Creating Ethical Algorithms – Data on Purpose Live Blog

This is a live-blog from the Stanford Data on Purpose / Do Good Data “From Possibilities to Responsibilities” event. This is a summary of what the speakers at the talked about, captured by Rahul Bhargava and Catherine D’Ignazio. Any omissions or errors are likely my fault.

Human-Centered Data Science for Good: Creating Ethical Algorithms

Zara Rahman works at both Data & Society and the Engine Room, where she helps co-ordinate the Responsible Data Forum series of events. Jake Porway founded and runs DataKind.

Jake notes this is the buzzkill session about algorithms. He wants us all to walk away being able to critically assess algorithms.

How do Algorithms Touch our Lives?

They invite the audience to sketch out their interactions with digital technologies over the last 24 hours on a piece of paper. Stick figures and word totally ok. One participant drew a clock, noting happy and sad moments with little faces. Uber and AirBnb got happy faces next to them. Trying to connect to the internet in the venue got a sad face.  Here’s my drawing.

Next they ask where people were influenced by algorithms. One participant shares the flood warning we all received on our phones. Another mentioned a bot in their Slack channel that queued up a task. Someone else mentions how news that happened yesterday filtered down to him; for instance Hans Rosling’s death made it to him via social channels much more quickly than via technology channels. Someone else mentioned how their heating had turned on automatically based on the temperature.

What is an Algorithm?

Jake shares that the wikipedia-esque definition is pretty boring. “A set of rules that precisely deinfes a sequence of operations”. These examples we just heard demonstrate the reality of this. These are automated and do things on their own, like Netflix’s recommendation algorithm. The goal is to break down how these operate, and figure out how to intervene in what drives these thinking machines. Zara reminds us that even if you see the source code, that doesn’t help really understand it. We usually just see the output.

Algorithms have some kind of goal they are trying to get to. It takes actions to get there. For Netflix, the algorithm is trying to get you to watch more movies; while the actions are about showing you movies you are likely to want to watch. It tries to show you movies you might like; there is no incentive to show you a movie that might challenge you.

Algorithms use data to inform their decisions. In Netflix, the data input is what you have watched before, and what other people have been watching. There is also a feedback loop, based on how it is doing. It needs some way to figure out it is doing a good thing – did you click the movie, how much of it did you watch, how many star did you give it. We can speculate about what those measurements are, but we have no way of knowing their metrics.

A participant asks about how Netflix is probably also nudging her towards content they have produced, since that is cheaper for them. The underlying business model can drive these algorithms. Zara responds that this idea that the algorithm operates “for your benefit” is very subjective. Jake notes that we can be very critical about their goal state.

Another participant notes that there are civic benefits; in how Facebook can influence how many people are voting.

The definition is tricky, notes someone else, because anything that runs automatically could be called an algorithm. Jake and Zara are focused in on data-driven algorithms. They use information about you and learning to correct themselves. The purest definition and how the word is used in media are very different. Data science, machine learning, artificial intelligence – these are all squishy terms that are evolving.

Critiquing Algorithms

They suggest looking at Twitter’s “Who to follow” feature. Participants break into small groups for 10 minutes to ask questions about this algorithm. Here are the questions and some responses that groups shared after chatting:

  • What is the algorithm trying to get you to do?
    • They want to grow their user base, and then shifted to growing ad dollars
    • Showing global coverage, to show they are the network to be in
    • People name some unintended consequences like political polarization
  • What activities does it use to do that?
  • What data drives these decisions?
    • Can you pay for these positions? There could be an agreement based on what you are looking at and what Twitter recommends
  • What data does it use to evaluate if it is successful?
    • It can track your hovers, clicks, etc. both on the recommendation and adds later on
    • If you don’t click to follow somewhere that could be just as much signal
    • They might track the life of your relationship with this person (who you follow later because you followed their recommendation, etc)
  • Who has the power to influence these answers?

A participant notes that there were lots of secondary outcomes, which affected other people’s products based on their data. Folks note that the API opens up possibilities for democratic use and use for social good. Others note that Twitter data is highly expensive and not accessible to non-profits. Jake notes problems with doing research with Twitter data obtained through strange and mutant methods. Another participant notes they talked about discovering books to read and other things via Twitter. These reinforced their world views. Zara notes that these algorithms reinforce the voices that we hear (by gender, etc). Jake notes that Filter Bubble argument, that these algorithms reinforce our views. Most of the features they bake in are positive ones, not negative.

