AI Notice Base Concept
Foundations / Best Practices / Global Patterns / AI Notice Base Concept
Intro
This guideline outlines UX and UI design methods that ensure users are clearly aware when interacting with AI-generated or AI-edited information. In accordance with the SAP Global AI Ethics Policy, transparency is crucial for building trust in AI-enhanced products, and the AI notice concept plays a key role in achieving this transparency.
Principles
Key principles of the AI notice base concept at SAP:
Discoverable
Ensure human oversight – users must be able to identify AI-created information confidently and accurately.
Informative
Manage expectations – users must have clarity about the characteristics and potential risks associated with the use of AI.
Transparent
Provide insight – users should be provided with options to educate themselves about the general terms of how AI operates and processes data.
When to Use
The AI notice base concept is foundational for differentiating AI-generated content and for any application where users interact with AI-created or AI-edited information.
Do
Use the AI notice base concept:
- To determine the right approach to mark AI results for your AI scenario.
- To determine the right life cycle of AI notice labels.
Don’t
Don’t use the AI notice base concept:
- In non-AI application scenarios.
Overview
The AI notice concept provides clear signals that content was generated by AI, using different methods and touchpoints throughout the user experience. However, it isn’t tied to a single UI component or interaction pattern. Because AI is used in diverse contexts, the recommended implementation may vary to reflect the specific impact AI has in each case. For more information, refer to the AI impact assessment section.
The following requirements need to be addressed:
- Consistent visual cues
For example, iconography, color, and other complimentary visual effects.
a. AI color
b. AI icon - Intelligible and explicit copy
For example, visible and invisible UI labels accessible via screen reader.
a. User assistance guidelines
b. Accessibility guidelines - Strategic marker placement with clear attribution
For example, element and/or area markers for AI-generated content.
a. AI notice base concept
b. Local AI notice pattern
Elements of user guidance for AI notice
Additionally, this concept is strongly linked with other experience concepts, for example explainability and user guidance. Therefore, we advise you to comply with the guidelines on message handling.
Best Practices
As a general rule, high-stakes decision scenarios should implement more rigorous safety measures to protect users. In scenarios with less severe consequences, the AI notice can be displayed less prominently. However, an AI notice label must always be present whenever AI has contributed to the results displayed on a screen.
- Label AI results: Provide explicit and unambiguous labels for AI-generated results.
a. Ensure users are informed about AI results by the AI notice label.
b. Use descriptive terms like “recommended”, “generated”, “predicted” in the UI copy for your AI results to clearly name the origin and characteristics of the given information. - Highlight AI results: Provide clear semantic color highlighting for AI results. For more information, refer to the using semantic and industry-specific colors guidelines.
a. Ensure users can distinguish AI results from other types of information through visual highlighting. - Persist markup information: Persist AI notice markups for every validating role (see Life Cycle definition).
a. Add metadata watermarks to AI-generated results to allow persistence of AI notice labels across users and applications.
b. Ensure that all changes made by AI are traceable and logged in the system change log.
Use descriptive UI copy
We recommend that you emphasize the characteristics of an AI-generated result and reflect this in the used UI copy. For more information, read the guidance on the contextualized AI notice label.
There are three main categories of AI content:
AI-curated content
Key characteristics
- Presents information pulled from other sources in an AI-defined order.
- Expressed through terms like “recommendation”, “proposal”, or “suggestion”.
- Helps to select the best available option.
AI-created content
Key characteristics
- Presents novel, generated information.
- Expressed through terms like “generated” or “synthetic”.
- Helps to create new information or content, based on available context.
AI predictions
Key characteristics
- Presents novel, inferred information.
- Expressed through terms like “predicted”, “simulated”, or “estimated”.
- Helps predict future situations and results.
Embedding AI notice labels into the main UI
It is mandatory to directly embed AI notice labels within AI-generated results in situations where users are required to make high-impact decisions based on those results as outlined in the AI impact assessment section of this guideline. Learn more about the correct placement in the local AI notice pattern guidelines.
