Recently, we gathered at the OSF offices for the Technology Salon on “Automated Decision Making in Aid: What could possibly go wrong?” with lead discussants:
- Jon Truong, and Elyse Voegeli, two of the creators of Automating NYC;
- Genevieve Fried and Varoon Mathur, Fellows at the AI Now Institute at NYU.
To start off, we asked participants whether they were optimistic or skeptical about the role of Automated Decision-making Systems (ADS) in the aid space. The response was mixed: about half skeptics and half optimists, most of whom qualified their optimism as “cautious optimism” or “it depends on who I’m talking to” or “it depends on the day and the headlines” or “if we can get the data, governance, and device standards in place.”
What are ADS?
Our next task was to define ADS. (One reason that the New York City ADS task force was unable to advance is that its members were unable to agree on the definition of an ADS).
One discussant explained that NYC’s provisional definition was something akin to:
- Any system that uses data algorithms or computer programs to replace or assist a human decision-making process.
This may seem straightforward, yet, as she explained, “if you go too broad you might include something like ‘spellcheck’ which feels like overkill. On the other hand, spellcheck is a good case for considering how complex things can get. What if spellcheck only recognized Western names? That would be an example of encoding bias into the ADS. However, the degree of harm that could come from spellcheck as compared to using ADS for predictive policing is very different. Defining ADS is complex.”
Other elements of the definition of an ADS are that it includes computational implementation of an algorithm. Algorithms are basically clear instructions or criteria followed in order to make a decision. Algorithms can be manual. ADS include the power of computation, noted another discussant. And perhaps a computer and complex system should be included as well, and a decision-making point or cut off; for example, an algorithm that determine who gets a loan. It is also important to consider statistical modeling and forecasting, which allow for prediction.
Using data and criteria for making decisions is nothing new, and it’s often done without specific systems or computers. People make plenty of very bad decisions without computers, and the addition of computers and algorithms is sometimes considered a more objective approach, because instructions can be set and run by a computer.
Why are there issues with ADS?
In practice things are not as clear cut as they might seem, explained one of our discussants. We live in a world where people are treated differently because of their demographic identity, and curation of data can represent some populations over others or misrepresent certain populations because of how they have been treated historically. These current and historic biases make their way into the algorithms, which are created by humans, and this encodes human biases into an ADS. When feeding existing data into a computer so that it can learn, we bring our historical biases into decision-making. The data we feed into an ADS may not reflect changing demographics or shifts in the data, and algorithms may not reflect ongoing institutional policy changes.
As another person said, “systems are touted as being neutral, but they are subject to human fallacies. We live in a world that is full of injustice, and that is reflected in a data set or in an algorithm. The speed of the system, once it’s computerized, replicates injustices more quickly and at greater scale.” When people or institutions believe that the involvement of a computer means the system is neutral, we have a problem. “We need to take ADS with a grain of salt, similar to how we tell children not to believe everything they see on the Internet.”
Many people are unaware of how an algorithm works. Yet over time, we tend to rely them and believe in them as unbiased truth. When ADS are not monitored, tested, and updated, this becomes problematic. ADS can begin to make decisions for people rather than supporting people in making decisions, and this can go very wrong, for example when decisions are unquestioningly made based on statistical forecasting models.
Are there ways to curb these issues with ADS?
Consistent monitoring. ADS should also be monitored constantly over time by humans. One Salon participant suggested setting up checkpoints in the decision-making process to alert humans that something is afoul. Another suggested that research and proof of concept is critical. For example, running the existing human-only system alongside the ADS and comparing the decisions over time helps to flag differences that can then be examined to see which of the processes is working better and to adjust or discontinue the ADS if it is incorrect. (In some cases, this process may actually flag biases in the human system). Random checks can be set up as can control situations where some decisions are made without using an ADS so that results can be compared between the two.
