How can you bring ethics to your work?
If you missed the last MLUX meetup on designing and using data science ethically, don’t worry! Here are our top four best practices from our expert panel that you can take back to your team.
Carry a handbook of case studies with you to start conversations.
Whether it’s in your head or written down, have a few case studies that you can talk to colleagues about. Use them as a platform to ask, “Is this us?” and start having conversations around the answers. Microsoft’s Tay didn’t start out nefarious intentions. Remember to think about how things might go wrong in the wild. The big takeaway from the panelists for this was: no one is going to be the ethical voice unless you are!
Build a multidisciplinary team for these conversations.
Going back to the idea that this is all of our responsibilities, regardless of job titles, start including everyone in these conversations! Some tangible ideas from our panelists are: have ‘Data Fridays’, where the whole team reviews the data, or build in a process during QA to talk about the data. The key for this to work is to stay away from shame and blame; that will only suffocate conversation.
Challenge your assumptions about your data and ask how its being collected, then test your model.
Part 1: In conversations with your team, ask critical questions about how the data is being collected, and what data may be missing. Ask “what assumptions are we making?” and “what could the answers to these questions mean for the model?” For example, predictive policing models often make the assumption that police department data is representative of all crimes committed. Because police recorded crime data is biased by the institutional context under which it is collected, there is potential for predictive policing models to perpetuate troubling patterns, such as over-policing in minority neighborhoods.
Part 2: Set up regular “data dips” to check that your data represents what you think it represents: meaning changes over time, this is a constant process not just a one-time check. Have the ability to test inputs and outputs so you can observe what the model is doing; question the outputs.
Use MLUX community to help you.
Lean on this community to gather stories for your case study handbook. If you’re not sure how to start these conversations internally, ask the MLUX community for advice on what worked for them and what didn’t.
Thank you to our amazing panelists:
Sherry Wong (moderator): Grant Writer at Electronic Frontier Foundation
Elizabeth Churchill: Director of User Experience at Google
Stuart Geiger: Staff Ethnographer at Berkeley Institute for Data Science
Kristian Lum: Lead Statistician at the Human Rights Data Analysis Group
Sign up for the newsletter to get first dibs on tickets to our next event!
Learn more about what MLUX does.