But who has the power the change these things? Not just on twitter, but health-care recommendations, Google, etc. One participant notes that in human interactions they are honest and open, but online he lies constantly. He doesn’t trust the medium, so he feeds it garbage on purpose. This matches his experiences in impoverished communities, where destruction is a key/only power. Someone else notes that the user can take action.

A participant asks what the legal or ethical standards should be. Someone responds that in non-profits the regulation comes from self-regulation and collective pressure. Zara notes that Twitter is worth nothing without it’s users.


Jake notes that we didn’t talk about it directly, but the ethical issues come up in relation to all these questions. These systems aren’t neutral.

Practicing Data Science Responsibly

I recently gave a short talk at a Data Science event put on by Deloitte here in Boston.  Here’s a short write up of my talk.

Data science and big data driven decisions are already baked into business culture across many fields.  The technology and applications are far ahead of our reflections about intent, appropriateness, and responsibility.  I want to focus on that word here, which I steal from my friends in the humanitarian field.  What are our responsibilities when it comes to practicing data science?  Here are a few examples of why this matters, and my recommendations for what to do about it.


People Think Algorithms are Neutral

I’d be surprised if you hadn’t heard about the flare-up about Facebook’s trending news feed recently.  After breaking on Gizmodo if has been covered widely.  I don’t want to debate the question of whether this is a “responsible” example or not.  I do want to focus on what it reveals about the public’s perception of data science and technology.  People got upset, because they assumed it was produced by a neutral algorithm, and this person that spoke with Gizmodo said it was biased (against conservative news outlets).  The general public thinks algorithms are neutral, and this is a big problem.


Algorithms are artifacts of the cultural and social contexts of their creators and the world in which they operate.  Using geographic data about population in the Boston area?  Good luck separating that from the long history of redlining that created a racially segregated distribution of ownership.  To be responsible we have to acknowledge and own that fact.  Algorithms and data are not neutral third parties that operate outside of our world’s built-in assumptions and history.

Some Troubling Examples

Lets flesh this out a bit more with some examples.  First I look to Joy Boulamwini, a student colleague of mine in the Civic Media group at the MIT Media Lab.   Joy is starting to write about “InCoding” – documenting the history of biases baked into the technologies around us, and proposing interventions to remedy them. One example is facial recognition software, which has consistently been trained on white male faces; to the point where she has to literally done a white-face mask to have the software recognize her.  This just the tip of the iceberg in computer science, which has a long history of leaving out entire potential populations of users.


Another example is a classic one from Latanya Sweeney at Harvard.  In 2013 She discovered a racial bias trained into the operation Google’s AdWords platform.  When she searched for names that are more commonly given to African Americans (liked her own), the system popped up ads asking if the user wanted to do background checks or look for criminal records.  This is an example of the algorithm reflecting built-in biases of the population using it, who believed that these names were more likely to be associated with criminal activity.

My third example comes from an open data release by the New York City taxi authority.  They anonymized and then released a huge set of data about cab rides in the city.  Some enterprising researchers realized that they had done a poor job of anonymizing the taxi medallion ids, and were able to de-anonymize the dataset.  From there, Anthony Tockar was able to find strikingly juicy personal details about riders and their destinations.

A Pattern of Responsibility

Taking a step back form these three examples I see a useful pattern for thinking about what it means to practice data science with responsibility.  You need to be responsible in your data creation, data impacts, and data use.  I’ll explain each of those ideas.


Being responsible in your data collection means acknowledging the assumptions and biases baked into your data and your analysis.  Too often these get thrown away while assessing the comparative performance between various models trained by a data scientist.  Some examples where this has failed?  Joy’s InCoding example is one of course, as is the classic Facebook “social contagion” study. A more troubling one is the poor methodology used by US NSA’s SkyNet program.

Being responsible in your data impacts means thinking about how your work will operate in the social context of its publication and use.  Will the models you trained come with a disclaimer identifying the populations you weren’t able to get data from?  What are secondary impacts that you can mitigate against now, before they come back to  bite you?  The discriminatory behavior of the Google AdWords results I mentioned earlier is one example. Another is the dynamic pricing used by the Princeton Review disproportionately effecting Asian Americans.  A third are the racially correlated trends revealed in where Amazon offers same-day delivery (particularly in Boston).

