Our journey to the effective use of artificial intelligence (AI) in international development has gotten stuck. Articles and white papers warn of potential abuses of personal information and assert the data and algorithms driving AI are biased.
Responsible data proponents want to agree on principles immediately, before we proceed. Meanwhile, everyone in the ICT4D community is trying hard to wrap their heads around AI’s potential uses and their implications.
So far, some of the most thoughtful work considering AI in development has come from the humanitarian community. Examples include:
- Artificial Intelligence in Global Health also from USAID
- Making AI Work for International Development from USAID
- AI Project Evaluation Criteria from Technology Salon
- Doing No Harm in the Digital Era by ICRC
- Ethical Obligations for Humanitarian Information Activities by Harvard Humanitarian Initiative
- Ethics Guidelines for Trustworthy AI from the EU
- Toronto Declaration the first collective human rights statement on AI
While these publications are well-intentioned, they actually seem to be stymieing creativity for AI in ICT4D. They are trying to apply an emerging technology to our current way of doing business and use AI to reinforce traditional ways of thinking, where:
- Projects are designed to achieve results that sound impressive in their magnitude
- Organizations have all the agency and provide technical assistance to “recipients” and “beneficiaries”.
As per the status quo, organizations seeking to set up data protection guidelines are doing so on behalf of others. A relevant quote from the Signal Code states:
“This duty of care…applies to how humanitarians research, develop, innovate, test, and integrate new approaches for utilizing data, information, and information communication technologies (ICTs) into their work.”
But, what about the person who understands privacy and wants to leverage their own data to improve their life? It seems a way to break the intellectual logjam around AI would be to shift our approach to where agency resides, potentially and hopefully turning our approach to development upside down for good.
Introducing Hyper-Personalized Development
Rather than applying AI to an agriculture, education, health, economic growth, etc. development project at the national level, imagine instead you used AI to identify the farmers, teachers, health workers or entrepreneurs who AI determined were sure bets.
You would set yourself up to only work with the most capable and amenable people, those who precisely followed your technical advice, repaid loans on time or successfully grew their startups. You would leave everyone else out until your focus or needs evolved and build your highly-focused project around those individuals.
You see this decision-making process in other fields. In the health field you triage to assign degrees of urgency. In business, you make a cost-benefit analysis. In education, you prioritize spending decisions based on the different abilities of pupils.
Using this focus in an AI-enabled approach could be described as hyper-personalization in development: In order identify these most-likely-to-be successful people, you would collect as much data about the individuals and their ecosystem as possible, then use some form of AI to identify their exceptionalness.
You would further build a detailed understanding of individuals’ abilities and ambitions, then use this understanding to guide investment in and provide technical assistance to those individuals to achieve overall development outcomes.
Only now the traditional roles would switch: These people, whom we should describe as “investees”, rather than “beneficiaries”, would determine their own direction. Donors could assume a supportive, back-seat role through their faith in the AI tools and the selected individuals. This model could be possible, since AI can isolate and measure the degree of influence individuals exert within complex ecosystems. It is critical to understand that this ability means that AI is unavoidably personal in its application.
How Hyper-Personalization Changes Development
Also, while we could use AI in ways to improve development programs, its capacities present us a chance to fundamentally change how we go about development. For one, we could use AI in a way that moves from the current standard of implementing partners being responsible for success to implementing partners playing open-ended supportive roles to the investees, who take credit for their own successes.
There is fundamental difference between using AI to identify desirable participants in a development program and using it to such a degree of confidence that you completely forego a typical project’s Expected Outcomes.
For this approach to international development, we would have to accept that our work would no longer be humanitarian, in the sense of impartially helping as many people as possible. This might be tough, because it cuts against the grain of our instincts.
Consider how you feel about the prospect of a development project benefitting only seven people, but those seven people doing exceptionally well. The ratio of funds to “beneficiaries” doesn’t seem right, since we typically try to help as many people as possible with available resources, with larger numbers of people equaling better value for donor money.
