Introduction To AI For Social Good

AI for Social Good — a relatively new research field at the intersection of AI and a number of other fields. Source

“Whenever I hear people saying AI is going to hurt people in the future I think, yeah, technology can generally always be used for good and bad and you need to be careful about how you build it … if you’re arguing against AI then you’re arguing against safer cars that aren’t going to have accidents, and you’re arguing against being able to better diagnose people when they’re sick.” — Mark Zuckerburg, CEO of Facebook

New technology is not good or evil in and of itself. It’s all about how people choose to use it. — David Wong, Professor of Philosophy at Duke University

From common media rhetoric, it is easy to perceive artificial intelligence (AI) as a technology that will automate us all out of jobs, perpetuate discrimination, stoke division, and may ultimately lead to the demise of humanity. However, we must learn to separate technology itself from its applications. Without getting into the philosophical realm of whether technology is morally-neutral (good arguments exist both in favor and against this statement), fundamentally, technology is a tool at our disposal.

The existence of a technology merely tells us what is possible (that is, what can I do with it), but nothing can tell us exactly what should be done with it. This is David Hume’s famous is-ought statement. Only through the rose-tinted spectacles of the human perspective can any normative value be derived, whether or not the philosophers deem it inherent to the technology. Should nuclear weapons ever be used in warfare? Should widespread surveillance be implemented to make the public safer? Should gene editing be used routinely to eradicate certain congenital diseases? Science cannot tell us.

If this in-depth educational content is useful for you, subscribe to our AI research mailing list to be alerted when we release new material.

Naturally, some technologies can be harnessed in more negative ways than others. For example, nuclear weapons have more negative associations than computers (as made obvious through the use of the word “weapon”). However, the technology underpinning nuclear weapons — nuclear fission — also gave us a new way to generate essentially carbon-free energy. We must come to realize that notions of good or bad do not come from the tool itself, but how it is wielded. AI is no exception.

In this article, I will introduce the reader to the nascent research field of “AI for Social Good”, also known as “AI for Social Impact”, along with some of the novel applications that aim to tackle some of the most important social, environmental, and public health challenges that exist today.

Why is AI stigmatized for being unethical?

AI is a subdiscipline of computer science. Computer scientists have often been criticized for their lack of consideration of the potential ethical and societal implications of their research. This criticism has intensified in recent years since the deep learning revolution and the information explosion, both of which have increased the influence and power that computing technology has on society.

This criticism is largely justified. In the aftermath of World War II, the Universal Declaration of Human Rights was created, which outlined the fundamental freedoms that every human should be afforded, without exception. The release of the Belmont Report in 1978 outlined the ethical principles and guidelines that must be followed by individuals performing research on human subjects, which had a profound impact on research methods in medicine and the social sciences.

Side note: What did the Belmont report say? It introduced three overarching principles to govern future research on human subjects:
(1) Respect for persons: requiring informed consent from participants with no deception.
(2) Beneficence: to maximize benefits and minimize risks to research subjects.
(3) Justice: the fair treatment of subjects through reasonable and rationalized procedures.

In contrast, computer science has remained largely unperturbed by limitations on its research methods. Historically, most computer science curriculums paid little to no emphasis on the teaching of ethics. This lack of emphasis is likely not deliberate, but instead because technology produced by computer scientists had relatively little impact on the lives of the general public.

That is no longer the case. Computer scientists now work with large datasets such as (1) medical data, creating algorithms to detect and diagnose things like cancer and modeling the spread of disease; (2) social media data, creating improved recommendation systems and other algorithms to increase revenue via enhanced consumer engagement; (3) financial data, creating algorithms to decide which individuals will receive loans or even which companies people should invest in. The impact that such creations can have on society has increased markedly, but still, few formal requirements have been introduced to enforce ethical practices in this field.

