Algorithms for Good

There’s a common misconception that technology is not helpful in the way it influences public opinion. One reason for this ubiquitous belief stems from the role that big social media platforms play: they continue to use algorithms to generate targeted ads, and promulgate bubbles to maintain user engagement. But technology is not nefarious, per se—rather, the problematic reputation of algorithms stems from the aims pursued by the creators of such platforms.

Platforms can indeed be designed for other purposes, such as facilitating constructive deliberation undertaken with mutual respect. Deliberation fosters an understanding of—more often than not—the complex nature of a given issue, and contributes to informed decision-making. But all too often, this opportunity to reach a decision is missed, and deliberation remains inconsequential. Deliberation and decision-making must be considered as two sides of the same coin.

In 2009, we formed a multi-disciplinary team (referred to as “we” moving forward) to create LiquidFeedback [LF], an open-source code that powers specialized internet platforms for democratic self-organization. LiquidFeedback integrates both deliberation and decision-making. It provides a permanent infrastructure that allows any given large-scale group to simultaneously and asynchronously discuss and decide on a great number of issues. Organizations, political parties, and other actors can make democratic decisions independent of physical assemblies with a straightforward and transparent process for initiatives, competing alternatives, and suggestions. Results can be used for information, recommendations, or binding decisions, depending on organizational needs or national legislation. LiquidFeedback is platform-independent: participants can access it using a web browser on desktop or mobile, regardless of their geographical location.

Based on the learnings from the LiquidFeedback project, this article will shed light on the importance of the purpose of applied algorithms and process design, which go hand-in-glove. Indeed, technology and algorithms aren’t doing anything, per se—the question is what the target functions of the algorithms are, what process design is put forward in combination with them, and to what ends they are used. It is instrumental to carefully select appropriate algorithms that pursue the right aims. Doing so creates a fair deliberation and decision-making process that works as a neutral platform between conflicting interests.

The article begins by outlining several of the goals that we sought to accomplish in order to create a fair deliberation and decision-making process. Then, the article’s focus turns to the area of process-design, where it describes the contours of LiquidFeedback and moves to one of its core concepts, known as liquid democracy. Algorithms then follow suit: the article pivots to exploring a few algorithms in the context of both deliberation and voting. Finally, the article concludes with insights about ways in which technology can be beneficial for public debate and have a positive impact on society at large.

Goals: Creating a Fair Deliberation and Decision-Making Process

If it is a question of the aims of process design and algorithms, then the first necessary step is to define the goals that we wished to achieve. In the case of LiquidFeedback, these included the following:

• treating every participant equally
• providing means to uncover and unmask political populism
• protecting against the dominance of noisy minorities
• curbing hate speech and so-called trolling
• ensuring that minorities are given the opportunity to adequately express their position
• avoiding incentives for tactical consideration.

It should be noted that minority in the context of preference aggregation refers to any set of participants holding a common position in regard to a given issue and representing less than 50% of the population. Minorities in this sense reflect the attitude towards an issue and may or may not correlate with properties such as ethnic identity. The term minority is not used as a synonym for marginalized groups. In fact, noisy minorities often consist of somewhat privileged and vocal participants. We thank Chelsea A. Moran, David D. Kim, and Natalia Garcia Tang at UCLA for alerting us to the fact that this could be a source of misconception.

After careful consideration of these goals, we looked at scientific findings from various disciplines. Based on our research, we developed a set of corresponding algorithms, and synthesized them into an integrated deliberation and decision-making process. Ultimately, by emphasizing the overall process design in combination with the use of sophisticated algorithms for preference aggregation, LiquidFeedback accomplished these goals.

