The Limitations of Aligned Incentives

The hype around cryptocurrencies like Bitcoin and Ethereum has led to a surge of projects trying to put seemingly everything on a blockchain including social media. Projects like Steem, Props and Kin all use blockchain to reward the users and developers of social systems. They market themselves as realigning incentives between developers, creators, and viewers of platforms so in theory, these sites should reward quality content rather than the attention-grabbing content that is prioritized on platforms like Facebook and YouTube. And because these are protocols that nobody owns, there is no worry that some centralized company will misuse user data. Discussing projects like these the often-cited article Protocols Not Platforms (Jack Dorsey for example cited it as an inspiration for Twitter's Blue Sky initiative.) argues that social cryptocurrencies like those mentioned above may be able to both create a sustainable business model, and align incentives between participants on the network.

For many, it seems aligned incentives are the be-all and end-all of network design—including us in our original version of Realigning Online Incentives—yet as we argue here, it is not that simple. First, these subjective rewards protocols are inherently flawed due to the impossibility of subjective consensus. Without consensus, social reward mechanisms are gameable. The most problematic way to exploit these rewards may be collusion. We show that identities are often incentivized to collude and that in many cases it is impossible to detect and therefore stop. Finally, we question whether economically aligned incentives would even lead to a healthy online social ecosystem where users create better quality content.

Although we are critical of projects in our space, we wish to do so with good intentions. Because these projects are clouded in esoteric terms and technology, we worry developers, users, and investors have not noticed the flaws that we see. And while we are critical of some projects mentioned here, we would like to note that there is also a lot these projects did well to become reputable. There is a lot that we have personally learned from them and ultimately, we hope that we too will grow large enough to be criticized in blog posts discussing social media. Lastly, we admit there is much more we don't know than what we do. The projects we criticize could understand things that we do not.

The Impossibility of Subjective Consensus

Let’s first take a look at how the aforementioned subjective reward protocols attempt to reward quality content and apps on their networks. Generally, these projects distribute rewards at set intervals to the creators, developers, or both based on some metric. In the case of Steemit, the blogging platform on the Steem blockchain, Steem tokens are minted and rewarded to content creators proportional to the votes on their posts during that interval. The Steem whitepaper calls this process ‘subjective proof of work’ (pg. 12) referencing Bitcoin’s Proof of Work consensus mechanism. But unlike Bitcoin’s objective consensus mechanism, Steem’s fundamental flaw is that there is no subjective consensus.

Consensus protocols are meant to allow distributed identities to agree on some sort of truth. There are limitations of what different identities can agree upon, though. Consider three people in the physical world sitting around a table with three pencils on it. Everyone agrees on a standard definition for a pencil and that there are 3 objects on the table that satisfy that definition. There is consensus because the number of pencils is objective. Yet now Alice, Bob, and Charlie want to figure out which one is the best pencil. Alice says it’s the first because it is the biggest so it will last the longest. Bob says it is the second because it has the best shade of yellow. Charlie, who considers himself a contrarian, likes the third. They argue for hours and cannot agree on the truth. There is no consensus.

Bitcoin’s consensus works because it is purely objective. The inputs of a transaction either have the same sum as the outputs or they do not. A Bitcoin block either only contains valid transactions or it does not. And a block is either on the longest chain or it is not. Resultantly, there are clear definitions of honest and malicious validators in the Bitcoin network. Just as importantly, Bitcoin’s miners are doing objective work for the network. Miners use their computation power to increase network security. The more honest hash power validating the network, the harder it is to attack. Miners provide objective value to the network and they are rewarded proportionally to the value they provide.

The same is not true for subjective consensus. Everyone will have different definitions of what constitutes good and bad content and therefore it is impossible to differentiate between honest and malicious actions. Therefore there is also no consensus regarding what content and actions may increase the quality of social networks and which do not. Without consensus regarding what is honest behavior, there is no way to detect or stop people from creating different accounts to vote on their own posts or from paying bots vote on them. The latter is already happening on Steem and we would be shocked if the former was not, too. The Steem whitepaper even acknowledges and tries to justify this flaw saying, “Eliminating ‘abuse’ is not possible and shouldn’t be the goal. Even those who are attempting to ‘abuse’ the system are still doing work” (pg. 15). They argue they are still doing work, yet there is no way to quantify the value of the 'work.' Resultantly, Steem rewards are gameable.

Props and Kin have similar issues. They both reward apps proportionally to the number of in-app transactions. Developers can game these mechanisms by setting up bots to send tokens between wallets in their apps. More transactions make their respective tokens more valuable, but they do not signify quality applications.

