Fairness & Inclusiveness

Apologies for Inaccuracy

leighbryant — Thu, 15 Oct 2020 00:00:00 +0000

Apologies for Inaccuracy

Application

Communication

leighbryant Thu, 10/15/2020 - 00:00

Topic

Fairness & Inclusiveness

Problem:

The AI in a system may generate inaccurate results that cause confusion, may be offensive, or otherwise unsettle the user. The user would like to know that this is a possibility and feel comfortable with the system regardless.

Microsoft's "How Old" AI-powered software guesses the age of the individual featured, but also includes an apology disclaimer for inaccurate guesses beneath the image.

Solution:

The system apologizes for any possible inaccuracy that may be present, especially in regards to sensitive topics that could cause harm or offence.

Discussion:

While it is important for systems to admit to their fallibility (via Setting Expectations & Acknowledging Limitations, for example), there is an additional requirement to address how this fallibility may affect the user. In reality, what we call "the user" is a million unique individuals each of whom want to be treated with dignity and respect, not just in this interaction but in every facet of everyday life. We need to use different lenses, with more expansive understandings of who "the user" is in terms of social, psychological, physical, and ideological variables that describe any one embodied individual that goes beyond standard task completion assessments. Apologizing for potentially harmful interpretations acknowledges the humanity of the individuals.

Automated Alt Text

mattiealston — Sat, 15 Aug 2020 00:00:00 +0000

Automated Alt Text

Application

Display

mattiealston Sat, 08/15/2020 - 00:00

Topic

Fairness & Inclusiveness

Problem

Users with vision impairments want to understand what image content exists in the interface. The system designers want to facilitate this without manually tagging alt text for every image.

Facebook generates alt text as content tags, but also allows the user to edit these if needed.

Solution:

Image recognition automatically tags images with alt text based on what the AI identifies in that image.

Discussion:

While the majority of patterns catalogued in this library are focused on the benefits to the user, in some cases, the value of AI may not be recognized by the end user but rather by those responsible for designing and maintaining the system. In this case, the primary users of alt text will care little how it is generated, as long as exists. Designers, on the other hand, should use patterns like this to broaden our ideas as to where data can be sourced from and ask which is the best approach in each case: from the initial uploader or content manager, crowdsourced across users, or generated by the system itself?

Each has advantages as per the accuracy, authority, breadth and depth of the data created, but the AI approach is unique in terms of requiring the least human effort. That said, it doesn’t pay to be one hundred percent hands-off— automated content tags of any kind benefit from an additional layer of human intervention, where end users can ad-hoc moderate the data by flagging anything they believe is erroneous or offensive.

Automated Subtitles

mattiealston — Wed, 15 Jul 2020 00:00:00 +0000

Automated Subtitles

Application

Communication

mattiealston Wed, 07/15/2020 - 00:00

Topic

Fairness & Inclusiveness

Problem:

Users want to know what is being said in video content. The system designers want to facilitate this without manually transcribing all videos.

While a manually created transcription might be more accurate, YouTube's automated subtitling is very effective at making videos accessible for users.

Solution:

Speech recognition automatically generates subtitles for video content. The user can choose to activate the subtitles as required.

Discussion:

Similar to Automated Alt Text, the benefits of this pattern are more in the efficiency gains to the designers than an improvement in the experience of the end users, who care more about the existence of subtitles than their source.

That said, we shouldn’t overlook the empowering potential of patterns like this — it isn’t that the users should be wowed by an AI automagically producing content, but rather that often the deployment of automation functionality like this can mean the difference between the content existing or not. In the case of automatically generated subtitles for video, this can make previously inaccessible content available to users with hearing impairments, which is of immense value to them.

Automated Text Summarization

mattiealston — Mon, 15 Jun 2020 00:00:00 +0000

Automated Text Summarization

Application

Communication

mattiealston Mon, 06/15/2020 - 00:00

Topic

Fairness & Inclusiveness

Problem:

In situations where there is a lot of body text, users with cognitive or literacy issues want to know what is contained within that content without having to read the whole text.

The Automatic Text Summarizer tool lives up to its name, demonstrating how well this functionality can work.

Solution:

The user can click to view a shortened and simplified version of the text that has been generated by an AI.

Discussion:

While patterns such as Automated Subtitles are certainly empowering, they are really just instances of automating the production of conventional content, and do not go far in demonstrating the true revolutionary potential of AI in this domain. Automated summarization, on the other hand, shows that AIs can actually produce valuable original content, and to varying specifications to match users’ unique needs.