Even in a fair decision making process, however, there are always winners and losers. The latter are more likely to accept the result if they can understand the process, and consider it credible. For that reason, we have always been aware of the importance of a transparent, deterministic decision-making process, as opposed to a personalized forum with algorithms hidden from public view. While this article will provide an overview of the process design and applied algorithms, a full disclosure is available in our book, The Principles of LiquidFeedback. [PLF]

Process Design: From Admission to Voting

For the overall process design, we had to take into consideration that the ability to determine the voting options is equally as important as the ability to vote. Therefore, we designed the process in a specific way, consisting of four phases:

• During the admission phase, LiquidFeedback measures if there is significant interest in discussing an issue, which consists of one or more initiatives.
• The discussion phase serves for the improvement of initiatives, allowing every participant to bring up alternative initiatives which can then become voting options.
• The verification phase protects against political hacks at the eleventh hour.
• Finally, the voting phase determines the winning initiative.

Admission (e.g. 4 weeks), 1st quorum (issue), Discussion (e.g. 8 weeks), Verification (e.g. 4 weeks), 2nd quorum (each initiative), Voting (e.g. 4 weeks).

Figure 1: The four phases of an issue in LiquidFeedback

In conjunction with these four phases, LiquidFeedback relies on transparent rules depending on the type of decision being made, for example, when deciding on political positions, organizational decisions, budget decisions, or changes to the organizational statutes. These rules of procedure are defined in advance by the organization using LiquidFeedback: they dictate the duration of each phase, any quorums, and the majority to be achieved.

In total, the structured deliberation generates the set of viable voting options, which are then voted upon in the voting phase. The overall process design permitted us to accomplish the initial goals, providing the framework by which the algorithms then operate. Indeed, a clear process design supports fairness and fosters a purpose-driven deliberation process.

Division of Labor: Liquid Democracy

When developing LiquidFeedback, we had full democratic self-organization for large-scale groups in mind. While it is a democratic ideal to allow everyone to have a say in every decision, it comes with a burden to the participants, many of whom can’t possibly be knowledgeable on all topics. Some critical questions arise here: Does everyone want to deal with every question? What if people are interested in different areas? It’s clear that the selection of topics in which people want to have a direct say, or be represented, may differ.

Thankfully, there was already an idea circulating at the time that provided a dynamic solution to these questions: the idea of liquid democracy. [LD] Its defining elements are:

• the ability to cast a direct vote
• the ability to delegate a vote to another participant
• topic-specific delegations
• transitivity of delegations
• instant alteration and recall of delegations.

Basically, one participates in what one is interested in, but for other areas, gives their vote to somebody acting in their interest. The notion of liquid democracy was not only convincing and practical—it was even the initial impetus to consider the development of LiquidFeedback.

LiquidFeedback implements liquid democracy in the most straightforward way: Participants can opt for representation by setting defaults, override them by topic-specific delegations, and even delegate for particular decisions. They can also opt against representation. In either case, changes of these preferences and direct participation are always possible. In order to directly participate, it is not even necessary to explicitly recall a delegation, given that the direct activity automatically suspends the delegation for a given issue.

In LiquidFeedback, liquid democracy is not only used for voting, but also during the structured deliberation applied as a debate empowerment. A dynamic scheme of representation unfolds, which facilitates a dynamic division of labor based on individual choice.

What is more, liquid democracy supports the self-organization of all factions and sub-groups, be they defined by gender, ethnic identity, or even by values. It also challenges the iron law of oligarchy, a tendency for amassing power by a few individuals. In the end, without putting too much of a burden on participants, liquid democracy allows everyone to directly participate whenever they see fit. As such, liquid democracy helped us accomplish our initial goal of equal treatment of all participants without rendering the process dysfunctional.

Algorithms during Deliberation

Process design isn’t everything. To combat the dominance of noisy minorities—which can cause great harm by creating a biased impression of majority ownership and drowning out views of other minorities—we developed aggregation algorithms [Evolution] to ensure a fair division of the screen area by ordering content based on current support.

