Augmented Democracy

exploring the design space of collective decisions


Augmented democracy

In this TED talk, César Hidalgo, Director of MIT’s Collective Learning group, presents the idea of augmented democracy. Below, you can find answers to frequently asked questions (f.a.q.). Here you can fin the talk in both English and Spanish.

This is a live document that will continue to be updated.

All the contents in this website are open source with a “by-” license with free commercial and noncommercial use allowed.

Homepage photograph by John Werner.
Art by Maria Venegas (download images 1,2,3,4,5)

$10,000 PRIZE

interested in designing an augmented democracy system?

Submit a proposal to the Augmented Democracy prize for a chance to win cash prizes (SEE DETAILS).


What is Augmented Democracy?

Augmented Democracy (AD) is the idea of using digital twins to expand the ability of people to participate directly in a large volume of democratic decisions. A digital twin, software agent, or avatar is loosely defined as a personalized virtual representation of a human. It can be used to augment the ability of a person to make decisions by either providing information to support a decision or making decisions on behalf of that person. Many of us interact with simple versions of digital twins every day. For instance, movie and music sites, such as Netflix, Hulu, Pandora, or Spotify, have virtual representations of their users that they use to choose the next song they will listen to or watch the movies they are recommended. The idea of Augmented Democracy is the idea of empowering citizens with the ability to create personalized AI representatives to augment their ability to participate directly in many democratic decisions. 

Are you proposing to replace actual democracies by Augmented Democracy?

 No. As I said in my TED talk, and I confirm here, my proposal is not to replace our current democracy with Augmented Democracy. My goal is to ignite a debate focused on expanding the design space of democracy. Democracy is a beautiful idea, but there are many ways in which it can be implemented. The number of possible implementations of democracy has increased with the introduction of new technologies, but we have not been diligent at exploring many of these alternatives. My proposal is to create unofficial experimental systems of Augmented Democracy that we can use to learn about the limits and potentials of this technology. 

How would one of these experimental “Augmented Democracy systems” work?

What an Augmented Democracy system would need to do is predict how each of its users would vote on a bill that is being discussed in that country’s congress or parliament. These predictions would provide an estimate of the support that the bill would have received if it had voted directly by the population of users of an AD system instead of their elected representatives. 

To provide these predictions, an AD system would need information from both users and bills. People participating in the system would provide information on a voluntary basis and have the ability to withdraw the information at any moment. This information could include active and passive forms of data. Active data includes surveys, questionnaires, and, more importantly, the feedback that people would provide directly to its digital twin (for instance, by correcting their twin when it made the wrong prediction). Passive data would include data that users already generate for other purposes, such as their online reading habits (e.g., New York Times vs Wall Street Journal), purchasing patterns, or social media behavior. 

The digital twin would then combine a user’s data with information about a bill to predict how that user would vote on that bill. The algorithm used to make a prediction would be chosen by the user from an open marketplace of algorithms, and the user would be able to change it at any time. These algorithms could also make predictions on more nuanced issues, such as which specific parts of the bill the user is more inclined to agree or disagree with, or what pieces of data about a user prompted the AI to suggest a decision. 

Experimental AD systems would be, by definition, unofficial, meaning that the decisions made by the agent would not be enforceable and have no bearing on official policy. The decisions would also be private and only visible to each individual. Yet, an AD system could also provide aggregate information about the level of support predicted for a bill, the geographic distribution of the support, and the estimated popularity, accuracy, and bias of the algorithms in the marketplace.

Would such a system centralize too much data and power on the system’s architects? 

In recent years we have all become well aware of important data privacy and monopoly concerns. These concerns involve primarily large platforms, such as the social media giants of today. These large platforms centralize large volumes of data, but they are not the only way to design an online system. 

My recommendation for an AD system is not to design it based on the platform paradigm that dominates today’s web but instead use the protocol paradigm that dominated the early days of the web. Platforms, such as Facebook, are almost natural monopolies, whereas protocols, such as email, allow the creation of distributed systems (more similar to markets than monopolies). 

In a protocol-based AD system each user can store their data in their own “personal data store,” or “data pod” (like the pods proposed by Tim Berners Lee in his Solid Project). The data can be in any of the many cloud providers of such a service or in a home computer. This is similar to email. Unlike social media, email services are provided by a large number of universities, companies, and other organizations. In a platform world (e.g., Facebook), some people have control over the entire platform. In a protocol system, like email, nobody has access to all of the email servers. It is a deeply fragmented and federated system by default.

