tutorial and less free when a manager describes it as part of the job. A subject-matter expert may accept internal reuse but reject a public sales campaign. A departing executive may allow old investor education to remain while refusing new statements in her voice.

A usable refusal path makes that choice observable. The employee can say no to a personal avatar without losing access to ordinary presentation work, and the team can offer a preset presenter, screen recording, voice-over, or another speaker.

Where possible, someone outside the direct reporting line records the choice. If the likeness is a stated job requirement, the employer should put the purpose, review duties, use period, and compensation in the role terms instead of treating a manager’s request as voluntary consent.

The consent record should state what it authorized, who can show that authority, and which change triggers another review.

Script approval belongs beside identity consent. Before release, the depicted person or a named approver should see the final words, visual context, language, audience, and publication date. If the company permits template-based production without approval of every script, the permitted categories and prohibited claims should be written down.

Role authority needs its own field. A safety specialist can explain a procedure because of a current job, not because her face was once enrolled in a generator. If she changes roles, the video may still be factually correct while its implied authority is stale. The company can add a visible title and review date, replace the speaker, or use a neutral synthetic presenter.

Compensation and work time also matter. Recording an avatar may take minutes. Reviewing the scripts it later delivers can become recurring labor. If the company expects the person to approve versions, translations, or public uses, that responsibility should sit in working time and ownership, not arrive as an informal request after the video is already scheduled.

Creation systems check whether the same human appears in the source and consent footage. Enterprise systems must check whether the same permission still covers the next use.

The employee leaves before the library

Synthesia’s newest guidance describes four roles around an enterprise avatar. The Talent supplies the face and voice. The Owner manages rights and permissions. A User creates videos. An Approver reviews content before release.

Those roles can belong to four different people. The employee whose face appears may not own the workspace asset. The communications producer who generates a clip may not approve the policy. The administrator who can delete the avatar may not know where finished videos were published.

This separation supports production. It also explains why normal offboarding misses the asset.

An IT checklist usually revokes login, removes group membership, transfers files, closes devices, and records data retention. A personal avatar needs those steps, but it also needs a decision about future generation and a search for existing content. Closing the account does not tell an L&D manager which modules contain the face. Transferring a Vids file does not tell the new owner whether the depicted person permitted continued use.

Synthesia states that avatars remain in a company workspace after a user leaves unless the user or an administrator deletes them. Shared users can continue generating videos after the original user leaves. The original owner can remove shared access during the avatar’s lifecycle, and administrators can manage permissions and deletion.

Previously generated videos remain because a rendered video is a separate content object, much like a recorded webinar that stays in a library after the speaker’s employee account closes.

Recorded webinars preserve words the person actually said at a date in the past. An avatar can create new performances after the source session. If generation access remains open, a former employee’s appearance can keep producing current-tense speech.

Offboarding should start with generation access before the person’s account or manager context disappears. Shared access should be listed. Content owners should receive a review task for every active video. The employee should know which uses will remain, which will be retired, and how to raise a later objection.

That handoff cannot wait for an exit interview. A content administrator needs time to identify the source projects, exports, LMS modules, and local owners while the employee and manager can still resolve ambiguous uses. The owner of each retained video then accepts responsibility for its script, review date, audience, and withdrawal path. File transfer without that acceptance merely moves an orphaned asset to a new account.

Role changes deserve the same mechanism. Suppose a chief medical officer becomes an adviser, a plant manager moves to another site, or an engineer no longer owns the product described in a customer tutorial. The person has not left the company. The apparent authority in the video has changed.

Review dates can catch that drift. Each published item should carry an owner, a depicted role, a script version, a policy or product dependency, and a next-review date. A change in job title, reporting line, jurisdiction, procedure, or public claims should reopen the item.

Stock avatars expose a parallel lifecycle. Synthesia’s retirement FAQ says an actor or voice provider can withdraw consent for future use. The vendor then retires the stock avatar or voice. Existing generated videos remain accessible and functional, but future edits require a replacement.

Employers can adopt the same separation for personal avatars: stop generation, preserve defined historical items, replace stale modules, and remove disputed or misleading copies.