Local embedded AI notice for high-impact decision scenarios and highlighted AI results
Replication of AI-generated content
AI notice markups are typically provided through interactive or non-interactive UI elements like labels or links. This is sufficient for non-replicable AI results. However, it is considered good practice to include the notice directly within the AI output when it is replicable. This applies to outputs like text, images, files, and other content that can be copied and pasted across different locations. There are two approaches:
- Hard embedding: For example, encrypting additional meta information directly into the file format of an AI output that can only be removed with considerable effort or under significant constraints.
- Soft embedding: For example, adding a text string to a copied AI text output that can be manually removed.
An emerging practice in many LLM services is to provide quick copy actions for AI-generated outputs. When pasted, the copied text includes an additional string, such as “Created with AI. Verify before use.” This added text acts as an extra layer of safety and can be manually removed by the user if desired.
Ensure through user testing that your end-users can clearly understand from the provided label whether the information they work with has been pulled from other sources (AI-curated) data or information newly created based on given input (AI-generated).
AI Impact Assessment
AI results are inherently limited by the quality of the input data and the accuracy of the trained model. Even with high-quality training data and careful optimization, there will always be a margin of error that can lead to mistakes during use. To address these risks, human oversight methods, such as human-in-the-loop, are essential to mitigate issues caused by incorrect AI predictions or hallucinations. Clearly marking and highlighting AI-generated results enables users to identify and verify the information, which is particularly critical in high-stakes scenarios where incorrect decisions can have serious consequences.
The individual impact associated with each decision scenario depends on several factors:
- Projectable consequences of incorrect decisions
- User competence to take an informed decision upon given results
- Complexity and clarity in the decision process
- Reversibility of applied incorrect decision
To determine the associated impact, we recommend that you get in touch with the AI Ethics Team and go through the official AI Ethics Self-Assessment process. This step should be applied for every stage in the decision-making process which requires the end-user to make decisions based on AI results.
Low impact
Low complexity, low stake
Definition
- Routine, straightforward tasks with little to no serious consequences.
- Simple tasks that are easy to automate.
- Predictable outcomes.
- Failure can be easily corrected.
Medium impact
High complexity, low stake
Definition
- Complex tasks with predictable consequences.
- Requires advanced skills.
- Failure might cause limited and manageable consequences.
- Failure can be easily corrected.
Guidance
- Users must be informed through a global AI notice label, which can be presented as a temporary message.
- Users should be able to toggle AI result highlights on or off at any time.
- Metadata watermarks to allow persistence of AI notice labels across users and/or applications are optional.
Guidance
- Users must be informed through a global AI notice label, which can be presented as a temporary message.
- AI content must be highlighted initially and then fade out. Users should be able to toggle AI result highlights on or off at any time.
- Metadata watermarks to allow persistence of AI notice labels across users and/or applications are optional.
High impact
Low complexity, high stake
Definition
- Simple tasks with potential for serious consequences.
- No special skills required but prone to improper handling.
- Failure could result in significant consequences, including reputational, financial, legal, or human well-being impacts.
- Failure is hard to reverse.
Very high impact
High complexity, high stake
Definition
- Complex tasks with potential for serious consequences.
- Specialized skills required and prone to improper handling.
- Failure could result in significant consequences, including reputational, financial, legal, or human well-being impacts.
- Failure cannot be reversed.
Guidance
- AI results must be marked using a locally embedded AI notice label.
- AI content must be highlighted by default. The user should have the option to toggle highlights for AI results on demand.
- AI results must be marked with a metadata watermark to ensure traceability of AI actions and enable persistence of AI notice labels across users and/or applications.
Guidance
- AI results must be marked using a locally embedded AI notice label.
- AI content must be highlighted permanently.
- AI results must be marked with a metadata watermark to ensure traceability of AI actions and enable persistence of AI notice labels across users and/or applications.
AI Notice Label
The design and properties of the AI notice label depend on the context and individual requirements of the use case. To give you some guidance, this section provides a non-extensive list of application scenarios for the AI notice label.
Global AI notice labels
Low–medium risk scenarios
A unified AI notice indicator that relates to one or multiple AI results in the current page. Individual AI results are made identifiable by specific local highlighting and descriptive labels.
