Requirements

Clearly communicate that generative AI is being used to create output

If you are using Generative AI to create output assets, you must be open and upfront. Do not try to mislead users.

You’ll have an opportunity when submitting your add-on to detail your add-on’s use of generative AI technology. You must mention in your listing that you are using Generative AI. You may also wish to indicate it in the add-on’s UI, although that is not a requirement.

Some possible examples:

• As part of your add-on’s description: “Create engaging comic book strips from text using the power of generative AI.” As part of your add-on’s UI: “Generate images” (call to action button)
• As part of your add-on’s UI: “Generate images” (call to action button)

Users should be aware that your add-on is using generative AI.

In the future we’ll call this out separately in the add-on’s listing page. But in the meantime, include at least the following text in your add-on’s description:

“This add-on uses generative AI technology to create content. You are solely responsible for the use of the content, and Adobe is not responsible for your use of this add-on.”

Content Requirements

Content created by your add-on must adhere to Adobe’s General Terms of Use and the Developer Terms of Use.

Your add-on must protect users against restricted content.

Your add-on must not:

• Contain or generate illegal content (such as CSAM, content that encourages illegal drug use, etc)
• Contain or generate adult content (such as sexual content, nudity, gore, intense violence, or strong language)
• Contain or generate hate speech or speech that promotes violence or bullying or cruel behavior to anyone
• Generate content that promotes, automates, or relates to highly regulated activities (such as financial advice, medical advice or diagnosis, legal advice or documents, contracts)
• Contain or generate code that could be malicious (such as viruses, malware, spyware, etc.)
• Participate in misinformation/disinformation campaigns
• Violate intellectual property rights, including copyright, of other persons and companies
• Automatically perform actions determined by AI that would be destructive or that can’t be undone without explicit user consent.

At minimum, you must apply filtering at the prompt level to protect against intentional attempts to generate restricted content. However, you might need to add additional filtering after generating content in order to pass Adobe’s automated tests.

NOTE: How a model is trained also has a large impact on what it might create. However, the model’s training is often insufficient to ensure that users are never exposed to restricted content.

As such, it’s often necessary to add additional filtering or post-processing steps. For example, you could have a set of words that would be immediately rejected, or a language classifier to detect violent speech, or an image classifier that detects sexual content in the model’s output.

Use cases requiring additional testing

In some cases, you will be required to confirm that you test your add-on to ensure that it does not generate restricted content. Adobe may also execute automated tests to validate the safety of your add-on.

Imagery (includes images and video)

• Text to image
• In-painting / out-painting
• Generative fill

Text (unless using ChatGPT and OpenAI’s moderation API)

• Text generated by a large language model
• Summarization
• Code generation

Audio

• Music containing generated lyrics
• Speech containing generated text (unless using ChatGPT and OpenAI’s moderation API)

Exceptions

Add-ons that meet the following criteria are not subject to automated testing by Adobe:

• Your add-on is privately listed (NOTE: We may still elect to remove your privately listed add-on if we receive credible reports of abuse)
• AI Models that the user installs and runs locally on their device
• Text-to-speech using Generative AI (provided the text is supplied by the user)
• Instrumental music or sound effects created using Generative AI (where there are no lyrics)
• Text generation if you use ChatGPT and OpenAI’s moderation API to filter content

Last updated 10/10/2023

Was this helpful?

Yes

No