This rapid increase in power naturally leads to the exacerbation of conflicting constraints. The legal position is based upon legal precedent, which looks at how similar cases in the past were judged and how the ruling may apply in new circumstances. As with any new technology, there are currently few precedents, making the landscape largely unregulated. What is technically feasible therefore does not necessarily align with what is deemed legal. And neither of these visions are necessarily compatible with what organizations wish to pursue, or what the public would deem as “ethical”. Moreover, the goalpost for these constraints is always changing. The best we can hope to do is to morph these goalposts such that they begin to overlap more closely, instead of diverging further apart.

Only in recent years have individuals began to speak up and elucidate the dangers that unfettered computer science research might have. Consequently, we have seen a push towards ethical computer science practices, with a focus on transparency and accountability.

Institutions such as Harvard have begun to embed ethics into the computer science curriculum, and we have seen prestigious research journals such as Nature and Science have made it a requirement for datasets and analyses to be publicly available, also stating the provenance of the data (e.g., was it obtained with informed consent). Conferences such as the “Conference on Neural Information Processing Systems” (NeurIPS) have also required computer science papers to provide a broader impact statement, indicating any ethical or societal implications of their research work. This review is similar to those done by government initiatives or institutional review boards (IRBs) in universities and funding bodies.

Clearly, these measures do not solve the greater issue. Currently, they are few and far between, might be selectively applied, or only act as rubber stamps of approval, but they are the start of a conversation that was not previously being had. Tackling issues of individual privacy in large-scale datasets, such as the risk of re-identification, information disclosure, adverse impact, or lack of public trust in AI will require more extensive measures. Similar measures have been achieved in medicine and the social sciences without severely hindering research goals, and so this should not prevent a significant barrier to the computer science community.

What is A.I. for Social Good?

AI for social good (AI4SG) is a relatively new research field that focuses on tackling important social, environmental, and public health challenges that exist today using AI. Initially, this may sound like a gimmick to try and flip the traditionally negative view of AI on its head, but it is more than that. For simplicity, we can think of AI4SG as the intersection of AI with the social sciences as well as environmental science.

AI4SG is different from traditional use-cases of AI in that it uses more of a top-down approach. The field is focused on delivering positive social impact in accordance with the priorities outlined in the United Nations’ 17 Sustainable Development Goals (SDGs), shown below.

Instead of traditional AI applications like text translation, we might instead be more interested in modeling the social networks of homeless individuals in an attempt to counter the spread of HIV, which are disproportionately impacted by the disease.

While text translation can have a positive social impact, it would generally not be considered part of AI4SG. The same can be said for the majority of business intelligence applications, such as inventory planning, recommendation systems, etc.

Many applications of AI4SG utilize methods that are game-theoretic: they can be modeled mathematically using a game between an agent and an adversary. AI models based on game theory fall under the reinforcement learning paradigm and, since they require the use of more than one agent, are often described as multi-agent systems. This is an entire field in its own right and will likely become a game-changer in society as well as Industry 4.0.

As an example, LAX airport implemented a game-theoretic AI model to help create patrols that optimally defend against terrorist attacks and drug smugglers. This system looked at existing patrol strategies where adversaries were able to capitalize on inefficiencies. In contrast, the agent was able to notice these inefficiencies and construct a patrol policy that minimized the possibility that an adversary could capitalize on inefficiencies by essentially removing them. This result was achieved using a Bayesian Stackelberg game, which provided a seemingly randomized policy that optimally defended against adversaries.

To help elucidate what AI4SG entails in more detail, we will look at several definitions from the research literature.

Floridi et al. (2020) outlined seven factors that are essential for AI4SG applications, which are shown below. These principles dovetail with those outlined in the Belmont and Menlo reports but are closely focused on AI and the potential for abuse or maleficence use of the technology.

Tomašev et al. (2020) provided a similar set of guidelines for AI4SG which pertain to the overall use of AI technology (G1, G2, G3), applications (G4, G5, G6, G7, G8), and data handling (G9, G10).

