Given the recent proliferation of artificial intelligence (AI) companion apps, such as Character.AI, policymakers, researchers, and journalists have sought to investigate and mitigate the potential harms. These same groups have also been concerned with the application of AI to generate nonconsensual intimate imagery (NCII), including through so-called “nudifier” apps that allow users, including those with no expertise in AI technologies, to upload images of anyone and generate semi-nude or nude images. However, little attention has been paid to the threat posed by the overlap of these systems. How the nonconsensual generation of intimate imagery is both enabled and encouraged in the AI companion app ecosystem must be considered in these discussions.
Likeness-based features in AI companion apps
In our preprint, “Examining Risks in the AI Companion Ecosystem”, we conducted an analysis of 30 apps selected from the AI companion ecosystem.
We found that, when creating a companion, in some cases users can upload photos to be used in companion profiles or as “reference photos” for AI-generated content. A few apps even offer a “face retention” feature that ensures the face in uploaded content is preserved in the resulting generated imagery. Users can also customize companions to resemble real people through text-based descriptors.
Once a companion is created, users can generate image or video content featuring that companion. Some apps also offer a variety of mechanisms to control this output, such as free-text prompts or toggle to make the output “spicy”. Some offer controls for bodily representations, including clothing, pose, and framing (e.g., “close up,” “full body”), as well as scene customization.
These companion creation and content generation functionalities raise serious risks related to the ingestion and use of people’s likeness, namely facilitating the nonconsensual use of someone’s image. For example, during our walkthroughs, we encountered companions modelled on real influencers, we recognized, which could constitute stolen labor by depriving creators of income they would otherwise earn through engagement on social media or other platforms.
Furthermore, the ability to generate content means users could produce AI-generated NCII. Some apps explicitly offer the generation of sexual or nude content, enabling the generation of NCII similar to nudifier apps. Moreover, what counts as “intimate” is highly contextual and cultural, so even content that is not sexual or nude may still be intimate and, in some cases, constitute NCII. Research has shown that in some non-Western contexts the bar for problematic content is lower. For example, imagery of people hugging or holding hands can constitute potentially damaging intimate content by insinuating an inappropriate sexual relationship, and survivors of this type of abuse have reported harms including coercion and death threats.
Advertising of likeness-based use cases
Our above findings are complemented by a subsequent investigation into the ecosystem’s advertising, the results of which we report here. We observed hundreds of advertisements on Meta’s platforms that explicitly advertised the capability to realistically model AI companions on real people’s likeness, including an “ex,” “roommate,” “gym crush,” and “colleague”, or, for 48 ads, “anyone from a photo.” For example, an ad states:
“I’ll send you photos and videos, whatever you ask for. Want me to look like an ex or like a colleague? Sure thing.”
In February 2026, we cataloged the information from Meta’s ad library for 146 ads that advertised this capability from 12 different accounts advertising 10 different apps or websites (ad catalog). 99 of these ads ran in either the EU or the UK, meaning the Meta ad library offers mandatorily reported figures about how many accounts each ad reached. Across the 94 ads ran in the EU there was a reach of over 1.1 million accounts. For the 78 ads that ran in the UK, over 260,000 accounts were reached. Our sample is far from comprehensive and was based on manually identifying and reviewing individual ads. For some accounts, there were hundreds of additional ads we did not review. Even within these limitations, the volume of ads we identified speaks to their ubiquity beyond our sample.
Only seven of these ads, all for one app, had been removed as of February 26th: six for “Adult Sexual Solicitation and Sexually Explicit Language” and one for “Unacceptable Business Practices.” We manually reported the 84 ads still running at the time of collection. Some reports were not reviewed because the ad had already been removed, either by the advertiser or by Meta. For the remaining reports that were reviewed, Meta responded that it chose not to remove the ads, despite some containing content identical to ads it had previously removed for policy violations. As of April 15, 2026, some of these ads were still running.
Putting it together
Taken together, our research into app functionality and advertised use cases paint an alarming picture. The convergence of companionship and nudification applications is further highlighted by a related investigation from The Indicator, which identified a network of companies that “along with nudifiers…appears to operate a group of at least 5 AI girlfriend/companion sites.”
These findings call for those working on NCII prevention and regulators to account for AI companion applications in their mitigation efforts. This is complicated by the fact that, unlike nudifiers, these apps advertise pro-social use cases (e.g., entertainment). Therefore, features that enable NCII creation are less easy to detect and it may require more justification to argue for their removal from app stores or advertising platforms. Given that Meta has explicitly committed to countering AI-enabled intimate image abuse in its March 2026 Adversarial Threat Report, targeting ads for both “nudifiers” and “kissing apps” that enable the generation of NCII, it must consider expanding its scope to treat companion app ads that encourage AI-enabled image abuse as in-scope violations, particularly given that our reports resulted in no action.
Likewise, those working on AI companion-specific harms should expand their focus beyond current mitigation efforts—which have largely centered on age-restriction, requiring disclaimers about AI-generated content, and directing users to crisis support resources when needed—to protect against the nonconsensual use of people’s likeness. A possible mitigation is to disable photo uploads until more robust safeguards to protect people’s likeness and deter NCII creation are in place.
Finally, the research raises deeper questions around likeness and AI-generated visual content. We must examine how and in what contexts different representations of likeness (e.g., photorealistic, cartoon) cause harm, how these representations might be leveraged by perpetrators to evade regulation and moderation, and how should we approach the use of an actor’s likeness in relation to a character they portray. Further, we must ask how the normalization of nonconsensual uses of likeness, in general, normalizes or acts as a gateway to creating NCII. Interdisciplinary exploration of these questions is essential to inform both technical and policy responses at the intersection of these ecosystems.