The algorithms used to make predictions are also distributed and not unique. They exist as part of a marketplace that is open for people wanting to contribute algorithms. Users can select which algorithms they allow to interact with their data, based on how accurate they think the algorithm predictions are, how much they trust the algorithm’s creators, and other criteria.

These decentralized architectures can help revert the data concentration problem, by bringing questions to the data, instead of centralizing all the data in a few places. These questions are answered in a decentralized manner. And yes, this could be an application of blockchain technology (although that may also not be the only alternative).

There are, of course, advantages and disadvantages of using protocols instead of platforms. The big advantage of protocols is their distributed nature. This allows protocol-based systems to avoid centralization of data and mitigates many of the privacy and monopoly concerns that are natural in platforms. The disadvantage is that, because of their decentralized nature, protocols are much more difficult to update and improve than platforms. 

How can you train algorithms if the data is not centralized in a platform?

While in principle it is easier to train a machine learning algorithm by having all of the data in one platform, in practice it is becoming possible to send algorithms to distributed data to train them without having to pull together the data first. This is the idea of Federated Learning. You can learn more about this idea in this recent paper.


How different is Augmented Democracy from our current democracy?

As a matter of fact, Augmented Democracy can be identical to our current democracy, or quite different, if we so choose. 

Consider an implementation of Augmented Democracy where everyone has the same exact algorithm for his or her digital twin: an algorithm that simply copies the votes of the congressman that was elected to represent your district. In that case, we recover exactly our present-day democracy. The difference is that Augmented Democracy provides a “dial” that we can turn to explore a larger design space, from a world in which only one out of a million decisions gets passed on to the AD system to a world in which most of them are.

What do you mean by a design space?

By a design space I mean the many possible ways in which one can execute an idea. It is the multiple “hows” satisfying the same “what.” 

Consider the idea of a chair. The idea of a chair (the “what”) can be implemented in multiple ways (the “hows”). Chairs can stand on four legs or on two arcs. They can be made of metal, wood, or even cardboard. They can have arms or not. They vary on the way they provide lumbar support. They can be rigid, or they can swivel and recline. In this analogy, democracy is the idea of a chair (the “what”). There are many possible ways in which we can think about implementing the right of populations to self-determine. Some of these different ways are very similar to each other and would be neighbors in this design space. Others are very different and would be far apart in the design space. But just like in the case of chairs, the design space expands as we develop new technologies. 

Why do we need an AI digital twin? Couldn’t people just vote directly without using AI?

Direct democracy based on communication platforms is not a new idea. It is actually an idea that has been tried repeatedly. Yet, few people know about these initiatives because they have failed to attract and sustain substantial levels of participation. 

 A beautiful and positive example of direct democracy is the “Decide Madrid” project, which empowers citizens of Madrid with the ability to vote and decide on dozens of millions of Euros of public investment (EUR 30 million in its last round). This is a big incentive; yet, the biggest issue that people in projects like Decide Madrid face is the lack of citizen participation. In Madrid, a city of millions, only a few proposals ever receive more than a few dozen votes. In Chile, the “Senador Virtual” project has been running for more than a decade. It presents citizens with simple questions about some of the bills discussed in the Chilean senate, together with the text of the bill. Despite its long trajectory, bills in Senador Virtual receive only a few thousand votes.

 As I explain in my TED talk, the level of participation in democracy is relatively low, even though the number of participatory instances is relatively small. If we were to expand democracy to more instances of participation (like participatory budgets or having direct democracy for parliamentary decisions), the empirical data suggests that the participation of people would be minimal and decrease with additional instances. So the technical problem of direct democracy is not one of limited communication but one of the limited time and cognitive bandwidth of people. To participate in hundreds of decisions, we don’t need additional communication technologies but technologies that augment the number of different things a person can pay attention to. 

Who would introduce new laws in an Augmented Democracy system?

In the beginning, the digital twins of an AD system could simply vote on the bills being introduced by professional policy makers (senator or congressmen). Yet, the technology opens the opportunity to have people introduce and propose bills directly, or to even have algorithms discover potential new regulations. To start, however, it makes sense to focus on digital twins and citizens voting, rather than proposing bills.

