idea42’s Josh Wright recently caught up with Tom Kalil, Chief Innovation Officer of Schmidt Futures, and former Deputy Director of the Office of Science and Technology Policy. One of the ideas that Tom is exploring is that science and technology can and should be playing a larger role in addressing societal challenges, particularly those challenges related to economic and social mobility.
As part of this exploration, Schmidt Futures is helping ideas42 launch a public call for ideas for how the integration of behavioral science and computer science can help enable upward mobility.
How did you get interested in the role that science and technology can play in addressing societal challenges?
I had an opportunity to work for two Presidents (President Clinton and President Obama) for 16 years. One of my jobs was to design and launch national science and technology initiatives that involved multiple agencies, like President Clinton’s nanotechnology initiative or President Obama’s BRAIN (Brain Research through Advancing Innovative Neurotechnologies) Initiative.
One of the things that I noticed over time is that some agencies, like DARPA, had the ability to set really ambitious goals and mobilize top researchers and entrepreneurs to meet them, while other agencies, like the Department of Labor, did not. The entire Department of Defense has a research budget of $66 billion, whereas the Department of Labor has a research budget of $4 million.
This huge difference in capacity shapes what agencies can do in important ways. For example, the Department of Defense was able to launch a program to reduce the time required for new Navy recruits to gain a technical skill from years to months, harnessing recent advances in AI. This is not something that the Department of Labor would have the ability to conceive of, let alone fund.
It occurred to me – what if the agencies that are responsible for promoting economic and social mobility (e.g. Labor, Education, HUD, the human services components of HHS) had a research arm like DARPA? What goals would it set? What research projects might it support?
While in the administration were you able to act on this insight?
I worked with President Obama’s Deputy Secretary of Education Jim Shelton on a proposal to create a DARPA for Education. President Obama included this in his budget, but we were unable to convince the Congress to support this idea.
Why do you think we should be investing more in science and technology to address societal challenges?
I think that if you are trying to make progress on a particular set of societal problems, and are generating a list of things to try, science and technology should be on the list.
That’s because we know more (e.g. the slow but steady increase in our understanding about how people learn or how we make decisions) and can do more based on the increased power and reach of technology, particularly digital technologies.
Having said that, I certainly don’t think it is the case that science and technology is going to solve all of our problems. I also think that we need to be more cognizant of the risks of technology, not just the benefits, and to develop technology-enhanced solutions in collaboration with the communities we are trying to empower as opposed to for them.
What do you mean by “societal challenges?”
In this context, I’m referring to problems that are making it more difficult for people to make the most of their talents and to enjoy a decent standard of living. Some examples include:
- By age 4, low-income children are exposed to 30 million fewer words than their more affluent peers. One study found that children of professionals also heard twice as many unique words and were engaged in twice as many “encouraging” as opposed to “discouraging” conversations.
- Only 20 percent of low-income students are proficient in 8th grade math.
- 36 million American adults are reading at the 3rd grade level or below.
- Twenty-five percent of Americans have no retirement savings or pension, and 40 percent do not have the savings to handle a $400 emergency expense.
What are some examples of science and technology applied to societal challenges?
Well, some examples are initiatives that ideas42 is currently pursuing – can the combination of behavioral insights and digital technologies reduce chronic absenteeism in K-12 schools, or reduce prison recidivism?
Greater progress may also be possible in workforce development. As you know, most workforce development programs have a modest impact on the wages of non-college educated workers. We have some preliminary evidence that it may be possible to do much better by combining AI and the science of learning to model the interaction between a novice and an expert.
Another example is depression. Entrepreneurs such Spencer Greenberg are exploring whether a smartphone app can deliver Cognitive Behavioral Therapy to help treat depression.
My hope is that the public call for ideas that ideas42 is issuing will encourage people in academia, civil society, and the private sector to suggest additional possibilities.
How could we begin to identify the opportunities to harness science and technology to solve societal problems?
I think a good next step would be to identify ideas and insights in four categories:
- Societal problems where the status quo is unacceptable, and where progress could reduce poverty, expand the middle class, and promote shared prosperity.
- Insights concerning how people learn, develop, make decisions, and form habits.
- Technological capabilities that can be used to amplify the impact of these insights about how people learn, make decisions, or form habits.
- Models for the definition, development, rigorous evaluation, and adoption of a successful intervention.
In other words, for a given societal problem, what is the role that a science-informed, technology-enabled intervention could play in addressing it? What roles should different public and private actors play in developing, evaluating and adopting these interventions?
For those of us in non-technical fields, what is different about the use of technology as opposed to other types of innovations in solving societal problems?