Being responsible in your data use means thinking about how others could capture and use your data for their purposes, perhaps out of line with your goals and comfort zone.  The de-anonymization of NYC taxi records I mentioned already is one example of this.  Another is the recent harvesting and release of OKCupid dating profiles by researchers who considered it “public” data.

Leadership and Guidelines

The problem here is that we have little leadership and few guidelines for how to address these issues in responsible ways.  I have yet to find an handbook for a field that scaffolds how to think about these concerns. As I’ve said, the technology is far ahead of our reflections on it together.  However, that doesn’t mean that they aren’t smart people thinking about this.


In 2014 the White House brought together a team to create their report on Big Data: Seizing Opportunities, Preserving Values.  The title itself reveals their acknowledgement of the threat some of these approaches have for the public good.  Their recommendations include a number of things:

  • extending the consumer bill of rights
  • passing stronger data breach legislation
  • protecting student centered data
  • identifying discrimination
  • revising the Electronic Communications Privacy Act

Legislations isn’t strong in this area yet (at least here in the US), but be aware that it is coming down the pipe.  Your organization needs to be pro-active here, not reactive.

Just two weeks ago, the Council on Big Data, Ethics and Society released their “Perspectives” report.  This amazing group of individuals was brought together to create this report by a federal NSF grant.  Their recommendations span policy, pedagogy, network building, and area for future work.  The include things like:

  • new ethics review standards
  • data-aware grant making
  • case studies & curricula
  • spaces to talk about this
  • standards for data-sharing

These two reports are great reading to prime yourself on the latest high-level thinking coming out of more official US bodies.

So What Should We Do?

I’d synthesize all this into four recommendations for a business audience.


Define and maintain our organization’s values.  Data science work shouldn’t operate in a vacuum.  Your organizational goals, ethics, and values should apply to that work as well. Go back to your shared principles to decide what “responsible” data science means for you.

Do algorithmic QA (quality and assurance).  In software development, the QA team is separate from the developers, and can often translate between the  languages of technical development and customer needs.  This model can server data science work well.  Algorithmic QA can discover some of the pitfalls the creators of models might not.

Set up internal and and external review boards. It can be incredibly useful to have a central place where decisions are made about what data science work is responsible and what isn’t for your organization.  We discussed models for this at a recent Stanford event I was part of.

Innovate with others in your field to create norms.  This stuff is very new, and we are all trying to figure it out together.  Create spaces to meet and discuss your approaches to this with others in your industry.  Innovate together to stay ahead of regulation and legislation.

These four recommendations capture the fundamentals of how I think businesses need to be responding to the push to do data science in responsible ways.

This post is cross-posted to the website.

Ethical Data Review Procesess Workshop at Stanford

The Digital Civil Society Lab at Stanford recently hosted a small gathering of people to dig into emerging processes for ethical data review.  This post is a write up of the publicly shareable discussions there.


Lucy Berholz opened the day by talking about “digital civil society” as an independent space for civil society in the digital sphere.  She is specifically concerned with how we govern the digital sphere in line with the a background of democracy theory.  We need to use, manage, govern in ways that are expansive and supportive for independant civil society.  This requires new governance and review structures for digital data.

This prompted the question of what is “something like an IRB and not an IRB”?  The folks in the room bring together corporate, community, and university examples.  These encompass ethical codes and the processes for judging adherence to them. With this in mind, in the digital age, do non-profits need to change?  What are the key structures and governance for how they can manage private resources for public good?

Short Talks

Lucy introduced a number of people to give short talks about their projects in this space.

Lasanna Magassa (Diverse Voices Project at UW)

Lasanna introduced us all to the Diverse Voices Project, an “An exploratory method for including diverse voices in policy development for emerging technologies”. His motivations lie in the fact that tech policy is generally driven by mainstream interests, and that policy makers are reactive.

