The team and timeline
Context
Microsoft Cognitive Services are AI services, which can be used in many ways. Whilst AI services can be extremely valuable to many users, such as allowing people with disabilities to be more independent (find out more about Seeing AI here), they can also be used in unethical ways, which could be harmful. Microsoft is passionate about responsible use of such services. This is why information on responsible AI use was going to be added to developer documentation for applicable AI services.
The overall goal of this research was to explore the understanding of the category labels that would be part of the ‘responsible AI’ node in the table of contents and assess what type of information would users expect to find under this category.
Source: https://www.microsoft.com/en-us/ai/responsible-ai-resources
Research plan and questions
Method: Survey
Research questions
I chose a qualitative survey, because I knew I wanted to collect a good number of responses across different countries. The survey was set up to collect open responses, which were analyzed by coding and applying thematic analysis.
For each label I asked:
1. What is the label about?
2. How confident are you that you know what the label is about?
3. What type of information would be expected under this label?
Preferences for wording for some of the labels were also explored.
Do developers know what responsible AI is about?
What information would they expect under this node?
Do other labels under this node make sense in terms of content that is planned to be published under them?
Outcomes and impact
As a result of this research, we:
Identified and reworded labels that have potential to confuse
Discovered what information users would expect under each label
Uncovered strong associations between content and expected terms, which informed label naming
Identified opportunities for education across different geographical regions
Responsible AI node is now a part of several Azure Cognitive Services, guiding developers on ethics in AI
https://azure.microsoft.com/en-us/services/cognitive-services/
Key insights
Responsible AI practices - what are they?
More than half of our participants had not heard about responsible AI practices, despite many working for companies that use AI services. This means that there is an opportunity to raise awareness.
Some geographical areas were better at knowing about responsible AI
I coded all the qualitative comments into three categories to see the level of knowledge participants had about responsible AI practices:
Has knowledge
Has some knowledge
No knowledge
European participants were more likely to correctly identify what Responsible AI referred to, possibly due to GDPR awareness.
Confidence in the knowledge doesn’t mean you know
Developers with AI experience were more confident in their knowledge of what responsible AI use is about, but more experience didn’t always result in more knowledge about responsible AI.
Responsible AI label was confusing
Participants had a hard time figuring out that Responsible AI label was about responsible use. Based on this, the label was reworded to Responsible Use of AI.
It’s about being responsible: Participants preferred labels with a word responsible in them
Labels ‘Integration and use’ and ‘Integration and responsible use’ were tested for preferences. First I created 2 decoy labels similar to the ones being tested, that mimicked these labels (one with a word responsible and another without). I gave participants description of the content that would sit under this label and asked them to choose the most appropriate label for content.
Participants liked the ‘Integration and responsible use’ the most, but we also noted that preferences were around labels with the word responsible in them.
Reflexions
This was challenging research, that generated some insights for me too. I realized that asking participants to name things does not always produce good data. One of the questions in the survey asked participants to suggest a name for a content we gave them a description of. We initially thought this would be a great idea and results would be full of good suggestions but we only got a couple of good suggestions. Participants mostly came up with recommendations that were closely related to words describing a content.
This inspired me to look more closely into methodologies for research around labels and terms.
Next steps
Tree testing
For next steps I would give participants tasks on responsible AI and see if they can find a label that would guide them to the relevant information. This is something I did later, for the Face API team.