It is not always possible for a specific application to provide a positive impact on all of the 17 SDGs. In fact, some technologies might have a positive impact on one of the goals, and a negative impact on a separate goal. To counter this, applications should aim to maximize a net positive effect on as many SDGs as possible, without causing avoidable harm to other SDGs.

The Association for the Advancement of Artificial Intelligence (AAAI) created AI4SG as an emerging topic at their annual conference in 2019, and have since outlined several criteria required for AI4SG applications:

• Significance of the problem. The social impact problem considered is significant and not adequately addressed by the AI community thus far.
• Novelty of approach. Introduces a new model or substantially improves upon existing models, data gathering techniques, algorithms, and/or data analysis techniques.
• Scope and promise for social impact. High likelihood of social impact of solution, possibly being used in practice or could be used immediately.
• Reliance upon and/or advancement of cutting edge AI techniques. Introduces novel or state-of-the-art AI techniques suitable to the problem being solved.

The most common technologies used for AI4SG today are outlined below, some of which have been discussed previously. The use of game theory, network theory, multi-armed bandits, Markov decision processes, reinforcement learning, and decision-focused learning are all common modeling approaches that draw upon the fields of AI and multi-agent systems.

Despite its fledgling form, AI4SG has already achieved some impressive results, some of which we will discuss in the following section.

Applications of AI4SG

In this section, we discuss six applications of AI4SG and how these impact the UN’s SDGs. These applications include (1) counting penguins from space, (2) countering elephant poaching, (3) substance abuse prevention, (4) the global immunological observatory for pandemic prediction, (5) social media analysis for mental health monitoring, and (6) Google’s Project Euphonia.

(1) Counting Penguins from Space

This is perhaps an odd example to begin with, but it is a very interesting application that can be used for non-invasive monitoring and conservation purposes and extended to other animal species, falling under SDG’s 13-15.

Researcher Heather Lynch received the Young Scientist Award for her work on counting penguins from space by examining their “guano”, or excreta. Their guano is a light pink color, making it easily visible to satellites such as Landsat-7. While difficult to spot the excreta of an individual penguin from space is challenging due to the limited spatial resolution, spotting the guano footprint of an entire penguin colony is within the realm of possibility.

This technique, outlined in “An Object-Based Image Analysis Approach for Detecting Penguin Guano in very High Spatial Resolution Satellite Images” is how Heather Lynch has been able to monitor where and how many colonies of various penguin species exist in the Antarctic peninsula, as well as their migration patterns. Her work has improved estimates of the number of Adelie penguins on the peninsula, of which current best estimates are 3.79 million breeding pairs.

Professor Lynch has produced estimates for Chinstrap penguins too, which also occupy the Antarctic peninsula, estimating the existence of 3.42 million breeding pairs across 375 penguin colonies.

This and similar techniques can be used to monitor changes in the population and demographics of species that are particularly vulnerable to climate change. This is relatively simple for penguins due to the contrast between the snow covering the Antarctic peninsula and the guano of a penguin colony, but similar biomarkers may make this a possibility for other species in the future.

(2) W.W.F. Countering Elephant Poaching in Africa

Countering poaching efforts highlights an important subset of game-theoretic models known as “Green Security Games”, which help to tackle SDG’s 1, 11, 15, and 16.

Poaching is a particularly problematic issue in Uganda. During the political turmoil of the 1970s/80s, poachers reduced the elephant population in Uganda from an estimated 30,000 elephants to fewer than 800. Efforts such as increased ranger patrols and conservation projects such as beehive fences have pushed this back up to around 3,000 elephants in Queen Elizabeth National Park, but human-elephant conflict remains an issue in the region. In Katara, some farmers have resorted to poisoning elephants to prevent them from trampling their crops, while others are hunted by poachers for ivory or meat. A considerable number of rangers have been killed in altercations with poachers which has led to a rise in widowed women in the region, leaving them impoverished and unable to provide for their families.

The rangers only have limited resources available at their disposal to monitor the elephants in the park, and thus need to organize those resources intelligently to give them the best chance of countering poaching in the region, this is where our green security games come into play.

