The Pulse
CDT Finds Major Gaps in Seven AI Firms’ Image Safeguards
A Center for Democracy & Technology report finds that only one of seven AI companies it examined explicitly defines non-consensual intimate imagery in its policies. The report also says firms have not shown that their technical safeguards r
AI.info Team · Written with AI ·
Only one of seven companies defines the harm
Only one of the seven AI companies reviewed by the Center for Democracy & Technology explicitly defines non-consensual intimate imagery in its policies, the organization says. The report, published October 15, finds that the companies’ rules and technical safeguards leave gaps in preventing AI systems from creating intimate images that depict identifiable people without consent.
The researchers examined Anthropic, Google, Meta, Midjourney, OpenAI, Stability AI and xAI. They reviewed company policies and technical documentation, academic research and 11 interviews with representatives of AI companies and civil-society organizations. The analysis covers practices from January through June 2026 and focuses on adult victims and adult content, not child sexual abuse material.
The authors distinguish non-consensual intimate imagery, or NCII, from sexual content generally. Their concern is that systems often target broad categories such as nudity or erotic material instead of the specific abuse at issue: using a person’s likeness in an intimate context without consent.
Broad sexual-content rules can miss likeness abuse
All seven companies prohibit NCII in theory, according to the report, but their policies classify it differently—as sexual content, sexual violence or misuse of someone’s likeness. Meta was the only company in the review to provide an explicit definition of NCII. The researchers say that vague or scattered rules make it difficult for users to understand what is prohibited and for outsiders to judge whether safeguards are being applied consistently.
That broad framing creates risks in both directions. A system focused on sexual content may fail to catch an image that uses someone’s likeness without being classified as sexual, while an imprecise filter may block lawful material involving sex education, health care, sex work or gender-related topics. The report says AI-based systems can apply those definitions inconsistently, adding uncertainty to decisions about both abuse and legitimate expression.
Protections can disappear outside company products
Many safeguards operate at the product level: a service may filter a prompt, review an output or restrict a user’s access. But developers who build products on top of a model may not use the same protections. The gap is particularly pronounced with open-weight models, whose downloadable components can be modified or used without the original company’s oversight.
The report says developers can fine-tune open-weight models for specialized uses and remove some restrictions. It also notes that current interventions can have unintended effects: filters aimed at sexual material may interfere with lawful content, while model-level changes may need redesign as systems change. Among training-data measures, the report found that companies said they filter for detected child sexual abuse material, but none publicly reported filtering training data for NCII; Anthropic alone listed such filtering on its roadmap.
Companies have not shown that safeguards work
CDT found that companies describe safeguards with varying levels of detail but generally do not publish enough information to establish how well they work against NCII. The researchers say testing methods are often opaque, making it hard to tell whether company evaluations reflect real-world attempts by people with different levels of motivation, skill and access to tools. The report calls for regular testing that accounts for tactics such as prompt evasion, image uploads and the use of downstream applications.
Academic research cited in the report has found substantial limits to technical interventions. The authors argue that independent testing is needed alongside companies’ own evaluations, while acknowledging that no test can cover every possible way of generating an image. They recommend that claims about a safeguard specify which tactics were tested and explain how those tests relate to real-world abuse and lawful expression.
Survivors’ experience is part of the test
The report argues that technical performance alone cannot measure whether a safeguard is working for people targeted by abuse. Companies might count flagged content, policymakers might look for enforceable rules, and survivors might judge the response by whether they can report harm and get help. Those measures do not necessarily describe the same outcome.
“Everything I think doesn’t work is because I usually work directly with survivors and it hasn’t f****** worked.”
Omny Miranda Martone of the Sexual Violence Prevention Association, quoted in the CDT report
CDT calls for a shared testing standard, clearer definitions of NCII and more access for independent researchers to assess protections. The report leaves a specific question for companies and policymakers: what test results should count as evidence that a model or product is sufficiently safe?