You can see an M&E person dismayed (or pleased!) at the prospect of no longer being able to report self-reassuringly high numbers of workshop attendees. You can also imagine donors initially shocked at AI programs and then beginning to permanently require the use of AI in all aspects of program design.
Of course, donor-funded hyper-personalized approaches would necessarily require individuals to express a desire to participate in a program and to give permission for a donor or project to access and analyze the data about them held by third parties. There would need to be a check on whether individuals understood what this permission entailed.
The point is investees would have significantly more autonomy under a hyper-personalized approach. And, it should be said that you could have reasons to use a hyper-personalized approach to work with the neediest, the least likely to succeed, even if it were difficult to obtain the necessary data.
The Private Sector Already Uses Hyper-Personalization
Hyper-personalization may become one of many examples where international development follows in the footsteps of the private sector. Since the seminal 2004 Wired article, “The Long Tail”, commercial marketers have used available data on personal internet and mobile phone use to developed hyper-personalized, person-specific approaches that tailor products and services to the specific preferences and needs of each individual.
The approach implies a death of generalization, that large-scale activities would be collections of unique, heterogenous pieces. Companies have long understood that hyper-personalization requires you to actively listen to each individual’s constantly evolving needs and expectations.
Globally, the sale of personal data for hyper-personalization is already a massive money maker. According to Hyper-Personalization with MNO Subscriber Data the sale of personal data in 2016 earned MNOs $9.6 billion, a far greater number than the previous year. MNOs are already using AI to improve network efficiency and customer services. And MNOs in emerging markets certainly have the profit motive to further monetize the personal data transferred via their networks by over-the-top (OTT) services. And if, as a 2017 Caribou Digital report suggests, the advertising economy won’t work for social media in emerging markets, social media companies and MNOs may find ways of monetizing personal data shared via social media.
An answer to this question, which supports the prospects for hyper-personalization in development, focuses on the monetization of individuals’ data. Companies like Wibson focus on enabling consumers (primarily in developed countries for now) to sell their own data to companies using AI for marketing. UBDI seeks to offer “a new, ethical way for people to be compensated for the value of the data they produce”. For several years Tala has been using algorithms to determine people’s credit worthiness “regardless of financial history”. Orange’s investment arm has backed an AI credit scoring startup.
Hyper-Personalization Enabling Environment Exists
Arguably AI is already at a stage where we could pursue this fundamentally new international development paradigm. For starters, global access and use of social media indicate there would be ample data resources to feed the algorithms driving a hyper-personalized approach in development.
While the ITU says 3.9 billion people, or 51.2% of the world’s population has Internet access, the GSMA says there are approximately 5.1 billion unique mobile subscribers and, according to Statista, there are around 2.5 billion individual Facebook users. Facebook usage alone generates 54 data categories that could be fed into an algorithm evaluating an individual’s suitability for a hyper-personalized development program. Call detail records (CDRs) include dozens of data types. Use of over-the-top (OTT) services creates potentially endless data points on individual interests, skills and knowledge.
Some in the ICT4D community have long been ready to exploit these data resources. It has been 6 years since Orange conducted its first Data for Development challenge, an experiment that provided 2.5 billion “anonymized” records of 5 million Ivorian mobile users to organizations that thought they could do something impactful with it.
Last year several organizations set up the OPAL Project to promote ethical and scalable approaches to using data from the private sector, including by using algorithms to analyze data without ever possessing it.
And contrasting to the cautionary approach of most publications on the topic, last year the Digital Impact Alliance (DIAL) published:
- Unlocking MNO data to enhance public services and humanitarian efforts
- Leveraging Data for Development to Achieve your Triple Bottom Line,
Both reports encouraged mobile network operators (MNOs) to sell their customer data to development programs. The report argued that “governments and NGOs often lack the awareness and resources to realize the cross-sectoral benefits made possible by the structured availability of MNO data.”
