Gemini 3.8 Live Avatar: how to check a convincing AI video

What Google announced on 24 September 2026

Google introduced Gemini 3.8 Live with Live Avatar on 24 September, pairing conversational speech with generated video in Gemini Enterprise. Its announcement describes a responsive visual character, rather than just a prerecorded talking-head clip.

Google Cloud reports enterprise approval for custom avatars and SynthID-watermarked output. These are vendor claims, not our test results.

The important verification question is not whether every avatar is deceptive. A clearly identified digital assistant can be useful. The problem arises when a viewer treats a believable face as proof of who is speaking or of what that person is authorised to do. This article explains a review method; it does not report a confirmed scam involving the new release.

• Google’s launch announcement
• Google Cloud’s availability and safeguards

Three questions an AI video checker must not blur together

Start with the decision you need to make. A publisher deciding how to label a clip has a different problem from a finance employee being asked to change a supplier’s bank details. Both may have the same video in front of them, but a detector score cannot answer both questions.

• What DeepfakePolicy’s analysis can and cannot establish

• Generation: does the supplied file contain signals associated with AI generation or manipulation? This is the question for media analysis.
• Origin: is there a verifiable watermark or provenance record linking the media to a particular production process? That record has its own scope and limitations.
• Identity and authority: who controls the account, and may they make this request? Check this through a separate trusted channel and the relevant approval process.

SynthID is an origin signal, not an identity check

Google DeepMind describes SynthID as an invisible watermark embedded in generated media. Its guidance directs users to upload supported media to Gemini and ask whether Google AI created or altered it.

A positive watermark finding is useful context. It does not tell you whether the person sharing the file has permission to represent a company, whether their story is correct or whether a payment is safe. Conversely, a missing watermark is not proof of a human recording: the check may not cover that generator or that particular copy.

Keep the watermark result separate from a general AI video detector result. If they appear to disagree, preserve both findings and inspect what each actually measured. DeepfakePolicy is not presented here as a SynthID decoder, and we have not established its detection performance on this new model.

• Google DeepMind: how SynthID verification works

How to check a suspicious video, step by step

If the clip accompanies an urgent request, pause that request before examining the face. Do not approve a transfer, release an account or share a code because the speaker looks familiar. Then work from the best evidence you can legitimately obtain.

• Preserve a digital media evidence trail
• Use the AI video detector

• Keep the received file unchanged. Record its source, the time received and the message or claim attached to it. Use a copy for analysis.
• Ask for the original where appropriate. A screen recording, cropped repost or short extract may leave out the context you need; label it as such.
• Check available origin information. Save the actual watermark or provenance finding, including an inconclusive result, rather than paraphrasing it as “real” or “fake”.
• Analyse a supported video file for AI-related signals. Record the analysed file, the result and any coverage or quality limitations reported by the tool.
• Verify the request separately. For a business contact, use a number or account already on record, not the contact details supplied in the suspicious message.
• Document the outcome. State what the media analysis supports, what remains unknown and who approved the next action.

A live call and a recorded clip need different checks

During a live conversation, a natural answer is not an identity credential. Asking an unexpected question can add context, but it should not replace your organisation’s established identity checks or approval rules. Equally, a delayed answer, odd eye movement or poor lip synchronisation is not enough to accuse someone of using a deepfake.

DeepfakePolicy examines submitted media files. It is not a live-call interception service, a liveness certification system or a complete KYC process. If a permitted recording is reviewed afterwards, the finding concerns that recording. It cannot retrospectively authenticate the entire session or every participant.

What this changes for claims, marketplaces and customer support

Google Cloud demonstrates damage-information intake, not proof of genuine damage or fraud detection.

Consider a practical, hypothetical case: a customer shows a damaged item and asks for a refund. The caller’s appearance, the item’s condition, the purchase record and the entitlement to a refund are separate pieces of the decision. A reassuring avatar or fluent explanation should not collapse those questions into one.

The same applies to a marketplace seller asking for an off-platform payment, a new dating contact requesting money, or someone claiming to be a colleague. Verify the action being requested, not just the face delivering it. For teams reviewing media repeatedly, keep the analysis report attached to the case and retain the reviewer’s reasoning separately.

• Review image and video evidence reports
• Integrate file analysis through the API

The useful conclusion is narrower than “real or fake”

An AI avatar can be honestly disclosed; an ordinary camera recording can accompany a dishonest request. Better verification asks a precise question, preserves the relevant evidence and leaves room for uncertainty. Use the detector to investigate the media, an origin check to investigate its production, and a trusted process to verify the person and the action.