Depending on the context, LiquidFeedback applies different sorting algorithms: Issues—except in the admission phase—are sorted by the remaining time in the current phase, i.e. the most urgent first. Issues still in admission are sorted by Proportional Runoff, and are placed behind all other issues. Competing initiatives in an issue are sorted by Harmonic Weighting, which utilizes the same counting scheme as Thiele’s elimination method. [Janson] Based on existing assessments from participants, suggestions are sorted by Proportional Runoff.

Given that knowledge of the algorithms cannot be misused to draw attention from another group, there is no need to conceal them. Likewise, it is not necessary to know the algorithms in order to benefit from them.

In sum, the algorithms assign an appropriate amount of attention to each minority, regardless of the intensity of the agitation. These algorithms address multiple of our initial goals: they protect against the dominance of noisy minorities, therefore curbing hate speech and trolling done by individuals or small groups. Ultimately, these algorithms also ensure that all minorities are given the opportunity to adequately express their position.

Subject area: Green Space. Decision #5. Phase: Discussion (29 days 23:52:15 left). i11: Paint park benches on demand (green colored) by Daisuke Tanaka, 324 supporters. i10: Park benches should be uniformly colored red by Martha Westerberg, 167 supporters. i12: Provide vandalism-resistant plastic benches by Lars Andersen, 284 supporters.

Figure 2: An issue in LiquidFeedback

Algorithms during Voting

Many different options can appear while deliberating on an issue. Some of them may be very similar, to the point that they are deemed “clones.” [Tideman] After consulting the comprehensive literature on voting algorithms, we decided to implement a clone-resistant algorithm for preferential voting. To determine the winner from a set of alternatives, LiquidFeedback implements Clone-Proof Schwartz Sequential Dropping, known as the Schulze Method. [Schulze]

In the Schulze Method, the winner is determined by a pairwise comparison of each voting option with every other voting option. The voting option preferred by a majority over every other voting option is the winner. If this criterion is not met, the winner is determined through further algorithmic mechanisms.

It’s worth noting that in some cases, there will be no alternative initiatives upon which to vote. In these cases, the voting becomes a binary decision (yes/no) in LiquidFeedback.

Through the Schulze Method, participants can express their real preference without needing to evaluate which clone has the best prospects to win. In other words, the existence of the same basic idea in different alternatives neither helps nor harms the very same idea. By using a clone-resistant voting algorithm, we were able to handle multiple voting options while avoiding the necessity of tactical voting. Once again, the right algorithmic choice was crucial, in this case avoiding encouragement for tactical voting.

i11: Paint park benches on demand (green colored) by Daisuke Tanaka, Reached > 50/100: 121 Yes (75%), 39 No (25), 0 Abstention (0%). Competing initiatives in pairwise comparison to winner: i10: Park benches should be uniformly colored red by Martha Westerberg [red bar greater than green bar]. i12: Provide vandalism-resistant plastic benches by Lars Andersen [red bar greater than green bar].

Figure 3: Voting result of an issue in LiquidFeedback

Conclusion

Although online debates oftentimes reinforce conflicts and political fault lines, they can also support reconciliation of interests and contribute to social cohesion. It all comes down to goal-driven process design and algorithms.

While this article provides an overview on how the LiquidFeedback project approaches the challenge of democratic self-organization, it can not cover every detail. A full description of the core concepts can be found in our book, The Principles of LiquidFeedback. [PLF] Over the course of the years, after carefully assessing their value and weighing their conformity with the postulated goals, we have added many optional extensions for LiquidFeedback, such as the issue limiter [IssueLimiter] and a fair distance function [FairDistance].

When it comes to process design and the choice of algorithms for particular purposes, there is no room for pure trial and error. Implementations need to be based on sound assumptions and have to take scientific findings into consideration. Ultimately, implementations should be further improved based on empirical studies.

If used in a decent and sensible way, the internet and technology, in general, create new opportunities for public debates undertaken with mutual respect. Just like the Parisian salons in the 19th century, but on a much larger scale, online platforms can expose participants to ideologically cross-cutting content, challenge views, facilitate informed decision-making, and be instrumental in the construction of a vision for the common good.