Discovery becomes the harder part. A face can appear in a Vids document, an LMS module, a downloaded file, a sales portal, a support article, a screen recording, or a presentation assembled from clips. File ownership alone will miss some of them.

Logging the generation event before distribution turns offboarding into a bounded review. Without an inventory, the company knows an avatar existed but has to search folders and memories to learn where it spoke.

Deletion has three different meanings

Google lets users retake or delete personal-avatar data through their account. Its Vids help page says the change also updates other services that received access to the avatar, including Gemini.

Google’s broader account guidance draws a precise boundary. Deleting the recordings prevents new AI content from being generated with that avatar and removes the captured selfie and voice data from Google’s systems. Videos or content already created and published with the avatar are not deleted automatically. The user must go to the relevant service to remove them.

Synthesia publishes a time boundary. A request to delete a specific avatar, video, or voice can take up to 90 days to remove all copies from its systems. Its deletion guidance says the company provides written confirmation. If an account ends without a specific request, associated avatar, video, and voice data are automatically deleted within 90 days.

These statements use the word deletion for different objects. An enterprise policy should not collapse them.

Avatar deletion removes future generation capability and source data according to the vendor process. Video deletion removes a rendered item from a named platform. Publication withdrawal removes a copy from an audience. None guarantees that downloaded or independently stored copies are gone.

A fourth action is access revocation. It stops named users from creating or editing. That can happen without deleting the avatar or any video. A fifth action is retention: the company deliberately keeps an item for training, legal, or historical reasons while preventing new use.

Automatic removal at the moment of departure would create a different operational failure. A validated safety module could disappear in the middle of a required training cycle. A customer tutorial could lose the version that explains what buyers were told. An audit file could no longer reproduce the material employees actually saw. Content teams also spend money on translation, captions, accessibility review, and learning-system integration after a clip is rendered.

Persistence is therefore useful when it is explicit. The company can freeze new generation, preserve an immutable historical copy for a defined reason, and keep an approved module visible until a replacement date. Those are separate choices. The vendor cannot infer them from an HR status change, and the former employee should not have to guess which choice the company made.

Put the vendor positions side by side and the gaps become visible:

Product surface	Creation gate	Persistence signal	Deletion boundary	Company decision still needed
Google Vids personal avatar	Secure verification through the user’s Google Account	Existing created and published content is not removed when avatar recordings are deleted	Source recordings, avatar use, Vids content, and published copies require separate actions	Script approval, content inventory, role-change review, and copy withdrawal
Synthesia custom avatar	Personal and biometric consent, with owner and sharing controls	Shared users may continue after the original user leaves; generated videos remain after avatar deletion	Specific deletion can take up to 90 days; account termination follows a separate process	Who removes shared access, which videos remain, and who receives deletion confirmation
HeyGen Digital Twin	The depicted person records a consent video; one identity can support up to 500 looks	One creation consent supports later looks of the same identity	The reviewed creation guidance does not itself define an employer’s offboarding inventory	Allowed scripts, audiences, reuse period, exports, and exit procedure

This is a lifecycle comparison, rather than a vendor ranking. Each product documents a different part. The buyer must connect those parts to its accounts, content systems, employment processes, and publication channels.

Provenance creates another boundary. Google says every generated Omni clip includes invisible SynthID so people can verify it was AI-generated. The European Commission’s July 20 guidance announcement separates machine-readable marking by providers from disclosure duties for deployers in defined situations. Those transparency rules began applying on August 2.

An invisible mark and a visible notice do different jobs. A machine can help detect origin. A worker watching mandatory training may need an immediate label that the presenter is generated, the depicted person approved the use, and a current human owner stands behind the instruction.

Whether a particular internal video falls under a legal definition depends on the jurisdiction, content, system, and use. A company should not infer compliance from a watermark alone. It can apply a simpler editorial rule across regions: disclose generated presentation when the realism or identity could change how a reasonable viewer interprets the message.

Finding a stale safety clip after 4,000 workers completed it leaves a correction problem even if the file is removed at once. The useful record includes exposure: where it appeared, when it was active, who viewed it, and whether a correction must reach them.

