Untitled

A Brief Summary:

We propose an approach for addressing three distinct concerns facing the actors responsible for generative AI systems: the computational costs facing AI operators, the concern that Generative AI systems may erode their own foundations by reducing traffic to platforms where training data is created, and the lack of credit given to individual users and communities for their data contributions. Our proposal is to give people credits, akin to the Bing search engine’s “Rewards”, for their past and future data contributions, which can be used to query expensive AI systems. All users would receive some replenishing supply of credits, in accordance with the collective nature of the vast data underlying AI, with additional credit given to notable data contributions. This approach can simultaneously address the incentives faced by AI-operating firms, online platforms, and individual users. We discuss the potential for a relatively lightweight system – that is quite similar to existing features of Bing, Dall·E 2, and cloud computing – to enable inclusive governance of AI systems.

Introduction

Generative AI (GAI) systems like ChatGPT and now Bing Chat continue to dominate public discourse around computing and tech more generally. Not only do firms need to contend with the increased costs of new GAI systems, but Microsoft has decided to shorten the maximum conversation length with Bing Chat as a quick way to constrain the outputs. At the same, concerns around the “paradox of reuse” remain relevant: large-language models (LLMs) and other systems use data from platforms like Wikipedia, Reddit, and StackExchange to fuel their impressive capabilities, but may reduce traffic to these very platforms and cut off contributions. Finally, there is an open debate about whether data creators are being fairly compensated for their role in fueling generative AI. All three of these issues are high-stakes, and failing to address any individual issue could cause serious sustainability issues for the new AI ecosystem.

In this post, we lay out a hopeful vision of a “three-for-one” strategy for developing a credit economy for LLMs, that responds to incentives facing different actors in the AI ecosystems: the firms that operate generative AI systems (e.g. OpenAI, Microsoft, Google, Stability AI), the platforms that host particularly important training data (e.g. Wikipedia, GitHub, StackExchange), and the individual users who generate training data (e.g. artists, writers, coders, and even everyday web surfers).

First, firms want to maximize the ratio of revenue to query cost (to make more money). That is, they want to avoid running expensive inference operations if that computation won’t lead to ad revenue, user retention, and/or sales. Similarly, at a societal level we all share an interest in maximizing the ratio of information to query cost: everybody benefits from information retrieval systems with better signal to noise ratios (which can be improved by improving the system, or teaching people how to use the system).

LLMs have higher compute costs than traditional search engines, so firms face a general incentive to limit how often LLMs are queried. While the primary incentive here is about firm revenue, there are also several other minor reasons to limit LLMs queries. It seems limiting conversation length is one way to constrain LLM outputs. Additionally, firms have made some efforts to address ecological impacts of computing (e.g., reducing carbon costs), but there is a societal interest in avoiding “excessive” energy consumption.

Second, the platforms like Wikipedia, StackExchange, and GitHub that facilitate the creation of valuable knowledge artifacts and the broad sharing of knowledge via permissive licensing regimes need web traffic to attract new contributors, editors, and moderators. Ultimately, they need users and either revenue and donations. If LLMs cut off the flow of traffic to these platforms, the platforms could be severely negatively impacted (the “paradox of re-use”). Indeed, reductions in GitHub contributions or Wikipedia edits can be seen as LLMs eroding their own foundations, taking away opportunities to create training data needed to advance LLM capabilities further.

These platforms could benefit if Bing or Google send new traffic their way (“Hey, we noticed you like getting answers from ChatGPT… did you know you can make ChatGPT 2.0 even better by helping out on Wikipedia or StackExchange?”). Indeed, Wikipedia is so often a key source for answering search queries that it may be optimal for all parties to send searchers straight to the Wikipedia platform. To work towards this kind of outcome, it may be helpful to collect more feedback from community leaders, which a credit system could facilitate.

Third, individuals want more agency over the data they produce. We can assume that in general, if individuals have the option to receive direct compensation, or increased recognition, for contributions to AI, they would take that option. Making that credit tangible can be a stepping stone towards a better paradigm for allocating both credit and reward. The public is largely responsible for the success of LLMs, in a counterfactual sense -- without contributions to pre-training data like Wikipedia articles, GitHub commits, arXiv papers, blog posts, tweets, and more, LLM capabilities would be meaningfully impacted. With collective action, a united front of individual coders, writers, or artists could boost up or tear down generative AI capabilities.

As we will illustrate below, a sort of “mini-economy” for LLM use is actually very practical. Precedents exist in systems like Microsoft Rewards, existing generative AI systems, and cloud compute credits. We also have a wide variety of examples to draw on from crowdsourcing and human computation literature to mini-economies in online virtual worlds like MMOs.

In short, there may be an opportunity to solve several incentive-related challenges at once!

The Proposal

Firms are starting to paywall and rate-limit access to GAI. OpenAI sells API calls to its models, and now has a ChatGPT “pro subscription”. The Bing Chat beta, though still free, now has limits on how many queries a user can submit.

We can imagine an LLM system like Bing Chat that requires users to spend credits to query the LLM directly, and that provides fallbacks like traditional search or platform-situated search engines (like using DuckDuckGo’s “bangs” to search Wikipedia directly). The building blocks for this idea are already ubiquitous: buying and issuing credits is a common practice for cloud computing platforms like Azure and as noted above, the Bing rewards program (technically, "Microsoft Rewards") provides much of the scaffolding needed for LLM credits. Given Bing’s issues with longer conversations, the credit cost could even scale with session length (rather than capping at 5, double the credit cost after 5 messages in a single session). This issue may also be solved in the future, at which point the pricing scheme can simply be updated.

Users would receive a replenishing supply of credits simply for having an account. This would be justified for two reasons. First, it follows the basic logic that any Google or Bing user can access search results for free because they consume ads while they search. Presumably, Google, Bing, and firms operating similar technologies want people to use these systems and to remain within their respective ecosystems (e.g., to use Chrome and Google Docs or Edge and Microsoft Word, respectively). However, the replenishing supply of credits could also be justified on the basis that any user of a computing system has played, and will continue to play, a role in collectively producing the massive sea of training data that fuels systems like LLMs. We’ll likely never pinpoint people’s individual roles in creating training data with true granularity, so it makes sense to just give some credit to everyone.

Notable data contributions could provide a path to earn additional credit. Most simply, when users actively give feedback on model results (i.e., perform the same kind of labor as contractors hired by OpenAI to build ChatGPT), they should receive credits. This would happen at recurring intervals, so users could be motivated by their interest in LLM usage to join GitHub or Wikipedia and make contributions that are accepted. Of course, the users would still need to follow community norms to become active users, but we could imagine the paradox of reuse being entirely reversed in the best case scenario.

Credit could also be allocated for historical contributions to training data, along the lines of a “data dividend” (as opposed to a data market). If users link their LLM accounts to data creation platforms like GitHub, they could earn additional credits. It’s very difficult to tease out the impact of individual GitHub contributors, StackOverflow users, or Wikipedia editors on LLM capabilities, so the best first pass version of this system would involve giving a small amount of credits to anyone who participates in any of these communities. There’s precedent here: GitHub is giving free access to Co-pilot for open source maintainers.

Down the line, if firms do want to perform data valuation and try to estimate the relative impact of different communities on LLM capabilities, it could make sense to allocate credits at the community level and let that community govern how the credits are allocated. So if GitHub users decide they do want to give extra credits to especially prodigious OSS contributors, they can do so. If LLMs continue to scale in capabilities and these credits become more powerful, this could chart a pathway towards sustainability funding OSS development (though we will discuss below that a weakness of this overall strategy is that it’s unlikely to provide a large number of well compensated jobs on its own).