On October 22, 2025, Search Engine Journal evaluated 2,709 news answers generated by various AI assistants and reached an alarming conclusion: 45% of responses had "significant problems" and, more broadly, 81% contained some level of issue.
The study did not treat "problems" as a blanket indictment but broke them down into actionable dimensions: missing or opaque sources, inaccurate citations, factual fabrication (hallucination), temporal errors, and misreading of context. The article notes many seemingly authoritative answers break the chain of evidence — for example, cited "studies" cannot be found in original sources, or provided statistics do not match public reports.
Among the tested products, the report specifically named Google's Gemini as having the most pronounced issues in "source management." Researchers found Gemini's answers more frequently included assertions that could not be traced back to original sources, meaning readers attempting to verify information often cannot locate the primary material from the clues given.
The authors emphasize the importance of a cross‑market design: the research sampled multiple markets rather than a single language or region to reflect how AI assistants perform across different languages and news ecosystems. Results show the distribution of problems is not confined to a particular topic or country's news, but represents a broader risk, especially in breaking events, complex data interpretation, or reporting that requires multi‑party verification.
For editors and content platforms, this dataset presents clear challenges: when AI is used to generate news summaries, draft reporting materials, or directly answer readers' questions, model output alone is insufficient. The article recommends two additional safeguards: improve transparency when models cite sources, and establish human verification or automated provenance mechanisms to ensure every key claim can be traced to a reliable source.
Search Engine Journal's report also reminds readers and developers to pay attention to time stamps and context — some AIs failed to specify the time range or data cutoff they used, causing readers to misjudge the timeliness of information. In a speed‑focused news ecosystem, such ambiguity amplifies the risk of misinformation.
This cross‑market analysis highlights AI assistants' weaknesses in real‑world news scenarios: not all model outputs are suitable for direct reporting or citation, particularly where sourcing and verifiability are deficient. The authors call on developers, media organizations, and users to prioritize "verifiability" as a core metric when adopting AI for news tasks.