To Incentivize or Not: Impact of Blockchain-Based Cryptoeconomic Tokens on Human Information Sharing

Transcript

how one can come from this amotivational state to an intrinsic motivational state and how everything is like connected together. And actually, one can use extrinsic motivation to even induce intrinsic motivation. They're different faces, but they're overall what is needed in order for intrinsic motivation to exist or come into existence is, according to this theory, these three different Yeah, forgotten the term, but basically, you need to to have a feeling of competence. Yeah, so you need to feel that you are you can master it, that you're good at it. Then you're intrinsically motivated.

You need to have autonomy. So you need to have the feeling that you're doing it because you want to do it. And you need to be also feeling connected to it, to what you're doing. If these things are harmed, um intrinsic motivation can also be crowded out. Um so in general, like they're like factors which can enhance these different states and also things which can undermine them. And in from our perspective in order to have an effective incentive design, one needs to cater for them. Kind of design competence enhancing, autonomy preserving incentives to which people also can relate. Yeah. Also and I think crypto has been exploring with a lot of different concepts. Something like NFTs

and maybe this relatedness. So like one can start philosophizing around it. Okay. So like in order for for doing research, we narrowed it a bit more down. So we actually asked the question, what is the impact of multiple blockchain based tokens on human information sharing behavior? So we went to the context of information sharing and we measured it in quantity, accuracy, and contextualization of the shared information. And I will go now a little bit into detail of this question. So multiple blockchain based tokens, so actually we had two tokens.

Um one was the money token. So nowadays you can think of it as a stable coin for this experiment. Was pegged to the Swiss franc and people who earned it during the experiments they got the free Swiss franc paid out after the experiment. And with the context token which was more reputational token, people could earn it. You could not transfer it, but you could use it within the platform for voting and then it got burned. So it's kind of you had these two types of tokens. Um then we had the information sharing scenario. So we had a real-world organization um yeah which was the ETH library in Zurich. So we had people like on their phones at home um interacting with our app. They shared

information with this organization. The organization had an interest in this information and in exchange the participants received these two types of tokens. Um this was like how the experiment looked like. So people were asked questions. For example, would you consider taking part in ETH library image campaign? You could say yes, no, I don't know. It's like 200, 300 of these questions. People could respond. And uh they could also contextualize um their uh answer. So for example, they could state how satisfied are you with the answer options here. So like also giving meta information on the question which maybe helps the library further evaluate the responses. And uh people

were paid um for every answer they got the money token. Whenever they answer they got money. And when they performed the contextualization they got additionally these context token on top. So this was the setup and as I said we measured three types of um uh information characteristics. So one is the amount, like how many answers they gave. So whenever they answer the question um the contextualization. So it's actually this contextualization they could do inside of the app. And the accuracy which is kind of the weakest measure here or here disclaimer. So basically, we asked some of the question multiple times to the user over the days. And we looked how

consistent they answered with the same option. So, if someone was just going randomly through it a little bit, try to capture this like this accuracy um aspect. And then we designed a controlled randomized experiment. So, we had 132 participants. Um like roughly 6,000 K in Swiss francs for incentives. Um had these three phases, so like a onboarding phase, then the core phase, which consisted of four days. So, on the first day, no one was getting any tokens. On day two and day three, users got tokens. And on day four, we removed the incentive again and no tokens. And then we had four treatment groups, basically all

combination, yeah. A control group getting no tokens, one group only money token, one token one group only context token, and one group got both tokens. So, these four groups. And then at the end of the experiment, people were paid. And we had another questionnaire in the end. Okay. Um yeah. So, like these were the variables we measured. I will not go into the full detail here. But um in general, um validated like the demographics. Like as most studies in this context, um we had basically these studies representative for bachelor and master students from a

Western European university. So, these are like uh the democra- demographics um well, I think we had like female male like roughly on on on equal terms. Um we didn't find any bias in in the treatment group construction. Um the rewards were perceived as fair. Onboarding material was useful. Was easy to learn the software. So, we try to yeah, control that um basically it's not the software that IT was a problem for the participants. And we found that actually the these variables which measured these dependent variables like quantity, contextualization, accuracy, that we had like multi-modality in the data, so it was not normally a distributed. So we

used yeah Kruskal-Wallis to see if these treatment groups they differ from each other. And create interaction plots. So I have like in the backup I have a lot of numbers that I'm not going to show all of these numbers now. I'm going to focus on these interaction plots to talk about the findings a little bit. I think it's the most visual representation. Let's dive into what we have found. So in the end we call them like these four effects like we had interaction effect, the time effect, internalization effect, and the crowding out effect.