What is the purpose of these experiments? 

The goal of these experiments is to expand our discussion on democracy, bringing it from the “who” to the “how.” Recent changes in technology have expanded the design space of democracy, but as a society, we have not been good at exploring this design space. Democracy should not only be about “who” runs the current system of collective decision-making but also about “how” we design systems for collective decision-making that support the right of populations to self-determine their future.

Are you concerned about the ethical implications of AI?

 Of course, I think AI ethics is an important topic, primarily because we know little about it. In recent years many opinion pieces, editorials, and essays have been written about AI ethics. But there is still relatively little data on the topic, except for a few noteworthy examples in the case of self-driving vehicles or computer vision bias. At the moment, my team and I are running dozens of experiments on AI ethics that we will release on a book later this year. For those interested in getting started in the topic, here is a short list of interesting papers to get you started: (1,2,3,4,5,6,7,8,9,10,11)  


How did the idea of Augmented Democracy come about?

 There are two paths that brought me to the idea of Augmented Democracy. The long story is about how I got to think about the use of AI in government. The short story is that of the “Aha!” moment.

 The long story starts nearly a decade ago. For many years, I’ve been creating platforms to integrate, distribute, and visualize public data. You may know some of these platforms, such as DataUSADataChileDataAfricaDataViva (Brazil), or the OEC, among others. The reason why I have been building these platforms is to support decision-making by public workers and support the use of public data by citizens. Over time, this work led me to imagine a future in which executive decisions, such as public investment and spending decisions, would be empowered by data and supported by algorithms. In fact, in Datawheel, a company I founded a few years ago, we are already working on projects that use public data to help identify gaps in public investment and can support such decisions (such as: which is the place in most need of a public hospital?). So the idea of Augmented Democracy grew naturally for me as I worked on the creation of multiple data integration, visualization, and analysis systems to support public decision-making.

 The second story is that of the “Aha!” moment. In January of 2018 I was in Santiago de Chile. I had just launched DataChile. After a hectic week of talks, meetings, and events, I met with a couple of friends and the editor of a newspaper to talk about potential projects. The day was January 11. They were concerned about the future of democracy and were wondering if I had any creative ideas. In the conversation, I proposed the idea of Augmented Democracy. 

 At that time, I had already confirmed that I was speaking at TED, and I met with my TED curator a few days later. He became very enthusiastic about the idea and began pushing me to focus my TED talk on it.

Was it difficult to share this idea?

 Extremely. While I think the idea of Augmented Democracy is important, and is a great topic of conversation, I also know very well that some people have a knee-jerk reflex when hearing AI and democracy in the same sentence. I know I am entering a very controversial territory. I’ve socialized the idea privately many times, and I know that some people are quick to respond with pushback, hostility, or aggression. In fact, I was so concerned about the controversy of the talk that I even tried to return my TED talk to my original storyline, based on my past projects (e.g., DataUSA). Yet, I was not allowed to do so. I also delayed the release of my talk to give myself more time to think about it and finish a number of other academic projects.

How do you think people will react to this idea?

I’ve shared this idea a few times, and I’ve seen how people react to it. The general reaction is to divide the room in half. The idea is controversial, but interestingly enough, it does not seem to divide the room in half based on traditional party lines but instead based on a somewhat orthogonal dimension. In fact, when I did my TED talk last year, TED tweeted a poll with a summarized version of the idea, and the poll was 47% in favor and 53% against. Many people argue that a 50-50 split is the sweet spot for a new idea since an idea with 100% support, or 100% rejection, is not an idea that leads to much discussion. 

What is also interesting about this TED twitter poll is that many people interpreted the tweet as a system that would vote on your behalf for a politician (which is very far from the idea presented here), and still the poll got to almost 50-50. Of course, TED followers are biased towards people that are more technology oriented.


Do you think collective decisions are always a good idea?

No. Collective decisions have many follies and foibles. The power of collective decision making ranges from the wisdom of the crowds, first described by Francis Galton, to herding effects that can generate mob mentalities and lead to nefarious outcomes. The fact that I am pushing a conversation on the design space of democracy is not equivalent to me believing that decisions made by many people are always better than decisions made by experts.

Isn’t democracy also about debate and deliberation?