All metrics regarding subjective consensus are gameable. Imagine if Alice tells a couple of pencil manufacturers that she will heavily reward whichever manufacturer can make the best pencil. As soon as the manufacturers discover that Alice likes big pencils, they will focus on maximizing that metric. They can skimp on other quality measures and be rewarded nonetheless. And there still will not be any consensus that these pencils are better. Simply stated, subjective consensus does not exist, and any attempts to measure subjective value with metrics are gameable.

The Collusion Problem

Subjective reward protocols work in theory under the assumption that all individuals are acting independently and purely in their self-interest. If everyone acts independently, then everyone will only vote on content that they like. Yet the ability to pay bots to vote on content brings up a larger problem with subjective reward protocols: the problem of collusion. There is no way to stop people from working together in distributed systems. It happens on Bitcoin in the form of mining pools so that miners can get more steady returns, yet mining pools do not undermine Bitcoin consensus. In the case of social rewards protocols, people can work together to get a disproportionate share of the rewards.

Even more common than payments as collusion will be coalitions as collusion. In most cases, coalitions are not considered collusion, but because we show they can undermine subjective reward systems, we refer to them as such. As Kevin Simler and Robin Hanson discuss in The Elephant in the Brain, all humans naturally form coalitions in their self-interest. It already happens on social media. Successful YouTubers often do collaboration videos with each other. Successful podcasts hosts often have each other on their podcasts. In many cases, successful creators Dan and Erin would never have worked together if Erin was not as popular. She could create just as awesome content and their collaboration would be just as great, but Dan may not care about working together if Erin has a lesser audience. Obviously, there are exceptions where established creators find budding talent, but those are the vast minority. Even more common than collaborations are liking someone's post because they've given your content likes in the past. We'll be the first to admit that we already do this on social media. It's natural human behavior. They scratch your back so you scratch theirs.

Not only do humans form coalitions in our self-interest, but Simler and Hanson also assert that humans subconsciously justify their selfish intentions as cooperative. Dan says that he loves Erin's content but maybe he's just justifying collaborating with Erin to expose himself to Erin's huge audience. If Erin had a lesser audience would Dan still love her content? The same goes for votes. Does Dan vote for Erin's content because he truly likes her content, or does he do so just because Erin also votes for his content? It is impossible to tell.

Now let's look at how coalitions could undermine Steem. Whales, those with a lot of Steem tokens, have the most voting power because voting power is proportional to the number of tokens voters hold. Everyone should therefore want whales to vote for their content, and whales will naturally want other whales to vote on their content so they can remain the rich. Whales should, therefore, be incentivized to trade votes with other whales. The whales will argue they like the content of other whales. Because we cannot see the inner workings of their brains, it will impossible to know if they truly do, or if they are just justifying actions in their self-interest. Steem, therefore, seems to be a plutocracy, governance by the rich. To reiterate, people subconsciously justify selfish behavior as cooperative. We believe that technology should be built to work with human nature, not against it.

Vitalik Buterin also wrote a great piece on collusion that greatly influenced our thinking on this subject.

Subjective Governance

The fundamental issues with subjective rewards protocols bring us to the larger problem of subjective governance in decentralized systems: there are no perfect solutions to subjective, distributed decision making because subjective consensus does not exist. The Zcash privacy cryptocurrency recently encountered this issue with their development fund. Initially, the protocol was designed so that 20% of all block rewards would further fund the development of the project. The Founders Reward is set to end later this year, but there is still a lot of development work needed. They must upgrade their protocol in a way that will hopefully satisfy everyone so that the blockchain does not fork. At first glance, the simple solution would be a vote on it, but decentralized voting is much easier said than done. One-identity, one-vote systems cannot work in distributed systems because there is no way to differentiate between a vote tied to a real person and a vote generated by a bot.^[1] Dishonest voters could create many identities to gain disproportionate control over the network.

More complex voting systems do not solve this issue either. Zcash founding scientist Eran Tromer explains there is an inherent trilemma between coin-weighted, permissionless, and whale-resistant voting systems. If a system is coin-weighted and permissionless—like Steem—then whales can have a disproportionate impact on the outcome, swaying the vote in their favor. You could try capping the number of funds someone can use, but then whales can split their funds among multiple wallets and gain the same voting power. You could create a coin-weighted and whale-resistant system by creating a permissioned vote where all identities are known. In that system, though, someone would inevitably be able to choose who can and cannot vote. In other words, such a voting system is centralized. It is also contrary to the ethos of a permissionless, privacy-focused blockchain.

