AI influencers and virtual creators are computer-generated personas that post, appear in ads and sometimes reply to comments, controlled by a studio or brand rather than a real person. In client campaigns they work best as a recurring brand character or a fast way to test creative angles, provided the audience is told the persona is synthetic and a human team approves every post.
The idea is older than generative AI. Lil Miquela, created by the studio Brud, has posted on Instagram since 2016, and Imma in Japan and Shudu, often described as a digital supermodel, followed. What changed recently is cost. Image models, video tools such as Runway and Pika, voice synthesis and avatar platforms like Synthesia and HeyGen mean a small team can now produce a consistent virtual face without a 3D studio.
That lower barrier invites rushed projects. This article explains the decisions that come before the first render, and how we structure the work.
What counts as a virtual creator?
The term covers several different things, and mixing them up causes poor briefs.
- Fully virtual personas: a fictional character with a name, a backstory and a feed of its own, like the examples above.
- Brand mascots built with AI: an existing mascot given a voice, a face that moves and the ability to appear in short videos.
- Digital twins of real people: an avatar of a founder or presenter, licensed by that person, used to scale video in several languages.
- Synthetic extras: AI-generated people in product shots or ads, who are not presented as creators at all.
Each type has different rules. A digital twin needs a written licence from the real person covering where and how long the likeness may be used. A fully virtual persona needs a character bible. Synthetic extras mostly need disclosure and a check that they do not resemble a real individual.
When does a virtual creator make sense?
They suit campaigns where consistency and control matter more than lived experience. A skincare brand cannot credibly have a virtual creator review how a cream feels on skin, because the persona has no skin. A fashion label showing a styling idea across thirty outfits, or a software company explaining features in six languages, is a much better fit.
They also suit testing. Producing ten variations of a hook with a human creator means ten shoots or a long day of retakes. With a virtual presenter, the team can change the opening line, the setting or the language and compare early results before committing budget to a human-led version.
They fit poorly where trust is the product: health claims, financial advice, testimonials and anything that implies personal use. Advertising rules in many markets, including UK CAP code guidance and the US FTC Endorsement Guides, focus on whether consumers are misled, and a synthetic person “recommending” a product they cannot use is close to that line.
What rules apply to disclosure?
Platforms have added their own labels. Meta, TikTok and YouTube all ask creators to mark realistic AI-generated or altered content, and TikTok and YouTube offer a toggle at upload. In the EU, the AI Act introduces transparency duties for deepfake-style content, and Turkey’s advertising board applies its general rules on misleading ads and influencer marking to sponsored posts.
Our working standard is simple. The profile bio says the persona is virtual. Every sponsored post carries the platform’s paid partnership label and the AI label where one exists. Captions avoid first-person claims about physical experience.
How Alien Road applies AI influencers and virtual creators
A typical engagement runs in five stages, and the first two take longer than clients expect.
Discovery
We start by asking whether a virtual creator is the right tool at all. The questions are practical: which platform, which audience, what the persona is allowed to say, and what a human creator would do better. Sometimes the answer is a hybrid, with a human creator for reviews and a virtual character for explainers.
Character bible
Before any image is generated we write a one-to-two page document: name, age range, tone of voice, topics the character covers and avoids, visual rules (palette, wardrobe, settings) and a list of phrases it never uses. This keeps output consistent across designers and tools, and it gives the client something concrete to approve.
Production setup
The toolchain depends on the format. Still images often use a fixed reference set and image-to-image generation to keep the face stable. Short video may use an avatar platform for talking-head pieces and an image-to-video tool for b-roll. Voice is either licensed synthetic voice or a real voice actor. Every asset passes a review for hands, text artefacts and likeness to real people before it goes into the content calendar.
Publishing and community
Posts go through the same approval flow as any client content. Comment replies are the sensitive part: we do not let a model reply automatically on the character’s behalf. A community manager writes replies in the character’s voice, using the bible, and escalates complaints to the client.
Measurement
Results are compared against a baseline, ideally a human-creator or standard brand post running at the same time with similar spend. We report reach, watch time, saves, link clicks and cost per result, and we read the comments for sentiment about the persona itself, since a character that annoys the audience can lift clicks while hurting the brand.
What usually goes wrong?
The common failures are predictable. The face changes between posts because nobody locked the reference images. The character says something off-brand because the caption was generated without the bible. Engagement is reported without a baseline, so a decent result is presented as a breakthrough or a normal dip as a failure.
- Lock references and keep a versioned asset folder.
- Keep a human approver on every post and every reply.
- Decide in advance what result would end the test.
- Plan an exit: what happens to the character, its followers and its assets if the campaign stops.
Where to start this week
Write a single page that answers three questions: what the virtual creator will do that a human cannot do as well, which platform it lives on, and what disclosure it will carry. If those answers are clear, draft a short character bible and produce five test posts for internal review only.
Show those posts to a few people from the target audience and ask what they think the character is. If they cannot tell it is virtual after reading the bio, fix the disclosure first. If they find it irritating, rethink the character before spending media budget on it.
Keep reading: Social Media Management · Social Media Marketing