Designing for AI-enhanced Experiences

Barriers are continually dissolving when it comes to the capabilities of machine learning programs. As an experience strategy and design professional, I find myself asking more frequently, “Are there relevant design principles I should consider when creating these new types of digital experiences?” Many of these platforms have an opacity that can affect the overall user experience in a variety of ways. In this blog post, I’ll suggest guidelines for constructing AI-enhanced experiences, and apply them to examples in the healthcare, consumer, and financial sectors.

Five design principles relevant to AI-enhanced experiences that engender trust, insight, and control

The principles I cover may not apply uniformly to all digital experiences, since other factors might predominate. For example, an internal application that facilitates portfolio construction may have a different “requirement” for a systemic recommendation than an athletic sportswear site that presents related sports gear to a triathlete. However, I believe that these design principles should become part of the dialogue as more digital experiences are mediated and impacted by forms of artificial intelligence.

1. Provide confidence metrics in your AI-enhanced experiences

There’s clearly a difference between the following two levels of confidence. Visibility of information helps users make an informed decision.

Confidence levels are key in AI-Enhanced experiences

Confidence metrics can make or break key decisions.

 

If you’re developing healthcare applications, for example, providing confidence metrics would be particularly relevant in high-stake diagnoses, like the facilitated interpretation of MRI scans for calcification signatures. A radiologist reviewing a mammogram would benefit from understanding the level of confidence in a set of diagnostic results, ranked in order of likelihood. If current systems don’t provide this metric, they are at the risk of over-identifying potential cancerous growths, and become like the boy who cried wolf.

One brand that effectively uses confidence metrics is Kayak, a travel search engine. Kayak provides a confidence projection in the form of a “buy now or wait” recommendation when guiding users who are deciding between two different cities to purchase an airline ticket.

Kayak confidence

Kayak uses an algorithm to estimate the confidence of when it is a good time to book a flight.

 

This type of metric is particularly useful in comparison or predictive situations, since there is no way to be 100% sure.


Experience Design Tip:

Consider prominently displaying confidence metrics, with recommendations based on an algorithm or heuristic. Consider enabling user control for confidence metric alerts.


2. Provide context for recommendation

Often, secondary factors are key to helping the user understand a particular choice over another. In online financial advisor applications, for example, more insight could help an investor choose one asset allocation model over another. This data could confirm that “Model A has a heavier weighting in international bonds.” This could provide valuable information over a chosen time period for investors. Without the deeper insight, the investor has no basis to make an informed decision.


Experience Design Tip:

Consider a natural language overview of the rationale. Try storyboard vignettes or another form of narrative structure to provide clear insight that isn’t bogged down by overwhelming details or complexity.


3. Allow visibility to underlying algorithms/heuristics

For a power user, understanding the core algorithm or underlying heuristic provides confidence and a deeper understanding. This can be as simple as understanding the relative weights of primary parameters, or as complex as rule extraction into decision trees or symbolic logic. This particular form of transparency is very dependent on the type of application and the user’s level of sophistication, but for certain expert systems, it will be essential to adoption and confidence.

In human mediated experiences, this will allow the expert to translate and answer client questions in a more thoughtful way. The radiologist who answers:

“The system believes that the fuzzy lump on the upper part of the right lung is pre-cancerous.”

is not as prepared to help the client determine whether a biopsy is required as one who can explain:

“When the system compared the density, level of margin irregularity, and the location of the fuzzy lump to its growing database of over eight hundred individuals with similar age/gender, it reported that the 68% who were biopsied returned as positive for pre-cancerous growth.”

The radiologist is also in a stronger position to communicate this back to the client in an appropriate way. The radiologist understands that the image processing is looking for correlations based on density of calcification signatures aka “dots,” margin irregularity, and location. This in turn helps them to extend the interpretation based on other factors that are not included in the image processing heuristic.


Experience Design Tip:

Where appropriate (based on user needs/skill, business model, etc.), document the algorithm or heuristic with key weighting and parameters defined and exposed.


Algorithm flowchart for AI-enhanced application

This is an example of a possible algorithm flowchart, with data to enable further decision making.

 

 4. Make sure your AI-enhanced experiences provide access to progressive details

Sometimes there is additional value in being able to drill down into the raw data that the machine learning is based upon. Again, this could mean designing far more specialized applications with internal or expert users. Employing a similar form of access and transparency can enable greater confidence and understanding.

The user experience would be similar to that where any data set that would benefit from the ability to perform faceted dynamic manipulation. Lung nodules, for example, can be identified with computer aided detection (CAD) algorithms that image process the cross-sectional “slices” of the computed tomography (CT) scans in search of uniform spheres.

AI-enhanced experience for lung scan

Make sure your AI-enhanced application provides a gateway to drill down into key details.