Technology and, more specifically, software and digital services or information have two overarching features that are helpful to consider when thinking about innovation.
- Low marginal cost: The cost of a traditional social service increases arithmetically with the number of people served. The marginal cost of making a digital product or service available to more people can be very low, although it may have a higher fixed cost.
- Continuous improvement: The productivity of public sector services is generally low or negative. For example, the U.S. has doubled real per pupil expenditures on K-12 education without significant improvements in learning outcomes. It’s plausible that digital techniques for rapid low-cost experimentation (A/B testing) and feedback loops could improve a service over time.
For those readers that are less aware of innovation concepts, can you explain what is meant by combinatorial innovation?
Google’s chief economist Hal Varian has made the case for the importance of “combinatorial innovation.” As he notes, occasionally we have a set of established existing technologies that “can be combined and recombined by innovators to create new devices and applications.” The Internet is a great example of combinatorial innovation because the components (programming languages, protocols, open standards, software libraries, etc.) are all “bits” that are faster and easier to combine and recombine to create new applications.
How might we foster combinatorial innovation for societal challenges?
By describing and cataloging relevant scientific insights from the social and behavioral sciences and by describing the capabilities of different technological building blocks, we could make it easier for people and teams to envision science-informed, technology-enabled interventions that can help address a given set of societal challenges.
For example, learning scientist Ken Koedinger has identified 30 mid-level theories about how people learn that have been validated experimentally and are specific enough to inform the design of learning technologies. Behavioral economist Sendhil Mullainathan has identified seven design principles for applying behavioral science to global development.
What are some examples of the technological building blocks that you have in mind?
Let me give 8 examples of the capabilities of digital technologies, with the caveat that this is just scratching the surface. What I think is important to do is to describe these in ways that are accessible to people who are not technologists.
For example, if you are involved in social policy and a computer scientist starts talking to you about convolutional neural networks and stochastic gradient descent – that is not terribly helpful. Instead, it might be useful for you to know that advances in machine learning are lowering the costs of prediction, or that with enough examples, we can train an algorithm to accurately map between an input (e.g. an image) and an output (a label that humans would use to describe the image).
- Sensing and measurement: Low-cost sensors or instrumented online environments are increasing our ability to measure everything. For example, companies have developed low-cost devices that allow parents to track the number of words that they speak to their children and the number of conversational turns that occur between parent and child, which could help address the word gap. Obviously, our increased ability to collect data raises important ethical issues, such as privacy, informed consent, and the fairness, accountability, and transparency of algorithmic decision-making.
- Personalization: Machine learning is improving our ability to deliver experiences that are personalized to the needs of an individual. For example, learning scientists are beginning to develop software that keeps students in the “Goldilocks” zone so that the material is hard enough to be challenging, but not so hard that they are frustrated and give up. Software can often recognize and address common student misconceptions that prevent them from mastering the material.
- Natural-language processing: Although computers are still very far away from human-level understanding of text and speech, natural-language processing (e.g. the ability to translate between two languages, answer questions, summarize a piece of text, etc.) is becoming increasingly useful.
- Prediction: With enough data, data scientists can develop predictive models that help prevent problems before they occur. Health plans are trying to identify which patients are at greatest risk of diseases such as stroke and diabetes to target prevention programs that can reduce the incidence of these conditions.
- Time on task: Time on task is clearly important for learning. Good game designers can create compelling and engaging experiences that increase time on task.
- Simulation: Simulation is also useful for training because it allows people to engage in learning by doing, in the same way that a would-be pilot uses a flight simulator before flying a plane.
- Matching: Some markets don’t rely solely on prices to clear. For example, if I want to attend a given university, it must be the case that I want to attend, and the university wants to accept me. Researchers at the intersection of computer science and economics are using “mechanism design” to solve societal problems that require this kind of matching, like helping a worker find a job that they would excel at.
- Location: Using satellites developed and deployed by the U.S. military, companies can provide location information that is accurate within 25 feet, and researchers have demonstrated systems that are accurate within one inch. For example, by collecting data on where asthma patients use their inhalers, Louisville and digital health companies have been able to identify and mitigate “hot spots” of asthma symptoms.
Are there any non-technological building blocks that people should consider?
There are some things that are not technologies in the traditional way we think about technologies, but are instead methodologies or ways of thinking (such as behavioral design, human-centered design, or system thinking) that I think are really important.
For example, when a Code for America team worked on increasing access to food stamps, they shadowed county workers, interviewed applicants, and actually went through the process of applying for food stamps themselves.
The combination of human-centered design and modern software engineering practices now allows Californians to apply for food stamps on a mobile phone, ask questions via online chat, take and upload images of documents needed to verify eligibility, and schedule any necessary interviews.