Deleting the object leaves a second task: tracing its effect.

A familiar face changes learner trust

Mike Reichert, Thorsten Jungmann, Ute von Jan, and Urs-Vito Albrecht published a small randomized crossover feasibility study in June. Thirteen undergraduate engineering students watched two content-identical videos about fuel cells. One used a Synthesia avatar; the other used a human presenter.

Both formats produced short-term learning gains. The group seeing the avatar first gained a median of five newly correct answers on a seven-item test, compared with 4.5 for the group seeing the human first. The difference was not statistically significant, with P=.51. Participants rated the human-presented version higher across the user-experience dimensions.

For an employer, that comparison matters more than rendering speed. Personal avatars may reduce production friction, while learning, calibrated trust, and correct action still require measurement.

Reichert and his colleagues warn against treating that result as proof of equivalence. The sample was 13 students. The task measured immediate learning on one technical subject. The second exposure had carryover and test-retest problems. It cannot establish equal performance across compliance, onboarding, safety, sales, or leadership communication.

It still offers a better buying question. Hold the script and visual material constant, then compare comprehension, delayed recall, user experience, and action. Production time is an input. Learning is the output. A second 2026 paper, Face value, tested how avatar identity changes trust in AI-mediated learning. One experiment included 102 participants and another included 294. The researchers measured credibility, warmth, competence, willingness to act, and whether learners incorporated the guidance into their own work.

Race, gender, age, and subject context changed those judgments and behaviors. Learners responded to social identity and stereotypes about expertise rather than rewarding realism uniformly.

An employee avatar adds another cue: organizational authority. A familiar face can tell the viewer who supposedly owns the message. If the depicted person is a respected safety lead, finance chief, or product architect, employees may assign weight to the guidance before checking its date or source.

Familiarity can help a current expert teach. It can also lend weight to stale or unauthorized speech.

Measure calibrated trust alongside approval. Ask whether viewers knew the clip was generated and who they believed approved the script. Test whether they can find a human contact for exceptions. Measure whether they follow the correct procedure and recognize situations outside the video’s scope.

For safety training, a completion rate is especially weak. A viewer can finish a three-minute module and retain the wrong sequence. A stronger test samples delayed recall, observes task performance, records near misses, and tracks questions by module version.

Onboarding teams can measure time to independent work and escalation quality. Benefits teams can track correct enrollment decisions and preventable support cases. A product-training test can ask whether people complete a task without repeating unsupported claims. Each use needs a result tied to the work.

Visible disclosure can support that measurement. A label such as “AI-generated presentation using an authorized avatar; script approved by Safety Operations on August 6, 2026” gives the viewer three facts: the media type, the responsible team, and the review date. It avoids pretending the person recorded every sentence while preserving a named chain of responsibility.

Sometimes a neutral synthetic presenter is the better choice. If the message is a high-volume procedural update with no need for personal endorsement, using a face that does not belong to an employee can reduce offboarding and authority confusion. A real likeness earns its extra complexity only when the relationship with the audience improves a measured result.

An avatar use and exit record

Keep the company record lean. The source selfie and voice belong in the vendor and identity systems under their own controls. The internal row connects permission, generation, publication, results, and exit.

One row should describe one avatar and one authorized use program. Each generated video then links back to that row and carries its own script, owner, and distribution status.