 

This particular technology requires many more “thin” scans than would normally be taken if the results were just being interpreted by a human who would use “thicker” scans. The systems that perform this actually do a pretty good job, but invariably the radiologist will want to review the raw data for themselves to ensure that it is not a false positive (most often in the form of normal vascular structures). There are approximately ten times the number of images generated in the CAD method, so facilitated access to the underlying data is essential to supporting efficient workflow.


Experience Design Tip:

Apply best practices for any ability to navigate large and/or complex data sets.


5. Provide simple mechanisms that enable user feedback on the quality of a recommendation

Waze, the traffic and navigation app, is a prime example of a brand that allows the user to easily provide information in the form of a feedback loop. This information is dynamically incorporated into the quality of the routing algorithms.

Feedback loops help with AI-enhanced experiences

Waze allows dynamic feedback channels to help users make driving decisions on the spot.

 

For instance, there might be an accident on a street. Because of new feedback data, those who normally travel on the street would now know to avoid it. When combined with the above four principles, a user will feel like they can understand the advice that an application gives them, and can actually improve its quality.


Experience Design Tip:

Provide a simple mechanism to contextually provide input, correction, supplemental feedback etc.


Summing It Up

Currently, it seems like the results of AI-enhanced experiences are opaque and mysterious. The desire to protect proprietary machine-based algorithms and heuristics is understandable.

However, where appropriate, we experience professionals are looking to evolve the design of AI-enhanced experiences to ones where users can truly understand what is being recommended. The sooner that users can get in and look “under the hood,” they will gain confidence and provide feedback.

Ultimately, this results in true hybrid solutions where humans and AI can collaborate and combine their unique expertise. Humans will supply narrative meaning and relationship context, while also providing common sense and the ability to spot emergent patterns. In turn, the AI will provide the power of correlation, big data, access to seemingly invisible patterns and the enormous horsepower to crunch through variations.

Image credit: “Aspergilloma CT scan” by Yale Rosen, licensed under CC 2.o

Barriers are continually dissolving when it comes to the capabilities of machine learning programs. As an experience strategy and design professional, I find myself asking more frequently, “Are there relevant design principles I should consider when creating these new types of digital experiences?” Many of these platforms have an opacity that can affect the overall user experience in a variety of ways. In this blog post, I’ll suggest guidelines for constructing AI-enhanced experiences, and apply them to examples in the healthcare, consumer, and financial sectors.

Five design principles relevant to AI-enhanced experiences that engender trust, insight, and control

The principles I cover may not apply uniformly to all digital experiences, since other factors might predominate. For example, an internal application that facilitates portfolio construction may have a different “requirement” for a systemic recommendation than an athletic sportswear site that presents related sports gear to a triathlete. However, I believe that these design principles should become part of the dialogue as more digital experiences are mediated and impacted by forms of artificial intelligence.

1. Provide confidence metrics in your AI-enhanced experiences

There’s clearly a difference between the following two levels of confidence. Visibility of information helps users make an informed decision.

Confidence levels are key in AI-Enhanced experiences

Confidence metrics can make or break key decisions.

 

If you’re developing healthcare applications, for example, providing confidence metrics would be particularly relevant in high-stake diagnoses, like the facilitated interpretation of MRI scans for calcification signatures. A radiologist reviewing a mammogram would benefit from understanding the level of confidence in a set of diagnostic results, ranked in order of likelihood. If current systems don’t provide this metric, they are at the risk of over-identifying potential cancerous growths, and become like the boy who cried wolf.

One brand that effectively uses confidence metrics is Kayak, a travel search engine. Kayak provides a confidence projection in the form of a “buy now or wait” recommendation when guiding users who are deciding between two different cities to purchase an airline ticket.

Kayak confidence

Kayak uses an algorithm to estimate the confidence of when it is a good time to book a flight.

 

This type of metric is particularly useful in comparison or predictive situations, since there is no way to be 100% sure.


Experience Design Tip:

Consider prominently displaying confidence metrics, with recommendations based on an algorithm or heuristic. Consider enabling user control for confidence metric alerts.


2. Provide context for recommendation

Often, secondary factors are key to helping the user understand a particular choice over another. In online financial advisor applications, for example, more insight could help an investor choose one asset allocation model over another. This data could confirm that “Model A has a heavier weighting in international bonds.” This could provide valuable information over a chosen time period for investors. Without the deeper insight, the investor has no basis to make an informed decision.


Experience Design Tip:

Consider a natural language overview of the rationale. Try storyboard vignettes or another form of narrative structure to provide clear insight that isn’t bogged down by overwhelming details or complexity.