DIAL’s Syed Raza said there are already examples of NGOs buying CDRs from MNOs for their projects. Perhaps this indicates we will we soon see solicitations from donors that include required of analysis of CDRs and other data types such as mobile transactions, mobile loan repayment rates, social network analyses, mobile phone type, etc.
Maybe one way USAID seeks to follow to the executive order encouraging government agencies to advance development of AI, issued by the White House on February 11th will be to begin requiring the use of mobile network data in projects.
Hyper-Personalized Project Plan
A hyper-personalized project would likely include the following steps (truncated for brevity’s sake, using a small business focus as an example):
- A donor decides it would like to support economic growth in a country by supporting new entrepreneurs. They evaluate whether they have sufficient data about local entrepreneurs. Social media use, mobile money transactions, and CDRs are obvious sources, but, depending on the technical focus, it may be useful and necessary to obtain unstructured data and use AI to extract data on willing individuals.
- Data sources identified, the donor develops its assumptions of how behaviors can be evaluated to predict individual entrepreneurs prospects the country’s economy.
- A donor advertises a development program and invites people to express their desire to be included. As a part of advertising program, it releases data literacy materials, explaining what the data will be used for, how it will be stored, etc.
- Individuals explicitly express their desire to have their data used.
- The donor obtains base data from individuals, companies, and others and analyzes it according to its pre-defined assumptions.
- The donor selects investees based on the initial data.
- Selected entrepreneur investees that decide to participate are provided in-depth data literacy training and given the opportunity of opting out or signing up.
- Investees give the donors access to their social media accounts, mobile money transfers, mobile phone usage data and other data identified as useful and available.
- The donor routes technical assistance to the investees. They may also choose to support important facets of the ecosystem around the investees. In an even bolder, yet minimalistic approach, they donor may initiate a cash transfer. Though, investees may benefit from other support as well.
- Investees proceed with their existing ventures. They take the lead but benefit from financial assistance, specific technical assistance, or other resources they request from the donor. To a great extent, the donor lets go of controlling inputs. It becomes a question of how much they trust the data and the algorithms to predict outcomes.
- As time progresses, the donor monitors outcomes to modify its assumptions and algorithms for potential future support. The investees may make some unexpected decisions, but generally are able to successfully grow their businesses.
- Eventually donors’ support comes to an end and, when it’s ready, a donor assesses the success of its investees, as well as any changes in the ecosystem stemming from the activities of their investees. It determines whether, based upon available resources its current assumptions, whether it will go another round.
Hyper-Personalization Development Benefits
There are several reasons why a hyper-personalized development approach is appealing:
Leverage and investment
If an algorithm is likely to identify fewer suitable candidates than a traditional development program, hyper-localization changes the character of the relationship between donor and recipient, because a program assisting them would proportionally rely on their success more. Hyper-personalization puts the recipient in the driver’s seat and transforms a project into an investment. It also potentially removes layers between them and allows for more iterative communications, rather than just reporting.
Efficiency
Why plant seeds in infertile ground? You choose the metaphor that questions why we should spend any type of resource on individuals who, for one reason or another, may not be a good investment. With hyper-personalized targeting abilities, spending money on people you knew wouldn’t be successful would no longer be justified. More positively, if your project is only going to be helping a small group, imagine how much more profound your discussions could get. And even though groups of investees could be small, it does not mean it your project isn’t designed for scale (see Digital Development Principle #3).
User-centered
A project with that few people would transform from being about the project itself to being about the individuals involved. Once you have determined that a woman entrepreneur is going to be successful because of her own merits, you should forego the strictures of a project, with its phases and end date, and embark on an open-ended journey with her, comfortable with the unknowns of time and direction.
Agency over Data
If an individual is interested in leveraging their digital track record to be considered by a development program, a hyper-personalized development program gives some productive usefulness to users of social media, mobile phones, etc. who otherwise would never benefit from its collection.
It’s as though years of phone calls, e-mails and postings were small deposits in information bank accounts and now they can cash in. Your hyper-personalized program enables them to leverage the metadata companies have about them into a virtual resume.