Yes, it would be possible for a copywriter to simply create two or three different versions of a text to suit different reading levels, but this is labour intensive and rarely done. With an AI automating this process, the opportunity is there to allow for multiple versions of the same text with little hassle, opening the door for true personalization that fits each user’s own abilities and needs.

Explaining Reductive Inputs

leighbryant — Fri, 15 Mar 2019 00:00:00 +0000

Explaining Reductive Inputs

Application

Input Data

leighbryant Fri, 03/15/2019 - 00:00

Topic

Fairness & Inclusiveness

Problem:

The user wants the system to respect their identity and not force them to conform to an identity that doesn't match their own simply to use the app.

By providing an explanation of why it requires a binary response, this health app acknowledges the identity of the individuals using it even when it has to limit them within the application.

Solution:

If the system does ask the user to provide a binary choice around identity—e.g. if a health app asks a user to choose between male and female—then the system explains why it necessary to reduce the options to only these choices, and qualifies what the implication of each is.

Discussion:

While ideally any app allows for the user to choose any gender, medical diagnosis normally requires a choice between binary male or female. The UI is not really the best place to discuss issues around the relationship between gender and sex and the social construction of both. Explaining the necessity of asking a reductive question in such a way as to make it clear that the system empathizes with the users concerns is tricky, yet it must try to do its best.

Gender Neutral Bot

leighbryant — Sat, 15 Dec 2018 00:00:00 +0000

Gender Neutral Bot

Application

Behaviour

leighbryant Sat, 12/15/2018 - 00:00

Topic

Fairness & Inclusiveness

Problem:

The user wants to interact with a bot that isn't characterized as being distinctly male or female.

An app with a user-generated bot profile allows the user to choose non-binary options for the persona being created.

Solution:

The bot character is portrayed as gender neutral, which can be conveyed via the pronouns it uses to refer to itself, the tone of voice it uses, the visual imagery or iconography, etc.

Discussion:

Is it a problem for the user if a bot is understood as male or female? Not directly in terms of task completion and a lot of the time it may have little impact on a user's impression of an app. But the perception of gender varies, so if we want to be in full control of the impression our bot is making and mitigate risks, we should avoid overt gender coding altogether. There's also a moral imperative not to contribute to harmful gender stereotypes (for example, by aligning a subservient bot with a female gender or a powerful one with male).

Focusing on gender is just one way to think about these issues; there are many other ways in which human-like traits in a bot can cause issue. Users who align with the traits may feel like their identities are being parodied or otherwise reduced to a stereotype, and users who don't align with the traits may feel they're not the intended audience and/or that the system is designed so they're dissuaded from greater participation. Steering clear of human-like traits altogether can reduce risk and is an effective way to avoid the Uncanny Valley effect if nothing else.

Gender Neutral Voice

mattiealston — Thu, 15 Nov 2018 00:00:00 +0000

Gender Neutral Voice

Application

Communication

mattiealston Thu, 11/15/2018 - 00:00

Topic

Fairness & Inclusiveness

Problem:

The user wants to interact with a voice interface which uses a voice that isn’t distinctly male or female.

The pioneering Q voice is designed to sound neither female nor male.

Solution:

The voice used by the interface is carefully constructed to sound neither male or female.

Discussion:

As with a Gender Neutral Bot, the perceived gender of a voice UI does not affect task completion and so may be overlooked. However, unlike bots which use interface, text, graphics, and emojis to communicate, the voice literally is the interface for a voice UI, so our entire perception of that interface is determined by the tone, expression, and vocabulary of it. The trend has been to default to female voices in these interfaces, which is especially problematic given that they often play the role of domestic servants or are otherwise subservient.

There are least two benefits to using a gender neutral or genderless voice like Q. One is immediately clear, that of avoiding the perpetuation of harmful stereotypes. The second one may be less obvious, but is also worth considering — that a significant number of people identify as non-binary or otherwise reject binary gender labels. Providing a non-binary voice option is a valuable recognition that people with non-binary genders exist.

Inclusive Voice Recognition

mattiealston — Sat, 15 Sep 2018 00:00:00 +0000

Inclusive Voice Recognition

Application

Input Data

mattiealston Sat, 09/15/2018 - 00:00

Topic

Fairness & Inclusiveness

Problem:

Users with speech impairments would like to use voice interfaces, but the software often fails to recognize their speech.

Google's Project Euphonia provides voice recognition to users with speech impairments.

Solution:

The AI that powers an inclusive voice interface is trained with voices from users with speech impairments so that it can understand their inputs. This is achieved to a baseline level through collective training, in addition to training for each individual so as to optimize the system for their personal speech patterns.

