An Ethical Analysis of AI and Flo Period Tracker

Microsoft created the “Guidelines for Human-AI Interactions” to help facilitate ethical, human-centered interaction between humans and AI. To provide an example of analyzing an interface with these guidelines, I will be looking at the mobile application “Flo.” Flo uses AI to track menstruation and menopause, create monthly calendars, and monitor pregnancy. I will be analyzing the app’s menstrual tracking function based on guidelines from each of the four stages: Initially, During Interaction, When Wrong, and Over Time. For each of the four interaction stages, I will use one related guideline to assess Flo on a 5 point scale (1=Clearly violated to 5 = Clearly applied).

Initially

• G1 — Make clear what the system can do
• Score: 3

Flo’s main function is the period tracker, which shows a user’s menstrual cycle on a calendar. However, when a new user opens the app the front page just shows an estimate of how many days later the user’s next period will begin. It is unclear that the user should click on the calendar icon to input their dates and receive a more accurate prediction.

During Interaction

• G4 — Show contextually relevant information
• Score: 5

The system shows information, articles, and insights entirely based on how the user describes their mood, physical symptoms, and behaviors that day. For example, if somebody inputs a series of negative symptoms right before their period, the system will show suggestions for managing PMS (premenstrual syndrome).

When Wrong

• G9 — Support efficient correction.
• Score: 4
• Every day, the user can “log period” for that day, “change period dates” from their past, or click the “plus” button to input and analyze other symptoms, all of which correct the system’s future predictions and encourage users to update information.

Over Time

• G13 Learn from user behavior.
• Score: 3

The app, because of the health-related personal nature of it, is all behavior based. They ask for information on exercise, food, water, sexual activity, and more. However, the system should learn more from a user’s behavior on the app itself because it relies on the assumption that users will consistently log in and update their information. If the system does not get updated information from the user during each menstrual cycle, it cannot learn from and provide accurate predictions.

More Examples

I will now provide examples of interfaces which clearly violated (1) or clearly applied (5) one of the Microsoft guidelines from each of the four stages of interaction.

Initially:

• G2 — Make clear how well the system can do what it can do
• Example of 5 (clearly applied) — Depop
• Depop, a clothing-selling app, creates a list of items “you might also like” based on a user’s search and response history. This language makes it clear that the system isn’t able to guarantee anything, but that they can offer a prediction or show some similar items.

During Interaction:

• G3 — Time services based on context.
• Example of 1 (clearly violated) — Headspace
• Headspace, a meditation app, sends alerts at the same time every day. Often users meditate, get up and get ready, and then get a notification from Headspace to “not forget to meditate!” This function would be more effective if they took into account whether the user has or hasn’t yet used the app for its function that day.

When Wrong:

• G11 — Make clear why the system did what it did.
• Example of 5 (clearly applied) — Amazon
• Amazon states why they are suggesting certain items by saying “related to items you’ve viewed,” “inspired by your shopping trends,” or “recommended items other customers often buy again.” These make the shopping experience very clear because they state what information they have gathered and what they have concluded from that information, rather than just making advertisements appear to a user with no explanation.

Over Time

• G15 — Encourage granular feedback
• Example 5 (clearly applied) — Spotify
• Spotify constantly checks in on a user’s feedback by providing the option to like or dislike suggested music, neutrally skip a song, or add songs to their library. They clearly state that this feedback helps them create more accurate suggestions for a user, creating an incentive for the user to engage with these controls.