Subjective consensus does not exist. There is no perfect solution to subjective, distributed decision making. Each has flaws that potentially allow some identities to gain disproportionate control over the vote. In Zcash's case, the singular decision of how to upgrade to the protocol is likely deemed a necessary one by the vast majority of members in the community and its outcome rather inconsequential to most, too. Most Zcash users probably will not care who further develops the protocol as long as it continues to function. But for protocols that must continuously make decisions as to what content and apps should be rewarded, it is a much larger problem.

Where Incentives are Counterproductive

Many may hear about the problems inherent to these reward protocols and think the answer must be to choose better metrics or to redesign the reward scheme. Well, that's at least exactly what we initially thought. But even if the perfect metric existed so that it is ungameable and everyone considered the reward scheme fair, would it necessarily work as intended? Would it lead to better quality content online and a healthy online ecosystem?

For many economists, it is a law of nature that rewards directly impact motivation, yet psychologists and behavioral economists believe it is more nuanced. Psychologists and behavioral economists often differentiate between intrinsic motivation, motivation for the activity in and of itself, and extrinsic motivation, doing an activity for an external reward. In many cases, extrinsic rewards like those in subjective rewards protocols may undermine intrinsic motivation for creating content and apps. If that's true, then subjective reward systems may be counterproductive, potentially leading to worse quality content.

Economists Roland Bénabou and Jean Tirole model how extrinsic rewards impact intrinsic motivation in their paper "Intrinsic and Extrinsic Motivation." Their findings generally agree with psychological research that promised rewards often undermine intrinsic motivation. Bénabou and Tirole argue that promised rewards signal a task is unattractive. The larger the promised reward, the more unattractive the task must be.

Psychologists John Condry and James Chambers provide a different explanation for why extrinsic motivation can counteract intrinsic motivation in "Intrinsic Motivation and the Process of Learning."^[2] The two studied how extrinsic rewards affect how adolescents learn and they discovered that when many people were promised rewards, they chose significantly easier tasks. They explain that rewards may shift people's focus from the activity itself to the reward they will receive after the activity.

Whatever the reason, promised rewards can undermine the intrinsic enjoyment of an activity. As a result, promising rewards to creators likely leads to more content but does not lead to better quality content. Promised rewards may even lead to worse quality content.

We have only focused on promised rewards so far, and Bénabou and Tirole additionally discuss ex-post rewards, rewards given unexpectedly at the full discretion of the giver. Unlike promised rewards which generally decrease intrinsic motivation, ex-post rewards generally increase it. They theorize that ex-post rewards signal that the work one did was valuable.

We consider subjective reward protocols to be in a gray area between promised and ex-post. If the developers or users creating on these protocols are purely honest, then the rewards are closer to discretionary. In the case of Steem, though, voting is of little cost to the voter, and thus not necessarily a signal the work is valuable. And as previously discussed, the metrics these protocols are gameable. If creators can pay vote bots on Steem then those rewards are more or less promised. If developers on Kin and Props can set up bots to send payments between different wallets, maximizing their number of in-app transactions, then those rewards also resemble promised ones. In both of those cases, we hypothesize that the ability to game the system will decrease intrinsic motivation, leading to worse quality content and apps.

Promising rewards do not necessarily create a healthy ecosystem for social media. We would, therefore, like to see next-generation social media market themselves as more than just social media with rewards. Networks with rewards are not 10x better networks that warrant users to leave existing platforms, and in many cases, we think they may be worse than traditional platforms that have little to no rewards for content creators.

We are personally advocates for paywalls and advertising. The revenue in both cases is entirely discretionary. We hope to create more opportunities for monetization so that more creators can realistically support themselves doing what they love.

Beyond Aligned Incentives

There is much more to designing next-generation social media than just aligning incentives. Social media platforms are not significantly better just because they reward creators and developers, and in many cases, we believe they may be worse. Because it is often impossible to reach consensus on subjective matters, subjective reward protocols must use gameable metrics. Additionally, these reward protocols ignore our natural inclination to form coalitions which would undermine some of these protocols.

We personally advocate that networks and governance systems should work with human nature, not against it. We believe that next-generation social media should not be drop in replacements for sites like Facebook, YouTube, and Medium. Instead, they should provide opportunities that these sites cannot.

---

1. See 'The Sybil Attack' for the proof behind this claim. ↩︎

2. Their paper is in The Hidden Costs of Reward: New Perspectives on the Psychology of Human Motivation. ↩︎