Recourse and redress. There should be simple and accessible ways for people affected by ADS to raise issues and make complaints. All ADS can make mistakes – there can be false positives (where an error points falsely to a match or the presence of a condition) and false negatives (where an error points to the absence of a match or a condition when indeed it is present). So there needs to be recourse for people affected by errors or for cases where biased data is leading to further discrimination or harm. Anyone creating and ADS needs to build in a way for mistakes to be managed and corrected.
Education and awareness. A person may not be aware that an ADS has affected them, and they likely won’t understand how an ADS works. Even people using ADS for decisions about others often forget that it’s an ADS deciding. This is similar to how people forget that their newsfeed on Facebook is based on their historical choices in content and their ‘likes’ and is not a neutral serving of objective content.
Improving the underlying data. Algorithms will only get better when there are constant feedback loops and new data that help the computer learn, said one Salon participant. Currently most algorithms are trained on highly biased samples that do not reflect marginalized groups and communities. For example, there is very little data about many of the people participating in or eligible for aid and development programs.
So we need proper data sets that are continually updated if we are to use ADS in aid work. This is a problem, however, if the data that is continually fed into the ADS remains biased. One person shared this example: If some communities are policed more because of race, economic status, etc., there will continually be more data showing that people in those communities are committing crimes. In whiter or wealthier communities, where there is less policing, less people are arrested. If we update our data only from communities where the crime rates are historically higher (because they are policed more and thus have more arrest records), we are simply creating a feedback loop that confirms our existing biases.
Privacy concerns also enter the picture. We may want to avoid collecting data on race, gender, ethnicity or economic status so that we don’t expose people to discrimination, stigma, or harm. For example, in the case of humanitarian work or conflict zones, sensitive data can make people or groups a target for governments or unfriendly actors. However, it’s hard to make decisions that benefit people if their data is missing from the database. It ends up being a catch 22.
Transparency is another way to improve ADS. “In the aid sector, we never tell people how decisions are made, regardless of whether those are human or machine-made decisions,” said one Salon participant. When the underlying algorithm is obscured, it cannot be reviewed for value judgments. Some compared this to some of the current non-algorithmic decision-making processes in the aid system (which are also not transparent) and suggested that aid systems could get more intelligent if they began to surface their own specific biases.
The objectives of the ADS can be reviewed. Is the system used to further marginalize or discriminate against certain populations, or can this be turned on its head? asked one discussant. ADS could be used to try to determine which police officers might commit violence against civilians rather than to predict which people might commit a crime. (See the Algorithmic Justice League’s work).
ADS in the aid system – limited to the powerful few?
Because of the underlying challenges with data (quality, standards, lack of) in the aid sector, ADS is still a challenge. One area where data is available and where ADS are being built and used is in supply chain management, for example, at massive UN agencies like the World Food Program.
Some questioned whether this exacerbates concentration of power in these large agencies, running counter to agreed-upon sector goals to decentralize power and control to smaller, local organizations who are ‘on the ground’ and working directly in communities. Does ADS then bring even more hierarchy, bias, and exclusion into an already problematic system of power and privilege? Could there be ways of using ADS differently in the aid system that would not replicate existing power structures? Could ADS itself be used to help people see their own biases? “Could we build that into an ADS? Could we have a read out of decisions we came to and then see what possible biases were?” asked one person.
How can we improve trust in ADS?
Most aid workers, national organizations, and affected communities have a limited understanding of ADS, leading to lower levels of trust in ADS and the decisions they produce. Part of the issue is the lack of participation and involvement in the design, implementation, validation, and vetting of ADS. On the other hand, one Salon participant pointed out that given all the issues with bias and exclusion, “maybe they would trust an ADS even less if they understood how an ADS works.”
Involving both users of an ADS and the people affected by ADS decisions is crucial. This needs to happen early in the process, said one person. It shouldn’t be limited to having people complain or report once the ADS has wronged them. They need to be at the table when the system is being developed and trialed.