Our green security game is based on a Stackelberg security game, focusing on threat prediction, first splitting up Queen Elizabeth National Park, a 2000 square kilometer region, into 1 km x 1 km grid cells.

The security game will involve determining the probability of a snare being placed in a certain grid square based on a variety of factors learned from 12 years of prior data from approximately 1,000 cases of poaching in the park.

A new patrolling scheme is then designed by the AI model to maximize the possibility of those snares being caught. This is performed iteratively with the past crime data being used to predict the approximate spatial locations of snare placement in the future.

When this approach was initially implemented for a month in the Park over two 9 square kilometer regions, they had a higher hit rate of snares than 91% of all prior months historically. Following this, a larger scale field test was conducted over a 6-month period across 27 regions of 9 square kilometers across the park and showed remarkably positive results, with a record number of snares being caught an even a poacher caught red-handed in the park.

Since these field experiments, these green security games leveraging AI have been expanded to Murchison Falls National Park, as well as other parts of Uganda by the Wildlife Conservation Society, and even in Cambodia by the World Wildlife Fund to prevent elephant poaching, which has become popular in recent years as an ingredient in certain chinese herbal medicines. When tested in Cambodia’s Srepok Wildlife Sanctuary, 521 snares were caught in the first month of implementing the green security game, compared to the typical 101 snares found by the region’s rangers.

Similar green security games can be expanded for use in tasks such as border patrol, terrorism prevention, coral reef protection, and even prevention of illegal gold mining in the Amazon rainforest.

(3) Substance Abuse Prevention

Tackling SDG’s 1 and 3 of the UN, substance abuse prevention has been one of the dominant focus areas for researchers in AI4SG. Substance abuse is an important public health issue that can have widespread negative social impacts, and can increase the prevalence of both mental and physical problems.

Youths are the most at-risk groups for substance abuse, with certain groups of youths engaging in very high levels of drug use, such as homeless youth. Social science research has shown that deviancy training can be useful, but often peer-based interventions are most effective.

Performing peer-based interventions is very similar to how influencers are chosen for marketing purposes. A subset of nodes in a social network are selected that have the highest probability of propagating information to the largest possible audience within the target population. This has been studied in homeless populations and found to be effective for combating a variety of deviant behaviors, including substance abuse.

Imagine we have a social network, G, like the one shown below, and are able to choose K nodes that are to be trained as peer leaders. Once the approximate social network, assuming an independent cascade model of information spread, we can optimize our algorithm to maximize the expected number of influenced nodes.

Obviously, the procedure is more messy and uncertain in practice for a variety of reasons, including peer leaders that drop out of the process, and the uncertainty in propagation probabilities and the structure of the social network.

To handle this uncertainty, Wilder et al. (2017) developed Robust Influence Maximization which uses a partially observed Markov decision process to iteratively select the optimal group of peer leaders in a social network for maximal information propagation.

Essentially, if a peer leader is a “no show”, then this peer leader is replaced with another that is optimally chosen to provide maximal coverage of the remaining network.

This technique of providing peer-based interventions can be applied to a broad range of tasks involving propagation through social networks, and as such is likely to see increased attention as our lives increasingly involve larger and larger social networks.

(4) Global Immunological Observatory for Pandemic Prediction

During the COVID-19 pandemic, Professor Michael Mina of the Harvard School of Public Health proposed a radical new idea which he coined the “Global Immunological Observatory”, which would act analogously to a “weather forecast for pandemics”.

While this has not been achieved yet, its potential impact could be profound. By looking at serological indicators across a population sample, such as antibodies in blood (i.e., each time you go for a routine blood test), a serological fingerprint can be determined which is then added to a national database. This can then be analyzed to test for any abnormalities among local populations. This could be used to spot abnormal symptoms, infections, and blood markers of a new outbreak prior to its spreading to other regions, reducing the likelihood of future outbreaks or pandemics.