3. Allow visibility to underlying algorithms/heuristics

For a power user, understanding the core algorithm or underlying heuristic provides confidence and a deeper understanding. This can be as simple as understanding the relative weights of primary parameters, or as complex as rule extraction into decision trees or symbolic logic. This particular form of transparency is very dependent on the type of application and the user’s level of sophistication, but for certain expert systems, it will be essential to adoption and confidence.

In human mediated experiences, this will allow the expert to translate and answer client questions in a more thoughtful way. The radiologist who answers:

“The system believes that the fuzzy lump on the upper part of the right lung is pre-cancerous.”

is not as prepared to help the client determine whether a biopsy is required as one who can explain:

“When the system compared the density, level of margin irregularity, and the location of the fuzzy lump to its growing database of over eight hundred individuals with similar age/gender, it reported that the 68% who were biopsied returned as positive for pre-cancerous growth.”

The radiologist is also in a stronger position to communicate this back to the client in an appropriate way. The radiologist understands that the image processing is looking for correlations based on density of calcification signatures aka “dots,” margin irregularity, and location. This in turn helps them to extend the interpretation based on other factors that are not included in the image processing heuristic.


Experience Design Tip:

Where appropriate (based on user needs/skill, business model, etc.), document the algorithm or heuristic with key weighting and parameters defined and exposed.


Algorithm flowchart for AI-enhanced application

This is an example of a possible algorithm flowchart, with data to enable further decision making.

 

 4. Make sure your AI-enhanced experiences provide access to progressive details

Sometimes there is additional value in being able to drill down into the raw data that the machine learning is based upon. Again, this could mean designing far more specialized applications with internal or expert users. Employing a similar form of access and transparency can enable greater confidence and understanding.

The user experience would be similar to that where any data set that would benefit from the ability to perform faceted dynamic manipulation. Lung nodules, for example, can be identified with computer aided detection (CAD) algorithms that image process the cross-sectional “slices” of the computed tomography (CT) scans in search of uniform spheres.

AI-enhanced experience for lung scan

Make sure your AI-enhanced application provides a gateway to drill down into key details.

 

This particular technology requires many more “thin” scans than would normally be taken if the results were just being interpreted by a human who would use “thicker” scans. The systems that perform this actually do a pretty good job, but invariably the radiologist will want to review the raw data for themselves to ensure that it is not a false positive (most often in the form of normal vascular structures). There are approximately ten times the number of images generated in the CAD method, so facilitated access to the underlying data is essential to supporting efficient workflow.


Experience Design Tip:

Apply best practices for any ability to navigate large and/or complex data sets.


5. Provide simple mechanisms that enable user feedback on the quality of a recommendation

Waze, the traffic and navigation app, is a prime example of a brand that allows the user to easily provide information in the form of a feedback loop. This information is dynamically incorporated into the quality of the routing algorithms.

Feedback loops help with AI-enhanced experiences

Waze allows dynamic feedback channels to help users make driving decisions on the spot.

 

For instance, there might be an accident on a street. Because of new feedback data, those who normally travel on the street would now know to avoid it. When combined with the above four principles, a user will feel like they can understand the advice that an application gives them, and can actually improve its quality.


Experience Design Tip:

Provide a simple mechanism to contextually provide input, correction, supplemental feedback etc.


Summing It Up

Currently, it seems like the results of AI-enhanced experiences are opaque and mysterious. The desire to protect proprietary machine-based algorithms and heuristics is understandable.

However, where appropriate, we experience professionals are looking to evolve the design of AI-enhanced experiences to ones where users can truly understand what is being recommended. The sooner that users can get in and look “under the hood,” they will gain confidence and provide feedback.

Ultimately, this results in true hybrid solutions where humans and AI can collaborate and combine their unique expertise. Humans will supply narrative meaning and relationship context, while also providing common sense and the ability to spot emergent patterns. In turn, the AI will provide the power of correlation, big data, access to seemingly invisible patterns and the enormous horsepower to crunch through variations.

Image credit: “Aspergilloma CT scan” by Yale Rosen, licensed under CC 2.o

Interested in partnering with us?

Send a message and we will work with you to understand your needs.

Related Insights from Our Experts
Related Consulting Solutions

Design Thinking & Ideation

Visualize your ideas and refine your thinking for the best path forward. We bring design thinking into agile programs and help stakeholder teams embrace a user centric model for digital product design.

Customer & User Research

Ground your CX initiatives on real insights uncovered via contextual inquiry.

Digital Application Architecture

Let us accelerate your technology shift by partnering with you to establish the best-fit front-end, service and back-end technology, process and skills that enable your IT organization to meet the needs of an Agile product organization.

Product & Platform Development

Leverage our proven Agile UX methodology to realize your customer experience. We accelerate the delivery of new digital products and capabilities or modernizing your digital platform by bringing expert high-performance teams – not an army but a swat team.

2017-06-02T09:25:35+00:00