Ten companies, including OpenAI, TikTok, Adobe, the BBC, and Bumble, have committed to voluntary guidance for developing and sharing generative AI. The recommendations focus on transparency: helping people understand what the technology can do and when they may be interacting with content it generated.
The guidance was assembled by the Partnership on AI (PAI), an AI research nonprofit, after consultation with over 50 organizations. Its early signatories include technology companies, media organizations, a civil society group, and synthetic-media startups.
Making AI-generated content easier to identify
A central recommendation is to research and introduce ways to tell users when content is synthetic. Possible approaches include watermarks, disclaimers, and traceable information connected to a model’s training data or metadata.
The guidance addresses both sides of the process: companies building AI systems and organizations creating or distributing synthetic media. In practice, that means being clearer about the technology’s capabilities and limits, as well as disclosing when an audience might encounter AI-generated material.
Claire Leibowicz, PAI’s head of AI and media integrity, said the aim is to prevent synthetic media from being used to harm or disempower people, while supporting creativity, knowledge sharing, and commentary. The recommendations also identify harms companies want to prevent, including fraud, harassment, and disinformation.
Voluntary commitments face questions about enforcement
The recommendations are not binding rules. Henry Ajder, an expert on generative AI who contributed to the guidance, described the current landscape as a “Wild West” and hopes the recommendations will help businesses identify issues as they adopt the technology.
Hany Farid, a University of California, Berkeley professor who researches synthetic media and deepfakes, welcomed the conversation but questioned whether voluntary principles can make enough difference. Companies outside the group may still offer tools that can produce inappropriate images or deepfakes. The source points to Stability.AI and its open source image-generating model Stable Diffusion as an example.
Farid argues that platforms and infrastructure providers should take a more active role. He says cloud service providers and app stores operated by Amazon, Microsoft, Google, and Apple should ban services intended to create nonconsensual sexual imagery using deepfake technology. He also argues that watermarks on AI-generated content should be required rather than left to voluntary commitments.
Experts say the guidance leaves development risks open
Some critics say transparency about finished content is only part of responsible AI. Ilke Demir, a senior research scientist at Intel who leads its work on responsible generative AI development, says the recommendations could say more about how models are trained, what data they use, and whether they carry biases.
Daniel Leufer, a senior policy analyst at digital rights group Access Now, highlights another omission: the guidance does not address toxic content in model training data. He calls this one of the significant ways these systems cause harm.
Demir also points out that the listed harms do not include every possible outcome. A model that consistently generates images of white people, for example, can cause harm through its narrow representation. That concern shows why a list focused on fraud, harassment, and disinformation may not capture the full range of risks.
The unresolved question is whether to build a system at all
Regulation is still catching up with generative AI. The European Union is seeking to include the technology in its upcoming AI law, the AI Act, which could require disclosure when people encounter deepfakes and impose transparency requirements on companies.
Until rules are established, the PAI guidance offers participating organizations a shared starting point for identifying and communicating risks. But the concerns raised by researchers point to unresolved choices: how to identify synthetic content, how to address data and bias, and how responsibility should extend to services that distribute or enable these tools.
Farid’s challenge goes beyond improving safeguards. If companies recognize that generative AI may cause serious harm and propose ways to reduce it, he asks, should they also consider whether the technology should be developed in the first place?