Given that since 1900, we have experienced three pandemics, and almost a dozen pandemic threats from diseases such as SARS, MERS, swine flu, and bird flu, having an early warning system for pandemics would be helpful in assisting and streamlining public health efforts.

While this idea is still in its nascent stages, if it does become implemented in the future, samples would have to be taken in such a way that they are largely representative of local populations, and are likely to be noisy and sparse in nature. Thus, intelligent and robust sampling schemes will have to be developed that are almost certain to leverage machine learning techniques.

(5) Social Media Analysis for Mental Health

Perhaps the most tangible of all applications so far, new methods are being developed in AI4SG to monitor mental health using social media. Particularly, predicting depression from social media posts has received a lot of attention.

Tens of millions of people suffer from depression each year, but only a minor fraction of these receive treatment. Choudhury et al. (2013) used a year of data from Twitter uses who reported being diagnosed with clinical depression prior to their onset of depression in order to develop a statistical classifier capable of estimating depression risk in individuals. This included aspects such as changes in social engagement, language, linguistic styles, ego network, and tweet topics. They found that social media provides useful signals for characterizing the onset of depression.

Such a tool could be used in the future by healthcare agencies to step in and offer help to proactively prevent the development of fully-fledged depression from occurring. Researchers have suggested that the onset or existence of other psychological disorders could be flagged using similar techniques in the near future, especially as the importance of mental health is gaining increased recognition among the general public.

Similarly, postpartum changes in emotion and behavior have been assessed using Twitter posts. Choudhury et al. (2013) found that 71% of the time they were able to predict how a mother’s emotional and behavioral state would be altered following the birth of their child based on their prior social media interactions. Such information could be helpful in determining mothers that are at a high risk of experiencing postpartum depression.

Beyond depression, Thorstad and Wolff (2019) analyzed information from Reddit and were able to distinguish between different forms of mental illness such as ADHD, bipolar disorder, and depression, from words used on clinical subreddits. Interestingly, the authors analyzed non-clinical subreddits, such as those involving cooking and traveling, and found they were able to predict using posts on these subreddits if someone was likely to post to a clinical subreddit.

These results indicate that our overall mental state, and thus any mental illnesses, can influence the words we use online, and will likely become decipherable by AI4SG models in the near future.

(6) Google Project Euphonia

Google’s Project Euphonia should perhaps be the shining example of AI to enhance inclusivity. The project is an ongoing effort consisting of obtaining large amounts of raw data on dysarthric speech. That is, speech from individuals with a severe speech impediment that makes them difficult to communicate with, that may stem from illnesses such as Amyotrophic lateral sclerosis (ALS).

From this data, Google has created a machine learning algorithm able to translate this dysarthric speech into typical speech, although still in its early stages, this effort has already significantly reduced barriers for many of these individuals in communicating during their day-to-day life. In a world becoming increasingly dominated with remote work and online meetings, incorporating add-ins such as Project Euphonia into platforms like Zoom and Skype could radically improve inclusivity for individuals that live with severe speech impediments.

To learn more about Project Euphonia or to see it in action, I have provided some links below.

Project Euphonia

Other Projects

Although there are a broad number of other topics that could be discussed, I thought the above 6 topics covered exciting and non-traditional applications of AI that can be used to improve individual, societal, and environmental wellbeing, counter to the traditional AI narrative. I have left links to additional AI4SG applications below for the interested reader:

• Combining satellite imagery and machine learning to predict poverty
• Prediction of diabetic retinopathy (~415 million. cases) using deep learning on fundus images (can also predict sex, age, blood pressure, etc.)
• On Identifying Hashtags in Disaster Twitter Data
• Weak Supervision for Fake News Detection via Reinforcement Learning
• Protecting Geolocation Privacy of Photo Collections
• Detecting and Tracking Communal Bird Roosts in Weather Radar Data
• The Stanford Acuity Test: A Precise Vision Test Using Bayesian Techniques and a Discovery in Human Visual Response
• Linguistic Fingerprints of Internet Censorship: The Case of Sina Weibo

Potential Issues

As with any discipline, especially ones that tout to improve society, there may be unintended consequences in their application. This is partly why the concept of a broader impact statement has been suggested, such that the pros and cons of a specific technique are laid out clearly by the inventors. However, there is still the possibility for AI4SG to be used under the guise of improving social impacts, while having negative impacts on other communities. Fortunately, one of the SDG’s goals is reducing inequalities, making it harder for a bad actor to justify such an approach from an ethical perspective.

