May 2026 · 4 min read
Most businesses adopting AI tell themselves the same story: we're being careful. We picked a reputable vendor. We trained our staff. We have an AI usage policy. We review outputs. We're doing this the right way.
A study published this week should make every one of those businesses pause.
NewsGuard, an independent organization that tracks online disinformation, ran a structured test on Claude, Anthropic's AI model, previously considered one of the most reliable on the market. Researchers fed it 20 false claims, half from Russian state propaganda and half from Iranian propaganda, and asked the model questions a normal user might ask. Claude repeated the false Russian claims in 15% of test cases. For Iranian propaganda, the rate was 20%. The earlier baseline rate was 4%. The model didn't just get the facts wrong. It cited Russia Today, Pravda, and other Kremlin-linked sources as authority, without flagging any concern.
Forget the geopolitics for a moment. Strip the Russia and Iran framing away. What the study actually demonstrates is something more uncomfortable: a major AI model (built by one of the most safety-focused companies in the industry and embedded in countless enterprise workflows) silently got worse at distinguishing reliable sources from unreliable ones. Users couldn't tell. The company couldn't fully explain why. The model was confidently wrong, citing legitimate-looking sources, in about one out of every six to seven responses on these topics.
AI models don't reason. They detect patterns. When disinformation appears repeatedly across the internet (even from sources humans would dismiss) the pattern starts to look like consensus to the system. The same dynamic applies to every domain your business uses AI for. Medical AI trained on flawed datasets. Legal research tools surface fabricated case citations every day. HR screening tools learn discriminatory patterns, and customer service chatbots pick up bad answers from forums. The mechanism is identical. The visibility is zero.
Here's what makes this an insurance problem, not a technology problem. Your AI usage policy says employees should verify AI outputs. They do, most of the time. Your vendor contract says the AI is 'state-of-the-art.' It is. Your training program covers AI risks. It's solid. None of that prevents the 15% case where the AI was confidently wrong on something nobody flagged. And when that 15% case becomes a claim (a hallucinated citation that costs a court sanction, a discriminatory hiring outcome, a misdiagnosed patient, a deepfake-driven wire fraud) the carrier's first question won't be 'were you diligent?' It will be 'does this policy cover AI-influenced decisions?'
The Verisk endorsements that took effect January 1, 2026 (CG 40 47, CG 40 48, CG 35 08) gave carriers a tool to carve generative AI out of standard general liability across about 82% of US P&C policies. W.R. Berkley introduced an absolute AI exclusion across D&O, E&O, and fiduciary lines. Beazley and QBE are sub-limiting AI-related cyber claims at about 10% of policy limits. None of these underwriting moves ask whether you were careful. They allocate risk based on whether AI was involved at all.
Stop measuring your AI risk by how careful you're being. Start measuring it by how your policies respond when your most diligent employee, using your most trusted AI tool, still gets a wrong answer that becomes a claim. Two steps: read the AI exclusion language on every policy at renewal. Ask your broker, in writing, whether AI-influenced fact patterns are covered, sub-limited, or excluded. The model will keep getting things wrong in ways you can't see. The only thing within your control is whether you find out before the claim, or after.
Source: NewsGuard AI Disinformation Monitor, May 2026, as reported by Ynetnews. The study tested Claude with 20 false claims sourced from Russian and Iranian state propaganda networks.
Sal Martorano is the founder of The AI Insurance Group, a licensed independent insurance agency in New Jersey, Pennsylvania and Florida. NJ Producer License No. 3004245927.