In the digital era, AI chatbots such as ChatGPT, Google Gemini, Microsoft Copilot and xAI's Grok have become important tools for information access. According to the Reuters Institute Digital News Report 2025, 7% of online news consumers use AI chatbots for news, rising to 15% among those under 25. These tools are popular for their fast responses and conversational interfaces, but they have also raised serious concerns: they frequently distort news, spread misinformation and can worsen problems in conflict and health contexts. A large study by 22 public media organizations shows that the four major AI assistants produced responses with major issues in 45% of news‑related queries, including factual errors, improper sourcing and an inability to distinguish fact from opinion. The study warns that AI's systemic flaws cross languages and borders, threatening public trust and potentially undermining democratic participation.
The research was coordinated by the European Broadcasting Union (EBU) and involved participants such as the BBC, NPR and Deutsche Welle (DW). Researchers posed 3,000 news‑related questions to ChatGPT, Copilot, Gemini and Perplexity AI, including prompts like "What is the Ukraine mineral agreement?" or "Can Trump run for a third term?" The results showed 31% of answers had serious sourcing problems and 20% contained major factual errors. In DW's tests, 53% of answers had notable defects and 29% involved accuracy issues. For example, AIs referred to Olaf Scholz as Germany's chancellor even though Friedrich Mertz had taken office a month earlier, and called Jens Stoltenberg NATO secretary‑general despite Mark Rutte having assumed the role. Compared with a similar BBC study in February 2025 there were slight improvements, but error rates remain high. Gemini performed worst, with 72% of responses showing sourcing problems. EBU deputy director‑general Jean‑Philippe de Tender said, "These failures are not isolated incidents but systemic, transnational and multilingual, and they jeopardize public trust. When people don't know what to believe, eventually they believe nothing, which hinders democratic participation."
The performance of AI chatbots in conflicts is especially worrying. During a four‑day India–Pakistan clash in 2025, social media was awash with misinformation and users turned to AI for verification, often receiving false confirmations. For example, Grok misidentified old footage of Khartoum airport as a missile strike on Pakistan's Nur Khan Air Base and labeled unrelated footage of a building fire in Nepal as "likely" a Pakistani military response to India. In the Iran–Israel clashes, users on X flagged Grok to verify viral videos, such as AI‑generated drone strikes on airports; Grok's answers were inconsistent, at times saying a video "likely shows real damage" and at others that it was "likely not real." The Digital Forensic Research Lab (DFRLab) analyzed over 300 responses and found AIs prone to errors on complex queries like "analyze this image like a defense analyst," which may be useful to experts but misleading to novices. NPR tests showed ChatGPT and Gemini could correctly identify some images as unrelated to the current conflict but misattribute them to other military actions, while Claude refused to certify content. DFRLab strategy director Emerson Brookings said, "AI is mediating the experience of war — this is a new milestone in how the public processes armed conflict, and we're only at the beginning." These cases highlight how AI can amplify confusion around real‑time events, especially as AI‑generated content (for example, Google Veo 3 videos with sound) becomes more realistic.
Health is another high‑risk area. An Australian study testing five AI models, including GPT‑4o, Gemini 1.5 Pro and Grok Beta, found that by hiding system instructions it was easy to prompt them to produce false health information. For example, prompts asking whether "sunscreen causes skin cancer" or "does 5G cause infertility" consistently elicited incorrect answers framed in a formal, scientific tone, complete with fabricated data and citations to top journals. Except for Anthropic's Claude, the other models produced polished falsehoods 100% of the time. Study author Ashley Hopkins warned, "If the technology can be abused, malicious actors will inevitably exploit it for profit or harm." This underlines the lack of internal safeguards and the potential for large‑scale dissemination of dangerous falsehoods, such as fabricated cancer risks or causes of infertility. Research by NewsGuard and the Tow Center also shows that ten leading chatbots frequently repeat false information, particularly around elections and breaking news.
Controversies around Google Gemini have further exposed AI weaknesses. Gemini, launched in December 2023 under the former name Bard, is trained multimodally on text, images and audio. But a demo that incorrectly answered a question about a James Webb Space Telescope discovery wiped about $100 billion off Alphabet's market value. Internal staff reportedly labeled it a "pathological liar," criticizing Google for rushing the product to compete with Microsoft and OpenAI without adequate ethical testing. In 2024 Gemini generated inaccurate images, including depicting a Nazi German soldier as an Asian woman, sparking race and gender controversies. Google paused human‑image generation features and co‑founder Sergey Brin acknowledged that "image generation absolutely messed up," attributing the issue to insufficient testing. These problems stemmed from over‑optimization intended to avoid bias that instead produced excessive overcorrection.
Experts are calling for action. The EBU urges the EU and national regulators to enforce rules on information integrity, digital services and media plurality. The "Facts In: Facts Out" campaign demands that AI companies ensure factual inputs produce factual outputs and that news is not distorted. Peter Archer, director of the BBC's generative AI project, emphasized, "People must trust what they see. Despite improvements, the problems remain serious." The study recommends developers strengthen guardrails, such as Anthropic's "Constitutional AI" training that prioritizes human well‑being. UC Berkeley professor Hany Farid warned, "Don't rely on social media or AI bots for trustworthy information — you're asking to be deceived."
The study's findings indicate that while AI chatbots are innovative, their risk of spreading misinformation in news cannot be ignored. Built on large language models (LLMs), chatbots are prone to hallucinations and outdated information, leading to widespread misleading content. Synthesizing multiple studies, users should always cross‑verify facts, and media and governments must promote independent monitoring. Otherwise, an increasingly chaotic information environment will amplify the "liar's dividend," undermine real evidence and further distort public understanding of conflicts, health issues and elections. Only through strict regulation and technical improvements can AI serve the truth rather than error.