AI Nutrition Fact Labels
The AI Nutrition Facts Label is intended to give consumers and businesses a more transparent and clear view into ‘what's in the box’ - how their data is being used - especially when it comes to training LLMs with vendors like AWS, Google, and OpenAI. These labels empower customers to make informed decisions about which AI Powered capabilities they are prepared to adopt.
This page is here to help you understand more about the AI Nutrition Fact labels and let you create labels for the features you are building.
How to read a label
Below you can find explanations for all fields and their meanings to better help you navigate these labels.
| Description | A description of the feature that the label is for |
| Privacy Ladder Level | A matching level based on the AI Privacy Ladder describing how customer data is used in the AI model. |
| Feature is Optional | Describes whether a customer can opt-in/opt-out of this specific feature |
| Model Type | Describes what kind of AI model is being used. For example a Generative AI model vs a Predictive AI model |
| Base Model | What version of model is the functionality is built on. For example "OpenAI - GPT-4" or "Anthropic - Claude 2" |
| Base Model Trained with Customer Data | Was any customer data used to train the base model? |
| Customer Data is Shared with Model Vendor | Is any customer data sent to the vendor that trains the base model? |
| Training Data Anonymized | Is the data being anonymized before it is being used to train the model? |
| Data Deletion | Deletion of specific pieces of data that have been stored in the model. This could include removing training examples, text prompts, user interactions, or any other form of input that has been used to fine-tune or train the model. |
| Human in the Loop | Are there tools to review/change/block the output of the AI before any final action is being taken? For example generating an email and having a human approve it vs automatically sending the generated email. |
| Data Retention | How long data related to the feature will be retained. This data can include meta data about input or weights, interaction logs, etc. that might be retained for legal or auditing purposes. |
| Logging & Auditing | Does the feature provide logging and auditing tools that help a user understand what output the AI has produced and how the output was produced? |
| Guardrails | Are there checks and guardrails in place that check for harmful output including ethical, bias, violence, hate, etc.? |
| Input/Output Consistency | Does the feature produce consistently the same output given the same input? |
| Other Resources | Any other relevant resources such as links to privacy policies. |
About the AI Privacy Ladder Level
The following levels are based on Twilio's CustomerAI Privacy Ladder and are intended to clearly identify how customer data is used within an AI feature, and who the model can be used by. Even though we developed this ladder for Twilio products we believe it can widely apply to AI use cases. We think of these levels in terms of a ladder. The higher the level the more caution you'll have to practice.
Models with your data for your use, without PII
Level 1 is the lowest privacy risk option and identifies models that are trained only with your data for your exclusive use without any personally identifiable information (PII)—that is, information that, when used alone or with other relevant data, can identify an individual.
Models with your data for your use, with PII
Level 2 offers more value and requires more caution than the first level as it identifies models trained with your data for your exclusive use, including PII.
Models with your data for multi-customer use, without PII
Level 3 offers the highest level of value and need for caution as it identifies models trained with your and other customers' data for multiple customer use but does not include any PII. This may include personal data that has been anonymized prior to being used to train the model.
Models with your and other customers' data with PII
Level 4 is where custome's data is used to train the model with other customer' data that includes personal data/PII.
Create your own label
Creating your own label is currently not supported on mobile devices or devices with small screens. Please open the editor nutrition-facts.ai on a device with a larger screen.
You can get started creating your own labels using the editor to your leftabove. The individual field names are fixed to give customers the ability to easily compare different features. To change any value of a field click into the respective field you'd like to change. Some fields also allow you to add additional comments to give more details.
Once you created your label you can either download it as a PNG or export a JSON config that you can share with your colleagues to enable them to do more changes.
If you want to start from scratch at any point click the reload button at the top right of the editor.
Partner with us
Interested in working with us to help establish AI Nutrition Facts as an open standard for transparent, responsible, and accountable AI practices? Reach out to us at ai-nutrition-facts@twilio.com