They plan and convene “Diverse Voices Panels”, full of people whole live an experience, institutions that support them, people somehow connected to them.  In a panel on disability this could be people who live it and are disabled, law & medical professionals, and family members.  These panels produce whitepapers that document and then make recommendations.  They’ve tackled everything from ethics and big data, to extreme poverty, to driverless cars. They focus on what technology impacts can be for diverse audiences. One challenge they face is finding and compensating panel experts. Another is wondering how to prep a dense, technical document for the community to read.

Lasanna talks about knowledge generation being the key driver, building awareness of diversity and the impacts of technologies on various (typically overlooked) subpopulations.

Eric Gordon (Engagement Lab at Emerson College)

Eric (via Skype) walked us through the ongoing development of the Engagement Lab’s Community IRB project.  The goal they started with was to figure out what a Community IRB is (public health examples exist).  It turned out they ran into a bigger problem – transforming relationships between academia and community in the context of digital data.  There is more and more pressure to use data in more ways.

He tells us that in Boston area, those who represent poorer folks in the city are asked for access to those populations all the time.  They talked to over 20 organizations about the issues they face in these partnerships, focusing on investigating the need for a new model for the relationships.  One key outcome was that it turns out nobody knows what an IRB is; and the broader language use to talk about them is also problematic (“research”, “data”).

They ran into a few common issues to highlight.  Firstly, there weren’t clear principles for assuring value for those that give-up their data.  In addition, the clarity of the research ask was often weak.  There was a all-to-common lack of follow-through, and the semester-driven calendar is a huge point of conflict.  An underlying point was that organizations have all this data, but the outside researcher is the expert that is empowered to analyze it.  This creates anxiety in the community organizations.

They talked through IRBs, MOUs, and other models.  Turns out people wanted to facilitate between organizations and researchers, so in the end what they need is not a document, but a technique for maintaining relationships.  Something like a platform to match research and community needs.

Molly Jackman & Lauri Kanerva (Facebook)

Molly and Lauri work on policy and internal research management at Facebook.  They shared a draft of the internal research review process used at Facebook, but asked it not be shared publicly because it is still under revision.  They covered how they do privacy trainings, research proposals, reviews, and approvals for internal and externally collaborative research.

Nicolas de Corders (Orange Telekom)

Nicolas shared the process behind their Data for Development projects, like their Ivory Coast and Senegal cellphone data challenges.  The process was highly collaborative with the local telecommunications ministries of each country.  Those conversations produced approvals, and key themes and questions to work on within the country.  This required a lot of education of various ministries about what could be done with the cellphone call metadata information.

For the second challenge, Orange set up internal and external review panels to handle the submissions.  The internal review panels included Orange managers not related to the project.  The external review panel tried to be a balanced set of people.  They built a shared set of criteria by reviewing submissions from the first project in the Ivory Coast.

Nicolas talks about these two projects as one-offs, and scaling being a large problem.  In addition, getting the the review panels to come up with shared agreement on ethics was (not surprisingly) difficult.


After some lunch and collaborative brainstorming about the inspirations in the short talks, we broke out into smaller groups to have more free form discussions about topics we were excited about.  These included:

  • an international ethical data review service
  • the idea of minimum viable data
  • how to build capacity in small NGOs to do this
  • a people’s review board
  • how bioethics debates can be a resource

I facilitated the conversation about building small NGO capacity.

Building Small NGO Capacity for Ethical Data Review

Six of us were particularly interested in how to help small NGOs learn how to ask these ethics questions about data.  Resources exist out there, but not well written enough for people in this audience to consume.  The privacy field especially has a lot of practice, but only some of the approaches there are transferrable.  The language around privacy is all too hard to understand for “regular people”.  However, their approach to “data minimization” might have some utility.

We talked about how to help people avoid extractive data collection, and the fact that it is disempowering.  The non-profit folks in the group reminded us all that you have to think about the funder’s role in the evidence they are asking for, an how they help frame questions.

Someone mentioned that law can be the easiest part of this, because it is so well-defined (for good or bad).  We have well established laws on the fundamental privacy right of individuals in many countries.  I proposed participatory activities to learn these things, like perhaps a group activity to try and re-identify “anonymized” data collected from the group.  Another participant mentioned DJ Patel’s approach to building a data culture.

Our key points to share back with the larger group were that:

  • privacy has inspirations, but it’s not enough
  • communications formats are critical (language, etc); hands-on, concrete, actionable stuff is best
  • you have to build this stuff into the culture of the org