Yes, and the hope for an AD system is to help promote a positive debate. Imagine a university, newspaper, or non-profit that decides to implement an experimental AD system. The outcomes of an AD system could be used as the starting point of conversations. How would a “congress of avatars” vote on issues such as gun law, immigration reform, universal marriage, wealth tax? How would digital twins vote on proposed bills that have not yet made it to the floor of the house or senate?

The idea of Augmented Democracy is never to offload the democratic process to a senate of digital twins but to empower people’s ability to participate by technologically expanding their cognitive bandwidth. By doing so, we may generate more instances of informed debate, not less. Hopefully, having a better user interface for democracy, and providing people with the ability to participate, will encourage people to become more aware of the regulations being discussed.

Isn’t democracy about compromises?

Many people see compromises as an important part of democracy. To approve bills, parties negotiate different clauses, rewriting bills until they reach an acceptable middle ground. Will an AD system eliminate these negotiations? Not necessarily. Imagine a bill being considered in an AD system that is supported by 30% of the people, rejected by 25% of them, and abstained by 45%. In such a case, the bill would need to be rewritten and iterated until it receives enough support. 

Do you think people are prepared to make complex political decisions?

For the most part, no. I think many of us lack the knowledge needed to make informed decisions in many complex issues. This also includes many publicly elected policymakers. Some legislations can be complex and require expert knowledge. Other legislations must reflect personal values. Making policy decisions is not easy, yet I don’t think we learn without taking responsibility or trying. 

Experimental AD systems provide an opportunity to “play” before making any real changes. They provide a safe, unofficial opportunity to experiment and learn about the potential of a world in which everyone has direct representation.

Have you ever worked in government?

While I have never held an official government position, I have worked closely with the governments of many countries. In my work, I’ve had the pleasure to meet with several presidents and ministers, and also, I am a frequent speaker at events that are attended by many public officials. Because of my work on economic development, industrial diversification, and economic complexity, and because of my work on integration, distribution, and visualization of public data, I am no stranger to working with governments. 


Do you not know that algorithms can have biases?

Of course, and this is something that is very well known by anyone that has worked in data science and artificial intelligence. In fact, algorithmic bias is not only an important topic of discussion but also an active area of research. On the one hand, some people are working on efforts to audit algorithms and point to their biases. On the other hand, there are people working on ways to reduce the biases of algorithms, by either expanding or modifying the data that the algorithms are trained on or modifying how the algorithms are trained.

What are your views on AI for the next decades?

Researchers in AI like to distinguish between weak AI and strong AI. Weak AI is AI capable of executing narrow tasks, like those executed by manufacturing robots, autonomous vehicles, or the AIs that have dominated games like Go, Jeopardy, or chess. Strong AI, or Artificial General Intelligence (AGI), is an AI that can be simultaneously intelligent across a broad range of domains. At the moment, I think the learning curve we are riding is that of weak AIs, with strong AIs far in the horizon. The ethical and societal implications that we will face in the next decades relate to weak AIs, designed or evolved to solve narrowly defined tasks. In the world of weak AI, AI technologies are mostly assistive and are subordinate to humans.

Do you think our technology is ready for creating the digital twins you describe?

At home I have three “personal robots” right next to each other: an Alexa, a Jibo, and a Nespresso machine. Guess which one is the one we use the most. In my view modern software agents are extremely limited. Alexa cannot even understand song requests from me in my native Spanish. Jibo is a cute but expensive clock-radio that sometimes interrupts us for no reason. I also find the experience of Apple’s Siri disappointing and frustrating. Yet, the Nespresso machine gets the job done every morning (even though it is a very narrow task). So, honestly, I think we are far from software agents capable of creating good representations of us and mapping those representations to the goals and intent of policies. At the moment, our technology is not yet ready for creating competent digital twins. 

If both people and the technology are not ready for these systems, why experiment with them and discuss them? 

Because of the power of play. The key aspect of this proposal is that I am strongly advocating for experimental systems that are unofficial and non-binding. These “toy” systems can help us learn a variety of things about, both the way in which we think about governing the commons and about technology.

While our technology is not yet capable of building sophisticated digital twins, the technologies needed to do so are within the realm of weak AI. Natural Language Procession (NLP) techniques can create representation of texts and are getting better at that. Similar techniques can be used to create representations of people preferences, who we can later match with the representation of legal documents. Of course, legal texts can be complicated. So to create these systems, we will also need to think about how to improve and structure the digital representations of legal texts.