If trust is to be built, the explainability of an algorithm needs consideration. “How can you explain the algorithm to people who are affected by it? Humanitarian workers cannot describe if they don’t understand it. We need to find ways to explain ADS to a non-technical audience so that they can be involved,” said one person. “We’ve shown sophisticated models to leaders, and they defaulted to spreadsheets.”
This brought up the need for change management if ADS are introduced. Involving and engaging decision-makers in the design and creation of ADS systems is a critical step for their adoption. This means understanding how decisions are made currently and based on what factors. Technology and data teams need to be in the room to understand the open and hidden nature of decision-making.
Isn’t decision making without ADS also highly biased and obscured?
People are often resistant to talking about or sharing how decisions have been made in the past, however, because those decisions may have been biased or inconsistent, based on faulty data, or made for political or other reasons.
As one person pointed out, both government and the aid system are deeply politicized and suffer from local biases, corruption and elite capture. A spatial analysis of food distribution in two countries, for example, showed extreme biases along local political leader lines. A related analysis of the road network and aid distribution allowed a clear view into the unfairness of food distribution and efficiency losses.
Aid agencies themselves make highly-biased decisions all the time, it was noted. Decisions are often political, situational, or made to enhance the reputation of an individual or agency. These decisions are usually not fully documented. Is this any less transparent than the ‘black box’ of an algorithm? Not to mention that agencies have countless dashboards that are aimed at helping them make efficient, unbiased decisions, yet recommendations based on the data may run counter to what is needed politically or for other reasons in a given moment.
Could (should) the humanitarian sector assume greater leadership on ADS?
Most ADS are built by private sector partners. When they are sold to the public or INGO sector, these companies indemnify themselves against liability and keep their trade secrets. It becomes impossible to hold them to account for any harm produced. One person asked whether the humanitarian sector could lead by bringing in different incentives – transparency, multi-stakeholder design, participation, and a focus on wellbeing? Could we try this and learn from it and develop and document processes whereby this could be done at scale? Could the aid sector open source how ADS are designed and created so that data scientists and others could improve?
Some were skeptical about whether the aid sector would be capable of this. “Theoretically we could do this,” said one person, “but it would then likely be concentrated in the hands of these few large agencies. In order to have economies of scale, it will have to be them because automation requires large scale. If that is to happen, then the smaller organizations will have to trust the big ones, but currently the small organizations don’t trust the big ones to manage or protect data.” And what about the involvement of governments, said another person, we would need to consider the role of the public sector.
“I like the idea of the humanitarian sector leading,” added one person, “but aid agencies don’t have the greatest track record for putting their constituencies in the driving seat. That’s not how it works. A lot of people are trying to correct that, but aid sector employees are not the people who will be affected by these systems in the end. We could think about working with organizations who have the outreach capacity to do work with these groups, but again, these organizations are not made up of the affected people. We have to remember that.”
How can we address governance and accountability?
When you bring in government, private sector, aid agencies, software developers, data, and the like, said another person, you will have issues of intellectual property, ownership, and governance. What are the local laws related to data transmission and storage? Is it enough to open source just the code or ADS framework without any data in it? If you work with local developers and force them to open source the algorithm, what does that mean for them and their own sustainability as local businesses?
Legal agreements? Another person suggested that we focus on open sourcing legal agreements rather than algorithms. “There are always risks, duties, and liabilities listed in contracts and legal agreements. The private sector in particular will always play the indemnity card. And that means there is no commercial incentive to fix the tools that are being used. What if we pivoted this conversation to commercial liability? If a model is developed in Manhattan, it won’t work in Malawi — a company has a commercial duty to flag and recognize that. This type of issue is hidden if we focus the conversation on open software or open models. It’s rare that all the technology will be open and transparent. What we should push for is open contracting, and that could help a lot with governance.”
Certification? Others suggested that we adapt existing audit systems like the LEED certification (which allows engineers and architects to audit whether buildings are actually environmentally sustainable) or the IRB process (external boards that review research to flag ethical issues). “What if there were a team of data scientists and others who could audit ADS and determine the flaws and biases?” suggested one person. “That way the entire thing wouldn’t need to be open, but it could still be audited independently”. This was questioned, however, in that a stamp of approval on a single system could lead people to believe that every system designed by a particular group would pass the test.