Local AI notice labels
Low–medium risk scenarios
In addition to local AI content highlights and labels, the AI notice indicator is embedded next to the AI results to encourage user validation.
Preannouncement for global AI notice label
Local AI notice label embedded below the subsection title
Placement
For specific guidance about the correct placement of AI notice labels, refer to the respective guidelines on global AI base concept and local AI notice pattern.
Generalized AI notice label
This type is recommended for universal use in messages, below AI prompt input fields, and as a sub-label of AI results. Use this variant to apply a simple unified text markup.
Generalized AI notice text example: “AI can make mistakes. Check for accuracy.”
Contextualized AI notice label
When a more context-specific AI notice text is required, reference the examples provided here:
Examples
- Marking recommendations in a list: “Recommended with AI. Verify before use.”
- Marking generated content: “Generated with AI. Verify before use.”
- Marking a predicted value: “Predicted with AI. Verify before use.”
Visual Language
The AI notice should be displayed using subtle components such as labels or links. It must be clearly present, but not visually dominant, offering additional context to the AI experience while promoting user awareness and responsible engagement. All visual attributes and component states are inherited from the base components used to implement the AI notice label in various applications scenarios.
Read-only
Interactive
Highlighting AI results
The AI notice also relies on other visual cues like highlighting (using colors or complimentary graphical elements) and explicit user assistance (using clear and transparent labels) which are used together with the AI notice label to support strong user mental models. The aim is to reduce the need for repeating the AI notice indicator across the UI and instead use consistent highlights to visually group and differentiate AI results.
For more information on highlighting, see the guidelines for value states and form field validation. As of now, we are relying on existing capabilities in controls to display the information value state as the default to highlight AI results.
Life Cycle
Secure human oversight throughout all crucial validation steps and enable users to take informed decisions based on AI-generated information.
Responsible validator
AI provides results through automation or on user request. It is the responsibility of the respective validator interacting with the AI output to review and assess the accuracy of the results. The user has the option to accept, reject, regenerate, or disregard any AI output, using one of the two following modes:
- Explicit decision-making (using dedicated controls)
If the user does not explicitly choose what to do with the AI results, they must be prompted to make a decision (with a confirmation dialog) before exiting or completing the process. - Implicit decision-making
If the user does not change the AI results, they are considered implicitly accepted.
1 User – 1 Session
(n) Users and (n) Sessions
(n) Users, Sessions, and Applications
Prerequisites
- Potential risks and implications due to incorrect user decisions on AI outputs are classified as low.
- Evaluating the AI results remains entirely with the initiating user and within the same session.
Prerequisites
- Potential risks and implications due to incorrect user decisions on AI outputs are classified as low to medium.
- Multiple editors contribute to the end result.
Prerequisites
- Potential risks and implications due to incorrect user decisions on AI outputs are classified as medium to high.
- Content is processed across multiple users and applications.
Trigger events / enablement
- AI meta-attributes are transmitted to the front-end.
- AI-generated content is displayed on the screen.
Trigger events / enablement
- AI meta-attributes are stored/logged for given AI outputs.
- AI-generated content is displayed on the screen.
Trigger events / enablement
- AI meta-attributes are transferred when AI outputs are applied.
- AI-generated content is displayed on the screen.
Validation responsibility
- The initiating user.
Validation responsibility
- All editors / collaborators.
Validation responsibility
- All consumers or post-processors.
Persistence scope
- Persisted for the duration of the session.
- Limited to the one interaction scenario and application screen.
Persistence scope
- Persisted across multiple user sessions.
- Limited to the same interaction scenario and application screen.
Persistence scope
- Persisted across multiple user sessions.
- Persisted across multiple interaction scenarios, screens, and applications.
Potential impact and implication for the application scenario are assessed and classified in compliance with SAP’s Responsible AI Guidelines, SAP Global AI Ethics Policy and any effective regional legal regulations.
Related Links
Elements and Controls
- Message Handling (guidelines)
- Value States (guidelines)
- Design Principles for Generative AI (guidelines)
- AI Notice (guidelines)
Implementation
- Label (UI5 web component documentation)
- Link (UI5 web component documentation)
- Smart Link (SAPUI5 samples)