Ensuring Trust in AI with Inclusive Design
Trauma-informed design (TID) principles were established in architecture and service design with an emphasis on providing safe spaces for those who have suffered trauma. More recently, TID has been adopted within digital design of web and mobile applications. The core principles of TID also lead to usable design that respects the user’s privacy, security, and trust.
TID Core Principles
- Safety
- Trustworthiness and Transparency
- Peer Support
- Collaboration and Mutuality
- Empowerment, voice, and choice
- Cultural Consideration
We are at an inflection point with the integration of Artificial Intelligence (AI) within our digital applications and services. Incorporating the TID principles into AI-influenced design will ensure our products are supporting our diverse audiences and prioritizing their experience.
Do No Harm
The most important element of TID is to do no harm. Our work will focus on ensuring the customer understands the risk of generative AI content, recognizes when AI is being used to generate information, understands how the decision was generated, and the ability to exit the experience when there’s a concern.
Use an accessibility-first approach to establish the experience of assistive technology users and those with cognitive disabilities. Building upon this foundation will ensure our AI interfaces work well for everyone.
What is the risk level?
Some AI is low-risk. For example, if you ask AI to make a cartoon of a cat skateboarding, it probably won’t hurt the customer. Using AI to transform a sketch into an estimate/invoice and send to a customer has multiple opportunities for error and a high impact on the business and customer. A customer should understand generated content and actions can include errors and must be reviewed. The process of generating an answer must also be transparent.
To do no harm, we must be transparent about the role of AI and the potential risk of accepting content without a review. AI identification needs to be clear and accessible. High-risk pathways must confirm approval. Describe how a decision was made via Explainable AI (XAI). Provide multiple options.
Best practices for AI Design
Clearly identifying AI
Many products use a combination of sparkles, propellers, and rainbow borders to highlight information that is AI generated.
Alt text SHOULD include “AI”
The icon represents the influence of AI. Include this in the alt text.
Bad
<img src=”…” alt=”Sparkles”>
Good
<img src=”…” alt=”AI powered”>
Focus on purpose, not design
At Intuit, we treat SVG icons as decorative via the aria-hidden rule. If they provide context, they are wrapped in a container, such as a button, link, or span that includes what the icon represents.
Bad (SVG)
<title>Thinking and generating</title>
<desc>Brand animation for thinking and
generating state</desc>
Good
<div role=”img” aria-label=”AI Powered”>
<svg aria-hidden=”true”>
</svg>
</div>
Avoid visual indications
Rainbow borders and animations are often used for defining something that is AI powered. This is fine if there is also content to describe the AI connection. This also includes the use of AI brand names, such as Rufus, Ari, and Gemini. In chat interfaces, it’s important to label the responder as an AI agent, Amazon’s Rufus uses an image with alt=””.
<img alt=”” src=”..” class=”rufus-avatar-img” height=”20" width=”20">
Explainable AI
In many ways, Explainable AI (XAI) is the antidote to irresponsible, harmful bias in AI. Imagine you are fired from your job after many years of solid work, only to find out the decision was made by an AI service. If the decision used XAI, the process would be clearly defined and open for review.
Cash flow is critical for small business success. QuickBooks introduced an invoice reminder service that highlights past due invoices and suggests emails to send to customers. This includes several options for the voice used, such as fun or firm.
There’s also a clear explanation of why this invoice was highlighted and the customer’s history. This is used to define the style of letter being sent.
Customer Choice
Choice and control are critical for designing with a trauma-informed approach. Our goal is to reduce anxiety and ensure the customer has control of their interactions.
Continuing with the QuickBooks example, there are several sections that ensure choice and control.
Voice and Tone chooser
While there are some guidelines for choosing the tone of a reminder email. A friendly message is great for an invoice that is only past-due a few days. But a stern message is needed for those that are overdue by more than 30 days. The customer is always able to choose what they feel comfortable with.
Exit the AI Experience
There’s a button to exit the generative AI functionality and send a basic past-due invoice reminder. The customer can still send an email, but trust the content is not AI generated.
Editable text
Each of the email attributes can be edited. This allows the customer to review the generated content to spell-check, correct contact information, and add personal notes.
Agentic AI
Most Agentic AI work will not have an interface; it’s the collection of data from multiple sources. As inclusive designers, we need to make sure the resulting data presentation can be verified, is understandable, and the customer has options to expand on the information.
Conclusion
Integrating trauma-informed design (TID) principles into AI design is crucial for building trust and ensuring inclusive products. By prioritizing safety, transparency, and user choice, we can create AI experiences that aren’t only functional but also respectful and considerate of diverse user needs and potential vulnerabilities. As AI continues to evolve, a commitment to these principles will be essential for mitigating risks, preventing harm, and fostering a positive and accessible experience for all users.
