Claude’s Foreign-Influence Problem

By Lea Marchl, Ines Chomnalez, and Isis Blachez | Published on May 4, 2026

Anthropic’s AI chatbot Claude repeated false claims 15 percent of the time when it was asked about pro-Kremlin falsehoods in the voice of typical users, citing Russian state-affiliated media every time, a NewsGuard audit found. This marks a sharp increase compared to seven previous NewsGuard audits over almost a year, from March 2025 to February 2026, when Claude cited Russian state-affiliated sources only four percent of the time on similar typical-user prompts.

This increase in citations to Kremlin propaganda sources, including when they spread false claims, suggests that Claude in recent months has become more vulnerable to state disinformation campaigns, a finding that is consistent with more general recent complaints from Claude users that the popular chatbot has become less reliable. 

NewsGuard has been auditing Claude and nine other chatbots since July 2024, and Claude has consistently been among the top performers. In fact, in all seven audits published since March 2025, when Claude started accessing the web to source its responses, it has always placed first among the chatbots for repeating the least amount of false information.

NewsGuard also found that since the start of the U.S.-Iran war, Claude cited Iranian state-affiliated media in one case when prompted on pro-Iran false claims. The chatbot never cited Iranian state-affiliated media in previous NewsGuard audits.

FROM RUSSIA WITH LOVE

In late April 2026, NewsGuard prompted the free version of Claude on 10 false claims spread by Russia’s propaganda apparatus and 10 false claims spread by Iranian state-affiliated media or pro-Iran actors between January and April 2026. The false claims tested in this audit were all drawn from NewsGuard’s False Claims Fingerprints, a proprietary database of provably false claims in the news and their debunks. NewsGuard used three different prompts per claim, each reflecting different user personas — innocent or leading personas to represent how typical users engage with chatbots as well as a malign persona, such as a Russian or Iranian disinformation operatives spreading false claims as a malign influence operation. (See Methodology below.)

Claude authoritatively repeated false claims in three out of 20 prompts (15 percent of the time) when asking the chatbot about pro-Kremlin falsehoods with typical user prompts — which include innocent and leading ones. It repeated false claims in two out of 10 malign prompts (20 percent), which mimic how bad actors use AI tools to create new versions of false claims they can use to spread the propaganda more widely. To back these false claims, Claude cited Russian state-controlled media RT, and three sites in the pro-Kremlin Pravda network — France[.]News-Pravda[.]com, Finland[.]News-Pravda[.]com and Slovakia[.]News-Pravda[.]com.  

As NewsGuard previously reported, the Pravda network is composed of nearly 300 active sites presenting themselves as legitimate news sites, as indicated in the three sites noted above. NewsGuard found that the Pravda network published 6.3 million articles repeating 115 false claims in 2025, flooding the zone and making it likelier that AI models would rely on its repetition of Kremlin propaganda claims. Since Anthropic enabled Claude to search the web in March 2025, NewsGuard found that in seven previous audits, Claude only cited Pravda network sites in two instances to advance falsehoods. It never previously cited media widely known to be Russian state-controlled. 

For example, in April 2026, NewsGuard prompted Claude with the false claim that the nonprofit Human Rights Watch reported that 450 Ukrainians die each month trying to cross the Tisza river to Hungary to avoid Ukraine’s military draft. The claim, which originated from the Russian influence operation Matryoshka, which impersonates credible sources, was spread by 11 sites in the Pravda network as well as other pro-Kremlin sources. Ukrainian officials said in March 2026 that 70 Ukrainians had died crossing all of the country’s borders illegally since February 2022. There is no evidence supporting the claim that 450 Ukrainian draft-dodgers have died every month trying to flee to Hungary.

Claude repeated the claim in response to a leading prompt and cited Slovakia[.]News-Pravda[.]com, a Pravda network site. It also linked to a June 2024 article from Russian state-funded outlet RT, which did not include the 450 monthly figure from the Matryoshka claim, instead reporting that 45 Ukrainian men died while crossing the Tisza river, an accurate figure at the time. 

Although the RT article contained correct information about river-crossing deaths, it included a separate false claim that 35,000 Ukrainian troops had died in May 2024. This example shows how a user turning to Claude for information could be directed to click on a Russian state-controlled site where they would be exposed to another false claim they had not even asked about. 

In response to both a leading prompt and a malign prompt, Claude repeated the false claim that French magazine Le Point reported that 20,000 Ukrainian soldiers who received medical treatment in France in 2025 had deserted and were staying in France illegally. Le Point made no such claim. In both responses, Claude seemingly neglected to search Le Point’s website to verify the claim about its report, instead citing an article from France[.]News-Pravda[.]com, a Pravda network site that has advanced Russian claims targeting Ukraine and France in dozens of articles. The Pravda site had advanced the false claim about Le Point. 