Field	Minimum record	Decision it supports
Talent	Person whose face or voice is represented, with current employment or external-provider status	Who can approve, revoke, or receive notice
Business owner	Team accountable for the use program and its budget	Who answers when the depicted person changes role
Users and approver	Accounts allowed to generate, plus the person who approves final scripts	Separation of production access from message authority
Creation consent	Vendor, date, method, identity verified, and stored agreement reference	Proof that the avatar was created with permission
Allowed purpose	Named subjects, roles, and content types	Whether the next script fits the original use
Audience and region	Internal groups, customers, public channels, countries, and age limits	Distribution control and regional review
Language and voice	Approved languages, translation method, pronunciation owner, and voice-clone scope	Whether localized speech remains authorized and accurate
Use period	Start date, review date, and expiry or renewal rule	Prevention of permanent permission by default
Generated-item inventory	Video ID, script version, generation date, depicted title, content owner, and source project	Discovery during update, dispute, or exit
Distribution inventory	LMS module, shared drive, website, campaign, download, and downstream owner	Withdrawal beyond the generation platform
Provenance and notice	Machine-readable mark, visible label, named human owner, and approval date	Viewer understanding and audit
Learning result	Completion, comprehension, delayed recall, error, escalation, trust, and action measures	Whether the avatar improves the work
Change triggers	Role move, policy update, product change, language addition, new audience, or new claim type	When approval must reopen
Offboarding action	Generation disabled, sharing removed, files transferred, active videos reviewed, employee notified	Coordinated exit across IT, HR, L&D, and communications
Revocation and deletion	Request date, object requested, vendor ticket, due date, confirmation, copies retained, and reason	Proof of what stopped and what remained

One consent file cannot carry every future use. A linked inventory also stops a deletion confirmation from being mistaken for a complete takedown.

Ownership should follow the message. A safety operations team owns a safety script. HR benefits owns enrollment guidance. Product owns a customer walkthrough. Internal communications may produce the media, and IT may control the accounts, but neither should approve a technical claim outside its remit.

Likeness approval stays with the depicted employee. That person can flag a script that misstates the role or personal endorsement. The business owner remains responsible for the underlying instruction.

Generation starts a new item in the record: video ID, script hash or version, and intended destinations. Before publication, the approver checks the final render rather than the text alone. Visual context can create a claim: a uniform, laboratory, customer logo, or piece of equipment may imply access and endorsement that the words do not.

Each destination gets a responsible owner and an expiry. Downloads need special treatment because the originating platform cannot withdraw them. The recipient agreement or internal policy should require removal when the source item is retired.

A role change pauses new generation and opens all active items. Some can remain with a revised title and date. Others need a new speaker or an archive because the old role was central to the message.

Exit splits the work. IT removes account and group access, while the avatar owner or administrator removes sharing and future generation. Content teams review the inventory. The person receives a record of intended retained uses and a contact for disputes. Vendor deletion runs only after the retention decision identifies the exact objects.

The response clock completes the process. Synthesia may take up to 90 days to complete certain deletions, while an internal publication can usually be disabled sooner. Backend deletion should not leave a disputed module visible to employees.

Quarterly review is enough for a small library. A large program can automate alerts from HR role changes, content expiry, and vendor access logs. Automation should produce a queue for named owners, not silently delete safety, legal, or historical material.

Cost becomes measurable once the record exists. The company can compare filming and editing time saved with avatar setup, consent administration, script review, content inventory, learner testing, corrections, and offboarding. A cheaper render can still create a more expensive lifecycle. A more expensive personal avatar can earn its place if it improves learning or reduces update delay without losing trust.

Filming is one line in a program budget.

Monday’s safety briefing needs an owner

Return to the loading-bay video. Suppose the company generated 18 clips with the safety specialist’s avatar. Twelve sit in active LMS modules, three appear in supervisor toolkits, two were downloaded by regional sites, and one remains in a Vids draft. These counts are hypothetical, but they show why an avatar total is not a content inventory.

Friday is the specialist’s last day. Her exit record opens before the account closes.

Safety Operations reviews the 12 active modules. Nine scripts remain accurate and fall inside her documented internal-training permission through the end of the year. Three describe equipment that has since changed. Those are unpublished immediately.

Internal Communications removes generation access and confirms that no public or customer use was authorized. IT records the account action. The regional sites receive the two file IDs and a withdrawal date. The specialist sees the retained list, the visible label, and the contact who can act if another copy appears.

Content owners first decide what may remain and stop the uses that should end. The vendor process then handles the avatar data and named files.

On the following Monday, a worker begins the loading-bay module. The screen states that the presentation is AI-generated from an authorized employee avatar. It names Safety Operations as the current owner and shows the script review date. A link reaches the shift supervisor for an exception.

The familiar face still speaks. Beside it, the current owner and review date tell the worker who is responsible now.