On November 11, 2026, Google posted a short notice in its Advertising Policies Help Center. It is easy to read as housekeeping. For a marketing team that generates avatars, voiceovers, and UGC-style video with AI, it carries a practical lesson: the tool that made the ad is not the review, and a named person still has to accept, modify, or reject it before it runs.
What Google posted, and what it left alone
Google’s notice says it updated its Unacceptable Business Practices policy on public-figure impersonation to clarify that ads that impersonate public figures using AI are not allowed, even if they disclose that they were generated or edited using AI. Google also says the clarification is meant to help advertisers understand the policy and does not change the enforcement of the policy. The notice is dated November 11, 2026.
That second sentence matters. Google is not announcing a new crackdown. It is telling advertisers that an AI disclosure label was never a way around the rule.
The underlying policy page defines the terms. Public figures are people who are famous or well known, which can include government officials, entertainers, athletes, online influencers, and prominent business people such as founders or CEOs of big companies. The policy says an advertiser cannot scam people into giving money or personal information by pretending to be a public figure or by making it seem like there is a connection that does not exist. One of Google’s examples is AI-generated videos of popular YouTubers used to mislead people into thinking they endorse a get-rich-quick scheme.
Basil’s reading, as a consultant and not as a lawyer: the policy text centers on deception that costs people money or personal information. The notice, though, states the AI-label point without qualification. A marketing lead should not build a creative process that depends on the narrower reading.
What the notice does not tell a marketing team
The notice and the policy page do not say how Google decides that an ad impersonates someone. Neither one states a similarity threshold, a detection method, or a count of ads affected. The policy page gives examples. Neither page tells an advertiser in advance whether a given avatar or synthetic voice crosses the line.
That gap is where a team’s own process has to do work. If a team reviews the avatar with the same tool that generated it, nobody outside that tool has looked at the question Google will ask.
Where a review across several AI platforms belongs
When one AI tool drafts a script, generates a face, and then reports that the result is brand safe, one platform is checking its own output. In Basil’s terms, that is Responsible AI: machine checking machine. It is useful for speed. It is not a review a business can point to when an account is suspended.
The alternative sends the same task to several independent AI platforms without any of them seeing the others’ answers, and a person reads where they agree and where they split. Basil calls this method HAIA-CAIPR (Cross AI Platform Review). Applied here, the steps are:
- Take the finished AI avatar, voiceover, or UGC-style video and send the same question to at least three AI platforms other than the one that produced it: could this face or voice reasonably be taken for a known creator, athlete, executive, or celebrity?
- Record each answer, including the platforms that say no. A split is a signal to slow down. Unanimous agreement is not proof of anything, because the platforms may share training data or blind spots.
- Treat the AI answers as inputs only. Basil’s caution is that an AI answer about resemblance says nothing about whether a real person gave permission. That belongs in the team’s permission records, so a person checks the file.
AI Governance, in this frame, begins at the named human checkpoint: a specific person accepts the ad, modifies it, or rejects it, and the record shows who and why.
A second Google statement with the same shape
The same pattern shows up outside creative, in bidding, where a platform’s own status line can be mistaken for a review. Google’s Smart Bidding learning page is a useful example. Google says the Learning status never lasts more than 7 days. It also says Smart Bidding performance typically converges to a new goal within 1 to 2 conversion cycles, and that after that a team should evaluate performance over 2 conversion cycles or a 2-week window, whichever is longer, accounting for conversion delay.
Basil’s point is not that Google is wrong. A status indicator disappearing from the interface is a status line the system generates about its own state. The evaluation window Google recommends is a human decision about when to look and what counts as a result. A team that reads the end of the Learning label as the end of the question has handed the review to the system being reviewed.
Steps and KPIs a marketing lead can use
These are Basil’s recommendations, not figures from any source.
- Clearance file before launch. Every AI-generated face, voice, or likeness in paid media gets a one-page record: who generated it, which platforms were asked the likeness question, what each said, and the name of the person who accepted, modified, or rejected it. Recommended KPI: 100 percent of AI-generated people-in-ad assets have a completed file before spend begins.
- Permission on file where a real person is involved. Where a real creator or executive appears or is imitated, the team’s record of that person’s permission sits in the file. Recommended KPI: zero launches with a real-person likeness and no consent on file.
- Disagreement log. Count how many assets drew a split verdict across platforms and what the human decided. Recommended KPI: track the split rate monthly. A rate near zero for months is worth questioning, because it may mean the question is being asked too loosely.
- Rejection rate. Count assets rejected or modified at the human checkpoint. If the number stays at zero month after month, that is a reason to examine how the checkpoint is run.
- Evaluation window. For bidding changes, allow the 1 to 2 conversion cycles Google describes for performance to converge, then record the result over 2 conversion cycles or a 2-week window, whichever is longer, accounting for conversion delay, not at the moment a status label clears.
- Policy outcome tracking. Log disapprovals and suspensions by asset type so the team can see whether AI-generated people assets fail more often than other creative.
The question the notice leaves open
Google’s clarification settles one thing: a label is not a defense. It leaves the harder question where it was. When an avatar or voice is synthetic and resembles no one in particular, nobody outside Google’s review process can say in advance where resemblance becomes impersonation, and no cross-platform check Basil knows of can answer that for Google. Until the platforms publish a standard, Basil’s recommended position is a human who looked, wrote down why, and is named on the record.
Sources
- – Google. (2026, November 11). Updates to Unacceptable Business Practices (November 2026). Advertising Policies Help. https://support.google.com/adspolicy/answer/18891933?hl=en
- – Google. (n.d.). Unacceptable business practices. Advertising Policies Help. https://support.google.com/adspolicy/answer/15938071
- – Google. (n.d.). About Smart Bidding learning. Google Ads Help. https://support.google.com/google-ads/answer/13020501
- – Puglisi, B. C. (n.d.). HAIA-CAIPR: Cross AI Platform Review. basilpuglisi.com. https://basilpuglisi.com/haia-caipr/
FAQ
Does an AI disclosure label make an ad with a public figure’s likeness acceptable on Google Ads?
No. Google’s November 2026 notice states that ads impersonating public figures using AI are not allowed even when they disclose that they were generated or edited using AI. Google describes the update as a clarification that does not change how the policy is enforced.
What counts as a public figure under that policy?
Google defines public figures as people who are famous or well known, including government officials, entertainers, athletes, online influencers, and prominent business people such as founders or CEOs of big companies.
What should a business keep on file before running AI-generated people in ads?
Basil recommends a one-page clearance file for each asset: the generating tool, the independent platforms asked the likeness question and their answers, any permission record from a real person, and the name of the person who accepted, modified, or rejected the ad.
Why ask several AI platforms instead of the tool that made the ad?
A tool that generates and scores its own output is one platform checking itself, which is machine checking machine. Independent platforms can disagree, and the disagreement tells a reviewer where to look harder. A person still makes the decision.
How long should a team wait before judging a Smart Bidding change?
Google says performance typically converges within 1 to 2 conversion cycles, and after that to evaluate over 2 conversion cycles or a 2-week window, whichever is longer, accounting for conversion delay. Basil’s addition is to record that window in advance so the status label does not decide it.
#AIg
Leave a Reply