There may indeed be a market for personal algorithm services that assemble and organize these data as a service.
- The open-source Data Transfer Project was founded in 2017 so that “individuals…could easily move their data between online service providers whenever they want.”
- The for-profit me offers a service of securely transferring your data to 3rd parties.
These efforts to return control to the individual contribute to the idea that digital development should be about individuals—their specific needs and potential.
Ecosystem Approach
Inevitably and paradoxically, hyper-personal approaches address the ecosystem simultaneously. The ratio of funding to investee requires an even greater understanding of the ecosystem. This is because an analysis of a single person’s appropriateness for a program inherently means an analysis of their relationship to the world around them—their relative qualities.
Analysis of multiple people’s appropriateness may see some third-party individuals show up repeatedly. It shows that this approach should involve social network and value chain analysis by default. It follows that
- A project supporting a group of farmers, entrepreneurs, etc., may need to support maintaining or even enhancing the role certain nodes play in a system;
- An individualized approach is simultaneously an ecosystem approach, albeit limited to the individuals your criteria and analysis indicate are indispensable.
Even if you wanted AI to design your project, you would likely find it would include people – influencers, critical exchanges, would span sectors as we have traditionally divided them. For ecosystem mapping, you might use R or a tools from Graphika, which specializes “in mapping and analyzing the complex fabric of social network structures, or…‘cyber-social terrain’”.
Harder Homework
This approach challenges donors to increase the substance and specificity of their assumptions of what “works”.
Hyper-Personalization Questions and Concerns
There are likely many aspects of hyper-personalization in development that will raise concerns. Several are potential critiques worth mentioning and would benefit from detailed consideration. Here are a dozen potential critiques, questions and concerns, which are included in truncated form to serve as place markers for future discussions:
- Is hyper-personalized development just another example of categorical targeting
- Isn’t it true that the intermediary’s (implementing partner’s) role would still be needed and wouldn’t change that much, meaning that hyper-personalization really isn’t that different from traditional development?
- Isn’t this just another example of techno-utopianism, with countless likely unintended consequences?
- Will this extend digital data colonialism and exasperate digital disparities?
- Will providing technical and/or financial assistance to such a small group of individuals result in negative attention for them and the donor?
- Will offering such focused assistance in exchange for people sacrificing their privacy amount to coercion?
- Even with AI, will you miss potential investees? Will the habit of maintaining multiple SIMs and social media accounts cause confusion?
- Will you create perverse incentives for people’s online behavior, where they try to game the system?
- Even if a person wants to give you the ability to use their data, would it even be legal?
- Do you actually need AI to do hyper-personalized development?
- Isn’t meaningful consent impossible for people with low levels of digital literacy?
- Aren’t there development areas, such as humanitarian and democracy promotion programs, where this type of project would never work?
- If development is a manifestation of foreign policy, would bilateral donors ever give up this much control?
These are just the start of good questions and concerns we should have about hyper-personalization. Yet, for all these concerns, one should keep in mind that there are many ways to implement a hyper-personalized approach. And it is likely the approach to using AI tools will inevitably evolve over time, limiting the ability to game the system.
The limited ability of data in information-poor environments does seem like a valid concern and may bias programs towards those living in more connected areas. And there are many capable entrepreneurs in Sub-Saharan Africa who might miss out because their information sources don’t include Facebook, TV, newspapers, and other sources where donors traditionally advertise programs.
Also, transparency of selection criteria does seem challenging: On one hand you would expect donors to provide transparency around how people are being judged. On the other, your program’s impact would be limited if people were able to game the system. Yet, if donors decide they should be closed with their criteria, at least they could be transparent with their reasons for doing so.
It is also true that people’s behavior with their mobile phones and online accounts is less than clean. People often switch MNOs in pursuit of short-term deals. And, according to Caribou,
“It’s also common for emerging market users to create multiple or successive social media or email accounts—sometimes intentionally, sometimes simply because it’s easier than recovering lost credentials, sometimes because the person they purchased the device from set it up with random account information that the new owner will never use—which fragments profile activity and further dilutes the value of the data to advertisers”.
