AI, deepfakes, and the synthetic adult content debate (2026)

The conversation about generative AI in adult content has matured considerably since the first wave of consumer deepfake tools appeared in late 2017. What was initially framed as a niche concern about celebrity face-swap forums on Reddit has, over the intervening years, become a structural issue that touches platform policy, content provenance standards, criminal law in several dozen jurisdictions, and the production economics of the legitimate adult industry itself. The technology has improved monotonically; the response has been unevenly distributed.

Where the research literature actually is

Deepfake detection as a research field has produced a steady stream of papers since the FaceForensics++ benchmark established a usable shared dataset in 2019. The arXiv corpus on synthetic-media detection has grown from low-hundreds per year in 2019–2020 to several thousand annually by 2024–2025, with detection-versus-evasion now treated by researchers as an explicit adversarial loop. The honest summary of the literature is that detection lags generation by approximately one model generation: every major detection technique works on outputs from the previous-generation model family and degrades on outputs from the current one. This is not a failure of the detection field — it is a structural consequence of generators being optimised, sometimes directly, against the detectors.

Provenance standards have proceeded in parallel and on a slower timescale. The Coalition for Content Provenance and Authenticity (C2PA), formed in 2021 from a merger of Adobe's Content Authenticity Initiative and the Project Origin coalition, has produced a workable specification for cryptographically signed media manifests. Adoption has been concentrated in journalism, stock photography, and increasingly in default camera apps from major manufacturers. Adult content platforms have largely not implemented C2PA at the ingestion layer, which is one of the structural reasons synthetic adult media remains harder to detect at the platform level than synthetic news imagery.

The non-consensual intimate imagery (NCII) problem

The most pressing harm category in this space is non-consensual intimate imagery — pornographic content created using a real person's face or body without their consent. NCII predates generative AI by several decades (the harm has been documented since the 1990s) but generative tools have collapsed the production cost and skill barrier so far that the historical analysis has had to be redone. Several research groups, including Sensity AI and the Internet Watch Foundation, have published volume estimates suggesting tens of millions of synthetic intimate-image files in circulation by the mid-2020s. Independent verification of these counts is difficult by construction.

Legal responses have varied. The United Kingdom's Online Safety Act (2023) and a series of US state-level statutes (Texas SB 15, Virginia §18.2-386.2, California AB 602, and others) criminalise the creation and distribution of NCII regardless of whether it depicts a real or synthetic body, with the synthetic case explicitly enumerated. The European Union's Digital Services Act creates platform-level obligations to act on NCII reports within defined timeframes. The enforcement gap is wide, particularly across jurisdictions, and most reporting infrastructure still depends on victim-initiated takedown requests.

What this means for legitimate adult studios

The legitimate adult industry — meaning studios that produce content with paid, consenting, documented adult performers under written contract — has been pushed by the AI conversation into a position it did not initially seek: it now competes with synthetic media for viewer attention, and it has a structural interest in being clearly distinguishable from it. Several industry trade groups (the Free Speech Coalition in the US, the European Adult Producers Association in the EU) have issued position statements affirming the value of documented-consent production and pushing for ingestion-level provenance signals on distribution platforms.

Step Secrets is a directory layer over verified-human-performer catalogs sourced from independent studios and distributors. No part of the catalog is generated by image or video synthesis models, and the directory does not accept submissions from any synthesis pipeline. Visitors interested in the structural distinction can browse the catalog directly to see how a non-synthetic directory is organised — by performer, by studio, by genre and by date.

Open technical questions

Three technical questions remain unresolved at the time of writing. First: whether a detection system that works on present-generation diffusion outputs can be made robust to the next architecture shift (video diffusion is already moving away from the 2023–2024 temporal-coherence approaches). Second: whether C2PA-style manifests can survive standard platform transcoding pipelines, which currently strip metadata aggressively. And third: whether passive watermarking schemes built into model outputs (Google's SynthID, OpenAI's image-watermark research, Meta's Stable Signature collaboration with INRIA) actually survive the kinds of adversarial transformations users perform routinely (crop, recompress, screenshot, re-encode through a different codec).

Open policy questions

On the policy side, three further questions are still being negotiated. The first is what platform liability looks like for synthetic content that depicts a real person without their consent: Section 230 in the US, the safe-harbour provisions of the EU's e-Commerce Directive (now folded into the DSA), and equivalent provisions elsewhere were drafted before generative AI existed and apply only awkwardly. The second is what verification looks like for performers in legitimate adult content — record-keeping requirements under 18 USC §2257 in the US, and the equivalent compliance regimes in other jurisdictions, were designed for a world where the question "is this person real" had a trivial answer. The third is whether watermarking and provenance should be treated as a regulatory mandate (the EU's AI Act direction) or as a voluntary industry standard (the C2PA approach).

There is no clean conclusion to write yet. The technical and policy responses are still in early innings, and the question of whether the synthetic-media ecosystem ends up looking more like the early-2000s spam problem (eventually contained by infrastructure-level filtering) or the early-2010s ad-fraud problem (a permanent low-grade equilibrium) is genuinely open. The honest summary is that AI has not solved adult content moderation and has not destroyed it either; it has reshaped the cost structure of certain harms and the verification burden of certain claims, and the institutional response is still adjusting.