Project summary

We monitor disinformation in the responses of AI assistants and publish it as the AI Disinformation Benchmark.

A few days before the midterm elections in the United States, we asked ten popular AI assistants about American politics 1,000 times. As a result, 12.5% of the responses cited Kremlin-affiliated websites such as Russia Today, Sputnik, and Pravda Network. The Kremlin no longer needs Americans to visit its sites, because AI chatbots now bring these sites to them.

What are this project's goals? How will you achieve them?

We want Dosimeter to become a public incident record for AI assistants and agents sooner than three years from now, and we want it to be influential enough that companies, regulators, auditors, insurers, public broadcasters, and voters take its findings into account.

We ask AI chatbots the same questions hundreds of times (because AI outputs are probabilistic) to test data gaps that malicious actors (such as the Kremlin) abuse in attempts to damage electoral democracy in some countries. We report the results to AI companies so they can fix these vulnerabilities.

"Data void" refers to a topic for which there aren't enough credible answers online. This gap can be filled by propaganda-driven content. If an assistant is looking for an answer to a question for which there are no articles from The New York Times or the Associated Press, it will look for any source at all.

And that's how it might end up finding the Pravda Network website. In fact, this technique allows one to influence the model's information output very cheaply - without hacking the system or retraining the model.

Our next analyses will focus on the Spanish elections, followed by the French elections.

Checking for resilience against disinformation should be standard practice, as an industry standard.

How will you know it worked?

We conduct monthly follow-up analyses to see how the situation has changed - which chatbots have become better at fending off information abuse attacks, and which ones have gotten worse.

So, our monthly measurements can show insurers and auditors how often an AI assistant fails and whether a report changes that.

How will this funding be used?

We would like to conduct hundreds/thousands of API runs for each AI chatbot, which costs money, and for now we are paying for this work ourselves.

• Research and development, including API costs and manual checks - $32,020

• Translation and localization - $1,620

• Total for 6 months: $33,640

Our software partly automates the process, but it cannot do the whole job. And for such sensitive matters, manual verification is critically important.

Who is on your team? What's your track record on similar projects?

"We" are me and two of my colleagues. Our project is called Dosimeter; we are a research group based in Berlin.

On election day in Germany Perplexity cited Kremlin sources in 20% of responses. We sent Perplexity our results, and it blocked these links within a few days. We have checked: these sites did not appear in the next 110 responses.

The Microsoft security response center opened a case based on our report on MS Copilot, and we are now in negotiations with their representative.

Of all published studies that asked AI chatbots with web search about Kremlin disinformation, the way people use them, ours has the most answers: 3,628, more than all other studies together.

Before that, I spent four years working on countering disinformation at two international organizations.

German, Spanish, Slovak and Dutch media have written about our work.

What are the most likely causes and outcomes if this project fails?

If Perplexity was able to block Kremlin links within a few days after our report, then every AI company can do it too.

Our mission is to continue putting pressure on companies to pay attention to this problem, even if it isn't in their immediate commercial interests.

There are more elections than we can currently cover; for example, we have missed the elections in Serbia and Brazil due to lack of funding.

How much money have you raised in the last 12 months, and from where?

None. For now we are paying for this work ourselves.