And, while Call Data Recrords and activity on social media are undoubtedly useful, a need for even greater insight on the abilities of willing individuals may make AI necessary, so that unstructured data could be identified, analyzed and included in identifying the best candidates for a hyper-personalized development program.
A key final question relates to human behavior as expressed by a bureaucracy’s behavior. Hyper-personalization raises the further question of how close donor agencies want to be to the ultimate beneficiaries, especially if the donor-beneficiary ratio raises the stakes. In addition to delivering technical assistance, implementing partners play the role of keeping beneficiaries at arm’s length.
Hyper-Personalization Data Types & Specifics
One final important consideration for AI-enabled hyper-personalized development, is that there is a big difference between personal data extracted from social media use and from mobile network use. While social media use data exists “out there” in the more nebulous web, CDRS are possessed in databases owned by individual companies.
A heavier reliance on CDRs than social media usage data seems good, because there is arguably a closer connection to one’s CDR data than to the personal and meta data collected by apps and browsers. Why? Because the data types are simpler and capture 1:1 representations of your actions. You called your friend in Abuja from Lagos for 7 minutes and 3 seconds on Monday evening, starting at 5:43. Your phone’s call history mirrors exactly the CDR information.
Contrast this to the amount of time you spend on a section of a page before scrolling down, the frequency you click on news articles containing dozens of attributes you don’t consciously consider (though the browser remembers).
Thankfully there is a lot written about CDR use in development, and CDR analysis is also more accessible for anyone looking to implement a hyper-personalized project, including:
- Politics and Ethics of CDR Analytics
- From seconds to months: multi-scale dynamics of mobile telephone calls
- Flowminder’s now open-sourced FlowKit CDR analysis tool
- Considerations for Using Data Responsibly at USAID (forthcoming)
Are You Ready for Hyper-Personalization Development?
AI tools analyze the role and influence of the smallest units in an ecosystem, which in development programs are people and their actions. Its gift for granularity means AI should be thought of as necessarily personal in character. This means we can and should change the way we do development, bolding harnessing the fate of our programs on individuals.
The idea behind hyper-personalization is about boldly letting go of control over a program, trusting in AI, trusting in your ability to use it, while still holding responsibility for the results and purposefully adjusting where we fail. If we really want to do this in a way that empowers people, it should be on their terms. This opportunity to abdicate requires the courage to become a passenger on a journey that ventures off in unknown directions.
Yet, as an ICT4D sector, we might not yet have the expertise required to launch a hyper-personal development program. The present state of AI expertise within development organizations feels like the time when mobile data collection apps appeared: Their potential seemed obvious and tremendous, but there was little expertise or experience in using them. So as a first step, we should skill up by closer collaboration with experienced companies. And then we start building and learning from personalized approaches.
It is both too early and too late to work on principles for AI in development. And calling for principles now is indicative of our sector’s insularity, since private sector actors are way ahead of us and the recent literature from our sector doesn’t sufficiently seek to learn from those experiences. Even though our data sources are thinner and younger, we aren’t yet implementing enough to gain enough experiences that would help in developing principles. To date, “Do no harm” has mostly meant doing nothing and not figuring out through experience how to avoid harm.
For all the valid concern of what social media companies do with our personal data, this is not a call that proposes reforming those companies’ practices. Rather, this is a push to change what that data mean for people in developing countries who, to now, have been actors in the play for which we wrote the script, did the directing and wrote the reviews. If we trust more and move away from the inherently paternalistic approach in traditional development, we would be more likely to achieve an outcome we didn’t necessarily anticipate, but that works better for the people in question.
BE ON THE LOOKOUT FOR AN EVENT ON AI IN DEVELOPMENT AT FHI 360 THIS SPRING.
By Troy Etulain, FHI 360 Digital Development Department
Sorry, the comment form is closed at this time.