How an AI companion app compares with a licensed likeness marketplace on consent, control and compliance. An honest look at what each is for.
Photo via Unsplash
Search interest in the AI companion app category is large and stable — roughly 4,400 US searches a month on August 2026 Keyword Planner data, with a competition index of 17. Interest in the adjacent "AI girlfriend generator" runs at 14,800 a month. That is a substantial audience, and a good chunk of it is now asking a question the category was not built to answer: whose face is this, and did they agree?
This comparison sets out what companion apps are genuinely for, what a licensed likeness marketplace does differently, and where the two get confused. We build the second kind, so the interest is declared — the comparison is drawn on criteria you can check.
Key takeaways
- Companion apps built on invented characters raise no likeness question at all. That is a feature.
- The problems begin when a persona resembles a real person without an agreement.
- A licensed marketplace's product is the consent record, not the imagery.
- Test any consent claim with three questions: readable terms, named refusals, working revocation.
- The licensed catalogue is smaller. Pretending otherwise would be dishonest.
An AI companion app provides an ongoing conversational relationship with a persona. Usually invented — a name, a personality, an avatar that resembles nobody in particular. The product is the relationship, and the business model is engagement.
A licensed likeness marketplace provides imagery of real, identifiable people who have agreed specific terms. The product is the permission. The imagery is what the permission produces.
Those are genuinely different products with different customers, and most of the confusion in this space comes from treating them as competitors. They mostly are not. A companion app with wholly invented characters has no likeness problem, no consent problem and no publicity-rights problem, because there is nobody to consent.
Everything turns on one property: is the persona based on a real person?
If no, the interesting questions are about age assurance, content standards, and the wellbeing design we come to below. Real questions, but not consent questions.
If yes, an entire regime attaches. Right of publicity in many US states. Intimate-image offences where the content is sexual. Platform removal duties under the TAKE IT DOWN Act. Age-assurance and reporting duties under the Online Safety Act 2023. We set the full list out in the deepfake law round-up.
The awkward middle case is a persona that is not named as a real person but is recognisably modelled on one. Legally that is a poor place to be, since identifiability rather than naming is generally the test.
| Dimension | Typical companion app | Licensed likeness marketplace |
|---|---|---|
| Persona basis | Invented character | Real, consenting creator |
| Consent required | None (nobody depicted) | Yes, explicit and scoped |
| Terms readable before use | Usually not applicable | Yes — public boundary set endpoint |
| Refusal behaviour | Content filter, generic message | Named rule, quoted values |
| Revocation | Not applicable | Creator-initiated, affects issued output |
| Output record | Usually none | Signed manifest per output |
| Creator earns | No creator | Metered plus subscription split |
| Primary business model | Engagement and retention | Per-generation and subscription |
| Catalogue size | Effectively unlimited | Limited by creators onboarded |
Table: dimensions on which the two categories genuinely differ. The last row is the honest disadvantage of the licensed model and is not going away soon.
Three ways, all worth watching for.
Persona drift. An app launches with invented characters, then adds "inspired by" personas that are recognisably real people. The consent obligations arrive without anyone updating the terms.
User-uploaded reference images. A companion app that accepts a photo to shape an avatar has just made every user a potential source of unlicensed likeness. The upload is where the category changes.
Claimed partnerships without enforcement. "Officially partnered with creators" can mean a revenue-share agreement with no operative limits at all. A partnership is not a boundary set unless the limits are enforced at generation time.
Whatever a platform claims, these three answers tell you what the consent model actually is:
1. Can you read the terms before generating? On LikeWard, GET /api/v1/creators/{slug}/boundaries returns the permitted contexts, forbidden depictions and realism ceiling as structured data, before you spend anything. If a platform's rules only become visible when you break one, they are not really terms.
2. Does a refusal name the rule? A refusal reading realism_exceeds_ceiling with both values quoted is auditable. "Your request could not be completed" is a filter with a press release attached.
3. Does revocation reach content already made? The important one. Stopping future generation is table stakes. A creator who withdraws consent needs the licence on existing output to flip too — otherwise every copy still resolves as authorised. We shipped this wrong ourselves at first and wrote up the fix in content provenance explained.
Worth being straight about, because it applies to us as much as to anyone.
Companion apps make money from engagement. So does a marketplace selling generations. Both business models reward more usage, and neither business model has any natural reason to encourage a healthier relationship with the product.
LikeWard's partial answer is to make it a creator-controlled lever rather than a platform promise: creators can set session caps and variety nudges, enforced on the same server-side path as the content rules, so it is a mechanism rather than a wellbeing paragraph in the terms. It ships on with a soft nudge.
That default is a compromise and we would rather name it than dress it up. A hard block would serve users better and would reduce revenue, including the creator's. Putting the dial in the creator's hands does not dissolve the tension; it just puts it somewhere more legitimate than our roadmap. The evidence and the incentives are in the engagement data study.
Use a companion app if you want an ongoing conversational relationship with an invented character. That is what they are built for and a licensed marketplace is a poor substitute.
Use a licensed marketplace if you specifically want imagery of a real person and you care whether they agreed. The catalogue is smaller — meaningfully smaller — and every image carries a record showing what authorised it.
Avoid anything in between: platforms offering real-person personas without readable terms, named refusals or working revocation. That is the configuration with all of the licensed category's obligations and none of its controls, and the risk analysis explains where it usually ends.
More: the creator-owned companion use case, how to license your likeness for creators, and the legal status of AI girlfriend generators for where the wider category stands.