On July 28, 2026 (Beijing time), AI Forensics released an audit project on Hugging Face Spaces. The Verge, which was authorized to publish a summary of this report in its subscription feed on the same day, noted that seven out of nine popular image editing models tested by researchers could easily modify images of women into revealing ones; according to the official index of AI Forensics' project, 73% of monitored requests contained sexual content. Both figures come from media summaries and research institution indexes, not self-reported data from Hugging Face.
Audit Scope and Research Conclusions
This audit targeted Spaces on the Hugging Face Hub rather than all repositories on the platform. According to official policy, Hugging Face categorizes models, datasets, and Spaces as machine learning artifacts and defines the Hub as an online platform that can host models, datasets, and machine learning applications. The policy also distinguishes between public, restricted, and disabled repositories: public repositories are accessible by any internet user, restricted repositories require conditions or approval from maintainers for their artifacts, while disabled repositories are only accessible to owners. Therefore, the current materials cannot generalize all artifacts on the platform as having the same level of openness.
AI Forensics named the investigation "lack of review in design" in its official project index and stated that the investigated Spaces formed a collection of tools capable of generating non-consensual intimate images, as well as comparing, benchmarking, and optimizing such tools. AI Forensics concluded that this concentration would facilitate the scaling up of image-based sexual abuse; this is the research organization's assessment. The 73% figure published in the index only refers to the requests it monitored, with no total request count, monitoring duration, user locations, or sampling methods provided on the currently accessible pages.
Seven out of Nine Results and Content Policy
The seven out of nine test results appear solely in The Verge's authorized summary from its subscription feed. The summary attributes this conclusion to AI Forensics, stating that seven models could easily complete the test tasks without listing the names of the nine models, their ranking criteria for popularity, or the number of repeated tests and error margins per model. The research details link returned a 404 during this collection period, making it impossible to verify more comprehensive protocols from the original report. Existing evidence supports describing these nine tested objects but cannot infer that all Spaces or image models on Hugging Face exhibit identical behavior.
Hugging Face's current content policy restricts sexual material used for harassment or bullying, sexual content created without clear consent, nudity involving minors, and any sexual content involving minors. The policy page indicates an effective date of April 10, 2025, with subsequent modifications taking effect ten days after publication. Written rules cover models, datasets, Spaces, and related community content; however, the policy text delineates what is allowed and prohibited but does not equate to actual enforcement records for the nine tested models.
Report, Disposition, and Enforcement Gaps
According to official policy, users can report content through in-platform features or a designated email. In-platform reports for repositories form public reports that notify the Hugging Face team. The platform may also flag based on social media leads, automated detection, or internal discoveries. After individual case reviews, possible actions include requiring owners to modify content, lowering rankings, adding not suitable for all audiences tags, removing or disabling access, restricting interactions, and suspending or terminating accounts. The policy does not commit to a specific action in every situation nor provide uniform processing timelines.
Hugging Face also stated that it prioritizes collaboration with repository owners when feasible, especially if modifications or added safeguards could bring them into compliance with the policy; owners can also set access thresholds for machine learning artifacts and manage discussions. These mechanisms indicate that official rules retain multiple levels of action from modification to disabling. However, existing official pages do not disclose whether the Spaces involved in this audit have been reported, flagged, or reviewed, nor do they publish numbers and times of demotion, tagging, removal, disabling, or banning; these enforcement outcomes cannot be confirmed.
Policy Conflict and Unverified Facts
Public materials present clear tension: AI Forensics claims that tested tools can generate non-consensual intimate images, while Hugging Face's written policy prohibits explicit sexual content without clear consent and any sexual content involving minors. The Verge’s subscription feed summary further describes the platform's preventive efforts as minimal, but this is a first-line media summary rather than quantitative enforcement metrics provided by Hugging Face. Both can indicate that policy implementation needs verification; they cannot be used to assert intentional approval of related requests or extrapolate results from nine models to the entire platform.
The most direct follow-up verification point is whether AI Forensics can restore the July 28 project details page and publish its complete methods, sample size, and model list, as well as whether Hugging Face will issue a response and disclose specific actions involving the tested Spaces. Current sources do not provide the identities or geographic distribution of the users involved, the content-removal status, enforcement totals, or legal conclusions; they also do not say whether these models still operate as they did during testing. Until that information appears, 73% and seven out of nine must retain their respective sample and source boundaries and cannot be presented as company-wide abuse rates.