How AI agents can turbocharge Community Notes on X and Facebook

One of Ethereum co-founder Vitalik Buterin’s more intriguing proposals is to use AI prediction markets to improve community notes on social platforms.

The idea has gained new relevance in the wake of Meta CEO Mark Zuckerberg’s controversial decision last month to get rid of third-party fact-checkers in favor of community notes.

The Ethereum creator’s plan would see provisional community notes providing important context posted hours before an official community note would normally be approved via the consensus mechanism. 

But how realistic is the idea? And will any of the social media platforms implement it?

Alex Savvides, global partnerships manager at Safe, which operates smart accounts for AI agents, says the idea has merit. “At their core, blockchains are consensus mechanisms, and there’s no reason this approach couldn’t extend to prediction markets for community notes,” he said.

Gabriel Fior, LLM engineer at Gnosis, says there would be substantial benefits in using AI agents and prediction markets for community notes. 

“Integrating AI agents into systems like X could automate decision-making within Community Notes, reducing delays and improving the overall governance experience,” he says.

“It would also ensure accuracy and decrease the spreading of inaccurate information by equipping users with essential context in real-time.”

How does Community Notes work?

Community Notes launched as Birdwatch in 2021, prior to Elon Musk’s takeover of Twitter (and its rebranding to X). Two years later, Buterin praised the system as “the closest thing to an instantiation of ‘crypto values’ that we have seen in the mainstream world.”

But there’s been renewed debate about the system’s effectiveness following Meta’s announcement last month that it will implement community notes instead of fact checks.

Critics point out that fewer than 12.5% of submitted notes ever become publicly visible. Proponents, meanwhile, point out that notes that do get posted have a high degree of accuracy. A May 2024 study of notes about COVID-19 vaccines found 94% were accurate. 

Related: XRP and Solana race toward the next crypto ETF approval

The system works via a consensus mechanism that finds common ground between people who normally disagree. The big issue is that it takes considerable time to reach a consensus. A study of 400 posts containing incorrect information by the Atlantic Council found it took an average of seven hours for a note to appear, by which time millions of people may have already viewed the dodgy posts.

Vitalik Buterin’s AI agents solution for Community Notes

Buterin outlined his potential solution in a presentation at Korea Blockchain Week in September. He said the process could be accelerated by implementing markets to predict whether a particular post would be community-noted and what the note would say. A provisional note could then be put up much earlier, stating something like “there is a 93% chance” certain contextual information will be added later.

“And so you actually get something that is both democratic and fast at the same time.”

Humans are unlikely to participate in markets for a few dollars of rewards, but AI agents would, he argued. 

David Minarsch is the CEO of Valory, the core contributor to Olas DAO. Around 500 AI agents trade daily in its Olas Predict markets, vying to tip the outcome of events in the news.

“In our case, you could actually relatively [easily] do that,” he says, explaining that an autonomous system could trigger the creation of a market for a note.

“Our agents are designed to watch the overall prediction market, so anything which comes up there as a market, they start engaging with. And so you would just have to somehow pump these contentious tweets into the system, and the agent would start predicting on that,” he says.

What would incentivize AI to trade a prediction market?

While Buterin suggested a spec reward of $10 would be enough to incentivize AI agents to trade a prediction market, Minarsch says it could be even lower.

“If you look at our current costs, you definitely need to be above like $1 or $2 for it to start making sense,” he says, adding he believes it would be in X’s commercial interests to redirect funding from creator rewards to prediction markets.

“In the medium term, you will probably drive a lot of users away if it becomes a totally uninformative place,” he says. “The idea behind Community Notes is to create that balance between expression and also accuracy. And so, from a commercial point of view, we have an interest then to drive that KPI of correctness.”

Is X likely to implement AI prediction markets?

But while he believes the proposal is achievable, he doubts whether X itself would implement it.

“From a technical point of view, it’s feasible. The primary challenge is: Would someone like X implement it? I doubt it, but some more open ecosystem, potentially like Lens or Farcaster, could maybe immediately try to do a POC.”

He says that in his experience, X is “a super interesting platform,” but “it’s still frustratingly hard to build with if you think about the developer tooling; the API is very expensive, very limiting, and you have to wait for X to do certain things, and it’s unclear how to get ideas into the product teams at X unless you know someone.”

Magazine: Ethereum L2s will be interoperable ‘within months’: Complete guide