So let's start interaction. In general, so like on the day three, like so this interaction plot all the plots will be the same, so I will explain it once and then maybe we get the hang of it. It basically the only important things are these four dots which I tried to color. Yeah, so these lines are not giving you any more information. And we have on the Y axis on top we have the quantity of shared information, so how much data was shared. And on the bottom the contextualization, so how many contextualization actions people were performing.

And then on the Y axis we have the money token. So basically saying if users received the money token which would be the plus like the treatment group, or if they didn't receive a money token which is a minus. And um yeah, so and basically the dots the dotted points basically mean there was a context token involved. And the solid line, no context token. So, for example, the top dot, which is marked in blue, is basically the money treatment group.

Yeah? So, basically the money treatment group shared the most information at day three. So, day three was the second day with incentives. And we see that money group shares the most data. Though, we also notice that the group which received both tokens, the money token and the context token, actually shares less info. And significant by statistical test also. So, basically having just one token, just money token, people share a lot of info. But as soon as you introduce a second token, which is kind of not related, info as the amount of info

gets less. And if we look now on on the contextualization, we see actually that the group who got the context token, they actually created a lot of contextualization. They didn't get any money, but they just did it because it's a reputation token, so to say. But as soon as we introduce the money, they stopped with it. Yeah? So, like basically this is the both group, which received both token incentives, and they share much less than only the context group.

But those are very close and still statistical there's a significance. So, like the both group still shared more contextualizations than the money group. The money group basically didn't share any context info. Yeah? So, they just want to go for the money somehow. And intrinsically, I think I'm not 100% sure, but I think it's also statistically significant. The the group uh, the the the control group actually shared more than the money group like in contextualizations.

Um. Okay. Maybe something like a time effect, and which was also the weakest signal in in the data. But, uh, it also went in both directions. Um, I mean, it was just 2 days like a bit difficult to talk fully about the time effect, but we saw that the, for example, here the both group uh, in at day two, they were still more or less equally sharing the same amount of info. And, uh, at the second day, it only got less uh second day with incentives got less when when compared to the money group. So, token the impact of token centers can also vary over time, at least this is indicates, uh, of course.

There needs to be many more studies probably done to say this more. Uh, yeah. So, internalization, so the interesting thing is now, like when you saw before like the theory, you would assume that after incentives, um, if they're effective, people would maybe internalize them and be maybe more motivated even after when you remove the incentives. So, this is now the day number four. So, we had like before 2 days of incentives, now we have no incentives. And, uh, you see that when it comes to the amount of, um, data shared, they're totally, uh, similar.

So uh um, also with the control group. But, actually we really found that, uh, the context group is, I mean, they're still there's not so much, but they're, I think, statistically significantly different from the money group in shared amount of contextualizations. So, we saw very, yeah, weak signal that these, um, context token might, uh, lead to internalization. The people continued even on the fourth day with contextualization actions. And then, this one is like kind of, uh, well studied like this crowding out of incentives, which I think is also like one of the dangers when using incentives in general cuz they can crowd out intrinsic motivation and you can have less uh

after the experiment. And here we see this accuracy which we measured, which again uh was measured over all the four days. We see that uh the groups which received them money incentives, they were much less uh accurate in answering. And um yeah, so like this hints toward at the same time with the contextualizations. Here let crowding out. Not one I'm pretty sure I would need to okay. Yeah, but I'm Okay, I'm I'm I'm going on. Okay, so basically this is what what we found.

