Whose Face Is It Anyway? The Legal Reckoning Over AI Training Data and Biometric Consent

For most developers, the excitement around face-swap technology centers on model performance — swap quality, latency, identity preservation. The dataset that made the model possible tends to feel like a solved problem, something that happened upstream, someone else's concern. That assumption is becoming increasingly dangerous.

A quiet but accelerating legal conflict is unfolding across the AI industry, and face-swap systems sit at its epicenter. Courts, state legislatures, and plaintiffs' attorneys are beginning to scrutinize not just what these models do, but what they were trained on — and whether the people whose faces appear in those training sets ever agreed to participate.

For developers and product teams building on face-swap technology, understanding this landscape is no longer optional. It is a foundational engineering and compliance requirement.

The Training Data Problem Nobody Wants to Talk About

Most large-scale face-swap and face-synthesis models are trained on images scraped from the public internet — social media profiles, film stills, news photography, YouTube thumbnails. The logic has historically been that publicly accessible content is fair game. That logic is eroding rapidly.

In Illinois, the Biometric Information Privacy Act (BIPA) has been the most aggressive instrument for holding companies accountable for unauthorized biometric data collection. BIPA defines a "face geometry scan" as biometric data and requires explicit written consent before collection. Critically, BIPA has already survived multiple legal challenges and has generated hundreds of millions of dollars in class-action settlements, most notably a $650 million settlement against Facebook in 2021 and a $100 million settlement involving Google Photos.

What developers often underestimate is that training a model on facial images may itself constitute biometric data collection under BIPA's framework — not just deploying the model. If your training pipeline processes facial geometry to extract embeddings or identity features, and those faces belong to Illinois residents, your exposure may begin before you ship a single line of production code.

Texas and Washington have enacted similar statutes, and at least a dozen other states have introduced biometric privacy legislation in recent sessions. The patchwork is expanding faster than most compliance teams can track.

Celebrity Likenesses and the Right of Publicity

Beyond biometric statutes, a separate but equally significant body of law governs the commercial use of recognizable faces: right of publicity doctrine. Every US state recognizes some version of this right, though scope and remedies vary considerably.

Right of publicity claims become particularly acute when face-swap models are trained on celebrity images scraped without authorization. The argument, increasingly advanced by entertainment attorneys, runs roughly as follows: if a model learns to synthesize a celebrity's likeness from unauthorized training data, and that model is then used commercially, the original subject has a cognizable claim for misappropriation of their likeness — even if the output is technically a novel image.

This theory has not yet been fully adjudicated at the federal level in the context of generative AI, but the groundwork is being laid. Litigation involving AI image generators and artists' copyrighted works is already working its way through federal courts. Face-specific cases are likely to follow a similar trajectory, with right of publicity claims potentially layered on top of copyright arguments.

Developers who have integrated pre-trained third-party models into their pipelines should not assume that liability stops at the model provider. Downstream users of a model trained on unauthorized likenesses may face exposure depending on how the model is deployed and monetized.

The Federal Void and Why It Matters

The United States currently has no comprehensive federal facial recognition or biometric privacy law. The American Data Privacy and Protection Act (ADPPA) has stalled repeatedly in Congress, leaving the field to a fragmented state-by-state framework. This creates a compliance environment that is simultaneously under-regulated at the federal level and over-complicated at the state level.

For developers building products that operate nationally, the practical implication is that the most restrictive applicable state law effectively sets the floor. A product used by consumers in Illinois, Texas, Washington, and New York must account for each state's distinct requirements around consent, data retention, and disclosure — and those requirements are not always compatible.

The absence of federal preemption also means that private rights of action remain robust in states like Illinois, where individual plaintiffs — not just regulators — can sue. This is precisely what has made BIPA so consequential: the litigation risk is not theoretical, it is actualized through class actions with eight- and nine-figure exposure.

What a Defensible Training Pipeline Looks Like

Given this landscape, developers building or fine-tuning face-swap models need to treat dataset provenance as a first-class engineering concern. The following framework reflects current best practices among legally sophisticated teams.

Audit your data sources. If you are working with a pre-assembled facial dataset, obtain documentation of how consent was gathered. Reputable academic datasets such as VGGFace2 and MS-Celeb-1M have both faced scrutiny and partial withdrawal over consent issues. Treat the absence of clear consent documentation as a red flag, not a technicality.

Implement consent-positive data collection. For proprietary training sets, build consent workflows that are explicit, granular, and jurisdiction-aware. A checkbox buried in terms of service is unlikely to satisfy BIPA's written consent requirement. Consent should specify that facial geometry may be processed for AI training purposes.

Maintain data lineage records. Document every dataset used in training, including version, source, consent basis, and any geographic restrictions. This documentation is your first line of defense in litigation and regulatory inquiry alike.

Evaluate third-party model providers rigorously. When integrating a commercial face-swap API or pre-trained model, request information about training data sourcing as part of your vendor due diligence. Vendors unwilling to provide this information represent a risk that should factor into your build-vs-buy calculus.

Implement deletion and opt-out mechanisms. Even where not strictly required by current law, providing individuals with the ability to request removal of their likeness from training data is both ethically sound and strategically prudent. Courts and regulators tend to view proactive compliance favorably.

The Litigation Horizon

Legal observers tracking AI-related intellectual property disputes broadly agree that face-specific litigation is coming in volume. The combination of increasingly capable face-swap technology, high-profile misuse cases, and a plaintiff's bar that has already developed successful class-action strategies around biometric data creates the conditions for a significant wave of claims.

For developers, the time to build defensible practices is before that wave arrives. Retroactively auditing training data after a complaint is filed is exponentially more costly than establishing clean data governance from the outset.

Face-swap technology represents a genuinely powerful class of tools with legitimate applications across entertainment, accessibility, education, and content creation. Protecting the ability to build and deploy these systems responsibly requires taking the legal infrastructure as seriously as the technical architecture.

The question of whose face powers your model is not merely philosophical. It is, increasingly, a question with legal consequences — and the developers who treat it that way will be best positioned to build sustainably in this space.