Ethical frameworks could be a tool, yet which framework? A recent article cited 84 different ethical frameworks for Artificial Intelligence.
Regulation? Self-regulation has proven to fail, said one person. Why aren’t we talking about regulation? The General Data Protection Regulation (GDPR) in Europe has a specific article (Article 22) about ADS that states that people have a right to know when ADS are used to made decisions that affect them, the right to contest decisions made by ADS, and right to request that humans review ADS decisions.
SPHERE Standards / Core Humanitarian Standard? Because of the legal complexities of working across multiple countries and with different entities in different jurisdictions (including some like the UN who are exempt from the law), an add-on to the SPHERE standards might be considered, said one person. Or something linked to the Core Humanitarian Standard (CHS), which includes a certification process. Donors will often ask whether an agency is CHS certified.
So, is there any good to come from ADS?
We tend to judge ADS with higher standards than we judge humans, said one Salon participant. Loan officers have been making biased decisions for years. How can we apply the standards of impartiality and transparency to both ADS and human decision making? ADS may be able to fix some of our current faulty and biased decisions. This may be useful for large systems, where we can’t afford to deploy humans at scale. Let’s find some potential bright spots for ADS.
Some positive examples shared by participants included:
- Human rights organizations are using satellite imagery to identify areas that have been burned or otherwise destroyed during conflict. This application of automated decision making doesn’t deal directly with people or allocation of resources, it supports human rights research.
- In California, ADS has been used to expunge the records of people convicted for marijuana-related violations now that marijuana has been legalized. This example supports justice and fairness.
- During Hurricane Irma, an organization in the Virgin Islands used an excel spreadsheet to track whether people met the criteria for assistance. Aid workers would interview people and the sheet would calculate automatically whether they were eligible. This was not high tech or sexy, but it was automated and fast. The government created the criteria and these were open and transparently communicated to people ahead of time so that if they didn’t receive benefits, they were clear about why.
- Flood management is an area where there is a lot of data and forecasting. Governments have been using ADS to evacuate people before it’s too late. This sector can gain in efficiency with ADS, which could be expanded to other weather-based hazards. Because it is a straightforward use case that involves satellites and less personal data it may be a less political space, making deployment easier.
- Drones also use ADS to stitch together hundreds of thousands of photos to create large images of geographical areas. Though drone data still needs to be ground truthed, it is less of an ethical minefield than when personal or household level data is collected, said one participant. Other participants, however, had issues with the portrayal of drones as less of an ethical minefield, citing surveillance, privacy, and challenges with the ownership and governance of the final knowledge product, the data for which was likely collected without people’s consent.
How can the humanitarian sector prepare for ADS?
In conclusion, one participant summed up that decision making has always been around. As ADS is explored more in-depth with groups like the one at this Salon and as we delve into the ethics and improve on ADS, there is great potential. ADS will probably never totally replace humans but can supplement humans to make better decisions.
How are we in the humanitarian sector preparing people at all levels of the system to engage with these systems, design them ethically, reduce harm, and make them more transparent? How are we working to build capacities at the local level to understand and use ADS? How are we figuring out ways to ensure that the populations who will be affected by ADS are aware of what is happening? How are we ensuring recourse and redress in the case of bad decisions or bias? What jobs might be created (rather than eliminated) with the introduction of more ADS?
ADS are not going to go away, and the humanitarian sector doesn’t have to wait until they are perfected to get involved in shaping and improving them so that they support our work in ethical and useful ways rather than in harmful or unethical ways.
Salons run under Chatham House Rule, so no attribution has been made in this post. Technology Salons happen in several cities around the world. If you’d like to join a discussion, sign up here. If you’d like to host a Salon, suggest a topic, or support us to keep doing Salons in NYC please get in touch with me!