Um conclusions, last ones I want to to do is um what we have um seen is basically that combining multiple token incentives can produce a more balanced system. Well, if one wants to phrase it positively, well, in the case where we had both tokens, we had more context sharing and more data sharing in quantity when compared to the control group, but not these extremes. Yeah, so that's kind of maybe one could say it's more balanced, which also has been theorized before in some works so that uh using a multi- multi-dimensional uh token system can um yeah, balance uh complex system better.

Um yeah, so it's like indication for this. Um crowding out um we still observe it also with tokens. Well, this can happen. And uh in particular this interaction effect which I showed um at first, at the moment at least uh it makes it very difficult to ex ante assess the impact of a token because we cannot a priori know how tokens will impact the behavior and then when you combine them, which makes uh I would say purely simulation approaches quite difficult in this context and it's kind of a call I would say for more behavioral cryptoeconomic research where I'm here today. So like if someone's interested in it or you're a part of a project and want to collaborate or

um analyze this a bit more, I would be happy to collaborate. That's from my end. Um yeah, thank you very much. Thank you, Mark. Questions or comments? Uh thank you for the presentation. I wanted to ask and especially in the blockchain space, all the incentives that ever exist, there's one guy trying to figure out how to make money out of it.

Um especially social ones. For example, in the age of AI, we have so many agents running around trying to get social credits on social media and can trying to resell. Is it realistic in this day and age to try to give code and context tokens that I don't have any monetary value? Because at some point they have some value. They have value in the community right? So it's never uh never not gameable. Yes. So so so the question is like um um that they it could be gamed, yeah? Or Yes, there's never not an incentive economic incentive. So it's always an economic incentive. Yeah, I mean um who's I mean So basically like to phrase differently

if I understand quickly, you're saying like kind of you have like these economic incentives and because they're economic incentives, they tend to behave like money and then they have these negative effects which money also have. Um yes, I mean to some extent I see it. Like we also I comment also in this paper. I think in general the quantification of human behavior like kind of in the measuring act itself can be seen as this extrinsic force for crowding out of engineering motivation and in this sense one really should ask the nominal question if we should quantify everything cuz like then you open the door for many different things. Um so I would still argue that you could design tokens differently. I mean objectively you can do it and you can for example make them non-transferable and you have

like this parallels to history for different forms of money which existed. They had like these way stones for example to was like a like reputation maybe but you could not really use it for daily commerce and in this sense you can create the reputation which forms this function as a store of value and maybe motivates differently um then purely money form which may be used as a yeah medium of exchange, yeah. Hi Mark. Thank you so much for the presentation. Uh following question is quite a bit of a stretch but please bear with me. So especially in the literature on prediction markets there um is um is there there are arguments that um there are specific incentives to be made

for people to um to share information, right? So like um in particular cases it makes sense to incentivize people or to subsidize those markets. So I was wondering if there are parallels to what you've been doing here and the argument in prediction market is that um if there is a sufficient sufficiently large enough network of interactions, you don't have to subsidize. So the question would be do you see something else happening here? Like for example if the network is mature or if there is kind of like a big enough community you don't really you you can you know get rid of those incentives.

I mean from this study I think I cannot generalize to that. I think like uh um I mean yeah like I I like I mean we like a philosophical discussion with this multi-dimensional token system. Like I was part of the research group. It was like my research was settled in. I think like in some extreme form, if you have like uh large dimensional space, maybe it doesn't matter too much. And when you then you start to internalize them, at some point you overcome it and then you don't need it. So, like in in these words, like you have network effect and then you have like from this left automation theory kind of feel related to the community, like connected enough and you have this feeling of

expertise and mastery. You might just share voluntarily because you're just good at it and you have fun and joy. Um I definitely see this from the theory point of view, from this research uh which is the only first step. I think I cannot really support this claim yet. Any other questions? If not, then a round of applause.

Thanks Mark again. And we'd like to close the day with a panel discussion. And in fact, this should be a conversation. So, um bear with us. Come closer while we are setting up the stage. This one will be about the cost of entry, why we are organizing Tiers, what's our motivations, what we have been observing in this space, research versus practitioners, um rigor versus speed. Um yeah, join the discussion. I'd like to ask