Discussion:

Accessibility issues are well understood when it comes to traditional web applications, and the need for interfaces that are accessible to users with vision loss or physical impairments, for example, is widely acknowledged (even if fully accessible UIs are not provided as widely as they should be). By comparison, it can be easy to overlook the issues that users may have with new emergent interfaces, especially those that rely on AI or ML. Unlike the lack of alt text for screen readers, for instance, there are few indications to the able-bodied tester that the interface will fail for some users. Voice interfaces are a very useful case study in that regard — it is only through inclusive interfaces like Voiceitt or Google’s Project Euphonia, that we are made aware of how inadequate our standard provision may be.

Going further, we should not just consider how these projects can make ordinary software accessible to all users, we should embrace opportunities to leverage the unique properties of these emergent interfaces in order significantly improve the quality of life for people with speech impairments, such as those with cerebral palsy, amyotrophic lateral sclerosis (ALS), Parkinson’s Disease, brain cancer, or traumatic brain injury.

Sign Language Recognition

mattiealston — Wed, 15 Mar 2017 00:00:00 +0000

Sign Language Recognition

Application

Input Data

mattiealston Wed, 03/15/2017 - 00:00

Topic

Fairness & Inclusiveness

Problem:

Users who are deaf or hard of hearing want to use sign language to interact with a system in a similar way to how hearing users use a voice interface.

Image recognition such as used by Google can already capture hand gestures, which is half of the challenge of developing sign language recognition.

Solution:

Using video capture and pattern recognition, the system allows the user to communicate with it via sign language. This can be used to issue the system commands, as per a typical voice interface. Alternatively, this can translate sign language into other formats, allowing the user to communicate with other users via text or speech, for example.

Discussion:

When we think of machine translation, we typically picture one language translated to another via text or voice, but there are of course many other ways to communicate. While many sign language users can use other input methods, being able to communicate with a system through their preferred language allows them the same efficiencies and quality of experience as non-deaf users.

Traditionally there have been complexities around machine translation of sign languages such as understanding how the syntactical structures are different from spoken languages. But as advancements in AI unlock the capacity to use signing as input, we should challenge ourselves to consider a range of viable input methods, rather than privileging some and sidelining others.

Anti-Pattern: Algorithmic Gender Recognition

mattiealston — Tue, 15 Sep 2015 00:00:00 +0000

Anti-Pattern: Algorithmic Gender Recognition

Application

Behaviour

mattiealston Tue, 09/15/2015 - 00:00

Topic

Fairness & Inclusiveness

Problem:

The user does not want to be misgendered by an AI. In fact, the user does not want a piece of software to make any assumptions about their gender at all.

This demo by Microsoft is designed to predict gender, age, emotions, and whether the two photos are of the same person.

Anti-pattern response:

An image recognition AI attributes data points to an image of the user, one of which is the user’s gender. Variants of this include AI-powered systems attempting the same prediction via speech recognition, textual analysis, behaviour tracking, and so on.

Discussion:

By using data pulled from a collection of individuals, we train AI systems to make predictions about any one individual based on what it learns about the group. In some cases, these generalizations and assumptions will be accurate and useful. In other cases, there are faults in the dataset that can cause problems, either due to biases in the sample data or faulty extrapolation from that data. When the data points concern superfluous details such as whether the user wears glasses or not, that user will not be offended by incorrect data attached to them. When the data concerns matters of gender identity, sexual orientation, racial or cultural identity, or other matters that the user is deeply affected by, much greater care should be taken about generating accurate assumptions.

In the case of gender, there are two substantial problems to overcome initially. First, a male/female binary is incorrect and simply adding more options (non-binary, for example) will likely also be an inadequate simplification. Second, assuming we could create a satisfactory catalogue of gender positions, the challenge would be to ensure the dataset captured contains a sampling of enough diversity to be predictively useful. Ultimately though, this functionality is doomed to fail simply because many individuals’ gender presentation doesn’t match their actual gender. If a system deploys this pattern, it will inevitably be inaccurate and cause offence for a number of users.

This pattern is indicative of a core challenge in developing AI functionality. Given that almost everything will prove possible to build in due course, the question is not whether we can build it, but whether we should. In this case, whatever benefits there may be must be weighed against the fact that the harm caused is impossible to mitigate.

A second core challenge is that when we are concerned with capturing and generating data without a deeper understanding of the nature of that data, we are always in danger of not just replicating existing harmful biases, but amplifying them and solidifying them.