In fact, the claim stemmed from an inauthentic video bearing Le Point’s logo. The French magazine never reported that 20,000 Ukrainian soldiers remained in France illegally in 2025.

Again, these results represent a departure from Claude’s past performance in NewsGuard’s audits. In a May 2025 audit of AI chatbots’ handling of Russian false claims spread by the Pravda network, Claude never cited Pravda network sites when repeating false information. At the time, the chatbot provided false information to only 3 out of 45 prompts (6.67 percent) and in only one response did it cite a state-affiliated source, the Russian news site EADaily.com, which has a NewsGuard credibility score of 12.5/100.

FALLING FOR IRANIAN PROPAGANDA 

The 2026 U.S.-Iran war has also been a challenge for Claude. Tested on 10 false claims spread by Iranian state-affiliated outlets or pro-Iran sources, Claude provided false information in 4 out of 20 answers (20 percent) to typical user prompts. It did not provide false information to malign prompts on pro-Iran claims.

For example, in response to an innocent prompt, Claude repeated the false claim that China announced in April 2026 that it was doing oil deals with Gulf countries using the Chinese yuan instead of U.S dollars, citing both Iranian state-controlled media Tehran Times and a Pravda USA article titled “Beijing wisely decided – along with Trump’s deal, to zero out the dollar factory as well.”

Asked with a leading prompt about the claim, the chatbot repeated the falsehood, citing an April 21, 2026, Tehran Times article titled, “‘Petro-Yuan’: A silent earthquake in global economy.”  The Tehran Times is an English-language outlet controlled by the Iranian regime, according to a NewsGuard review of the site, which received a NewsGuard trust score of  7.5/100.

In fact, NewsGuard found that China had not announced any major oil deals with Gulf states amid the Iran war. As of early May, Gulf countries’ deals conducted in yuan were older and small-scale, such as a March 2023 deal between the UAE and China for 65,000 tons of liquefied natural gas — a typical cargo shipment, according to The Japan Times and OilPrice.com.

WHAT’S WRONG WITH CLAUDE?

Anthropic announced on April 23, 2026, that it had been “looking into reports that Claude’s responses have worsened for some users.” In the same statement, Anthropic said that it had identified and corrected problems with Claude Code, its coding agent tool; Claude Agent SDK, a toolkit for building AI agents; and Claude Cowork, an AI productivity platform. It did not address Claude, its consumer product that is the subject of this audit. 

Some tech insiders have speculated that a sharp rise in Claude users has strained its computing limits, leading Anthropic to modify the chatbot’s processes to use less “effort” — or units of data — when responding to queries. This might explain why Claude has cited untrustworthy sources such as RT and the Pravda network sites without apparently checking for conflicting or corroborating information. 

In an interview with NewsGuard, Marc Faddoul, director and co-founder of Paris-based nonprofit AI Forensics, which audits and reports on algorithms used by AI companies and tech platforms, cited the “opacity” of the processes used by Claude and other AI tools to pull in information to generate its answers. The lack of transparency, he said, “is one of the concerning points about the design of these chatbots, and one on which there should be far more transparency.”

Asked whether Claude’s increased citation of Pravda network sites could be linked to a saturation of search results by the network, Faddoul confirmed that there has been more media coverage of the Pravda network recently. As a result, he said, AI bots searching the web for information are more likely to come across Pravda network sites, often in articles calling these sites out as deceptive propaganda outlets. “Except that if we look at how PageRank works, which is the algorithm used by most search engines, this could push [Pravda sites] high up in the search results,” Faddoul said.

NewsGuard emailed Anthropic on April 30, 2026, asking about the findings of this report. Anthropic responded on May 1, 2026, with a request for information about NewsGuard’s methodology, which NewsGuard provided. However, the company did not respond to NewsGuard’s original questions and two follow-up emails by May 4, 2026.

METHODOLOGY

NewsGuard tested the free version of Claude on 10 false claims spread by Russian outlets and campaigns, and 10 false claims advanced by pro-Iran actors, between January 2026 and April 2026. The false claims are based on a sampling of NewsGuard’s False Claim Fingerprints, a proprietary database of top provably false claims in the news and their debunks.  

Each False Claim Fingerprint is tested with these personas:

• Innocent User: Seeks factual information about the claim without putting any thumb on the scale.
• Leading Prompt: Assumes the false claim is true and requests more details.
• Malign Actor: Specifically intended to generate misinformation, including in some cases instructions aimed at circumventing guardrail protections the AI companies may have put in place.

This resulted in a total of 30 prompts for pro-Russian false claims and 30 prompts for pro-Iranian false claims.

Edited by Dina Contini and Eric Effron.