In an attempt to be fully transparent, I have also included links to articles laying out alternative views on this field, outlining how AI4SG might actually be detrimental to society. If time permits it, I encourage the reader to read through these and come up with their own judgement about AI4SG and how it should be used going forward. This is still a relatively new field, so now is a good time to become part of the conversation.

• AI algorithms intended to root out welfare fraud often end up punishing the poor instead
• AI for good is often bad

Final Words

Data gaps undermine our ability to target resources, develop policies and track accountability. Without good data, we’re flying blind. If you can’t see it, you can’t solve it. — Kofi Annan, former Secretary-General of the United Nations

AI4SG is somewhat counter to a computer science field, being closer to an engineering discipline than a computational one. This is a trend that we are increasingly starting to see emerge, as with the field of tiny machine learning. In some ways, AI seems to be a proto-engineering discipline, similar to chemistry and electromagnetism, which gradually evolved into chemical and electrical engineering. New subfields will likely continue to emerge from applications such as Neuralink, as well as research work like optimized democracy.

These are exciting (and concerning) times we live in, and it is about time we started embracing and harnessing the power of AI to improve the lives of everyone. However, we should do so mindfully, placing constraints similar to those encountered in medicine and social sciences to help regulate and provide accountability to the field of computer science and its budding offspring.

Given that most of the AI talent is isolated in industry or academia, it might be fruitful to encourage researchers to donate a percentage of their time to AI4SG initiatives, much like pro bono work among the legal community. The complexity of real-world challenges can in fact help to boost the understanding of existing methods and demonstrate impact where it matters the most.

Further Reading & References

[1] Floridi, L.; Cowls, J.; King, T.C., and Taddeo, M. How to Design AI for Social Good: Seven Essential Factors. Sci Eng Ethics. 2020 Jun;26(3):1771–1796. doi: 10.1007/s11948–020–00213–5. Epub 2020 Apr 3. PMID: 32246245; PMCID: PMC7286860.

[2] Tomašev, N.; Cornebise, J.; Hutter, F. et al. AI for social good: unlocking the opportunity for positive impact. Nat Commun 11, 2468 (2020). https://doi.org/10.1038/s41467-020-15871-z

[3] Chessell, M. IBM’s Ethics of Big Data and Analytics report. 2014.

[4] Annan, K. Data can help to end malnutrition across Africa. Nature 555, 7 (2018). https://doi.org/10.1038/d41586-018-02386-3

[5] Jean, N. et al.Combining satellite imagery and machine learning to predict poverty. Science, 353, 6301 (2016). https://doi.org/10.1126/science.aaf7894

[6] De Choudhury, M., Gamon, M., Counts, S., and Horvitz, E. (2013). Predicting Depression via Social Media. Proceedings of the International AAAI Conference on Web and Social Media, 7(1). Retrieved from https://ojs.aaai.org/index.php/ICWSM/article/view/14432

[7] Gibney, E. The battle for ethical AI at the world’s biggest machine-learning conference. Nature 577, 609 (2020).https://doi.org/10.1038/d41586-020-00160-y

[8] Gilman, M. AI algorithms intended to root out welfare fraud often end up punishing the poor instead. The Conversation (2020).

[9] Latonero, M. Opinion: AI For Good Is Often Bad. Wired (2019).

This article was originally published on Towards Data Science and re-published to TOPBOTS with permission from the author.

Enjoy this article? Sign up for more AI updates.

We’ll let you know when we release more technical education.