The fact that these systems are unofficial, and non-binding, also provides the opportunity for people to play with other design paradigms. How would you design the user interface (UI) of democracy if you could start from scratch? How would you structure laws in the same way that we do now? Or would you simplify the interface by which people communicate their preferences regarding universal social dilemmas, like those of immigration, abortion, universal marriage, controlled substances, or guns? How would you keep people engaged in the use of the system? Would you give communicated preferences an expiration date? Would you make strong design distinctions in the UI to separate among important policy decisions (e.g., gun laws, immigration, substance control) and more mundane design decisions (renaming of a local federal library)? How often would you ping users for input? Would you open these “toy” systems to the world? Or would you restrict them to local communities?

 By focusing on play we can also learn more about when collective decisions work or when they don’t. For instance, the non-binding outcomes of experimental systems will help promote discussions in which people evaluate and compare the AD systems outcome with that of a binding congress or parliament. 

Play is an important path to learning in both new technologies and social problems. The reason why you experiment before you build is because you will never be ready if you expect to be ready on your first try. There are many lessons we still need to learn, and for that, play is key.

Is the world ready for this idea?

I don’t think so. I think the world is ready for an idea when that idea becomes obvious to many people. For most people, the idea of powering direct democracy by augmenting people’s cognitive bandwidth through personalized and decentralized digital twins is both a new idea and an unexpected idea. To me, the idea only became obvious after working on the creation of systems to integrate and distribute public data for nearly a decade. So, I think the world is not yet ready for this idea. Yet, I believe this idea, or some variation of it, may be obvious to young adults in 30 to 40 years.

Are there other people thinking about similar ideas?

While the idea of using digital twins to augment democratic participation is new for most people, there are other people thinking about similar ideas. In New Zealand, a virtual politician called Sam was designed to learn about the preferences of New Zealanders. Its conversational abilities clearly need improvement, but this is also a related idea (although Sam is not designed to represent individuals in a personalized manner, but as a collective). In a recent article on the Journal of Ethics, a couple of researchers also propose creating and training moral bots to represent the ethic preferences of individuals.

Is the idea of digital twins limited only to democracies, or does it have private sector applications?

Digital twins will probably go mainstream in private sector applications before they become part of our democracies. Imagine the natural evolution of the retail sector. In only 20 years much retail activity moved from physical stores to online shopping. These online shopping systems are powered by recommendation algorithms, which are personalized and drive an important fraction of purchasing decisions. Retail is now moving closer to the home through services that deliver items and groceries directly to our homes, or even our refrigerators. It is not hard to imagine a future in which your weekly grocery decisions are made and executed by digital twins that know how much butter you eat and when you are expected to run out of it. 

Aren’t you missing a few things? 

Of course! There is more to the idea of Augmented Democracy than what I can explain in a 10-minute TED talk and a ~4,000-word FAQ. In fact, I enjoy having conversations about this topic with other people, and learning about their views, but just like everyone else, I have a limited cognitive bandwidth. To help us imagine what an Augmented Democracy system would look like, I have opened a contest with USD 10,000 in prices. The contest will give two prizes of USD 4,000 to the best two team proposals and two prizes of USD 1,000 to the best two individual proposals. The design proposals would be judged by a panel of experts and will be shared online in this site. I encourage proposals that are both: imaginative and technically sound. The proponents would not need to implement the system, but they would have to design and describe how they would attempt an implementation. The proposals should include simulated screen shots. I would also encourage submitting a video describing how their version of the AD system would work. Any material from my TED talk or this FAQ (e.g., the idea of decentralizing the systems using protocols instead of platforms) can be freely incorporated into the proposed designs (think of these ideas as completely open source). For more details about the contest, click here


I thank the support of Gerry Garbulsky, for working with me during the months leading to TED and Chris Anderson for pushing this idea as an idea worth spreading. I thank Andrea Margit for supporting this idea with enthusiasm from the beginning, all members of the Collective Learning group at MIT for discussing this idea with me on multiple occasions, all members of Datawheel for supporting me in this and other ventures, and specially I thank the support of my wife Anna and daughter Iris. I also thank Alejandra Larrain and Tomas Gonzalez for supporting this idea from the beginning and Francisco Larrain, for a useful conversation when I was hesitant about presenting this idea at TED. All errors are my own.