What are the implications of this for academia?
A number of universities are creating centers or initiatives that are at the intersection of computer science (CS) or data science and some particular problem – a trend that some are calling “CS + X.”
- In response to the explosive enrollment growth in CS, Northwestern is recruiting 20 CS faculty, but 10 of them will be in “CS +X” areas such as the intersection of computing and education.
- University of Illinois Urbana-Champaign is creating CS + X degree programs for students interested in the intersection of CS and other fields, such as crop science.
- Cornell has created a research center in computational sustainability, USC has a Center for Artificial Intelligence in Society, and UC Berkeley is the lead campus for the Center for Information Technology Research in the Interest of Society.
I think that programs like this that explored the intersection between CS and data science and some of the societal challenges we’ve been discussing would make a lot of sense.
What do you mean by “models for the development, evaluation and adoption of effective interventions?”
One key question is whether there are market failures that reduce the incentive of the private sector to invest in development, evaluation and adoption of effective interventions, and if there are, what are ways to address these market failures?
For example, it may be that the private sector has limited incentives to develop an intervention for adult literacy because the relevant population has little or no discretionary income to spend on adult literacy applications. Or it may be that there is a market, but the size, profitability, or riskiness of the market severely constrain what companies and investors are willing to support in terms of R&D or rigorous third-party evaluation.
The global health community has developed the capability to (1) identify unmet needs; (2) create a performance-based specification for the products that would meet those needs; and (3) design the incentives that would make it attractive for firms to develop these products.
For example, pharmaceutical companies have limited incentives to develop vaccines for diseases of the poor. To solve this problem, five countries and the Gates Foundation pledged to purchase a vaccine that could save the lives of 7 million children over the next 20 years by preventing diseases such as pneumonia and meningitis. This “Advance Market Commitment” motivated drug companies to develop the vaccine. The public sector agreed to bear the demand risk by pledging to purchase the vaccine at a given price, and the private sector agreed to bear the performance risk, which was whether they could develop a vaccine which was safe and effective.
I think this is an approach that we should be experimenting with domestically. I think it is unfortunate that governments have the ability to make trillions of dollars of financial commitments that are contingent on failure, such as loan guarantees, but are just scratching the surface when it comes to making financial commitments that are contingent on success, such as advance market commitments. This is an idea that Schmidt Futures is interested in exploring.
So that’s one model. But there are many others as well, such as:
- Traditional grants and contracts;
- Incentive prizes;
- Milestone payments that reward teams for intermediate progress towards the development of an effective intervention;
- Hybrid value chains, where public or philanthropic support of one component of the value chain would enable a sustainable business model in another part of the value chain; and
- Impact investing in firms or social enterprises, where investors are motivated more by social impact than financial return.
What are the policy implications of the ideas we’ve been discussing?
I think there are at least three policy implications.
The first is that the Executive Branch and the Congress should increase the capacity of agencies that work on promoting economic and social mobility to harness science and technology. For example, the Department of Labor should have a research arm. It might have as a goal reducing the time for non-college educated workers to master a skill that is a ticket to the middle class from years to months. It might achieve this goal by sponsoring research in areas such as cognitive task analysis (what is it that top performers in a given job know and are able to do), and AI-based digital tutors. Since this is unlikely to happen in the next several years, this is an area where philanthropy can play a role by demonstrating what might be possible.
The second is that the government should increase its capacity to create incentives for innovations that would have a high social return, but a low private return. I was able to make some progress on this idea during the Obama Administration by working with Congress to pass legislation that gave every agency the authority to support incentive prizes of up to $50 million, but there is a lot more that remains to be done. The budget and appropriations process generally assumes that federal agencies are going to allocate funds in the same fiscal year that they receive them. If we want agencies to make financial commitments that are contingent on success, they may never need to spend any money (if no one achieves the relevant goal).
The third is that agencies that work on promoting economic and social mobility need the capacity to support interventions that have a high fixed cost but a low marginal cost. A science-informed, technology-enabled intervention that has been rigorously evaluated and that increases adult literacy and numeracy might have a high fixed cost, but a low marginal cost.
The reason that many agencies would find it difficult to support the development and evaluation of these types of interventions is that much of social spending is allocated as either payments to individuals or block grants to states and localities. In an area like adult literacy, no single actor is in a position to create a market for a high-impact technology-enabled intervention. That’s because the federal money that we allocate to adult literacy is divided between 50 states, and divided again in even smaller amounts to local service providers. I think there is a case for allocating some social spending to create markets for interventions that have high fixed costs and low marginal costs, and currently the public sector is not organized to do that.