ChatGPT, Claude, Gemini, and Grok Are All Bad at Crediting News Outlets, but ChatGPT Is the Worst (at Least in This Study)

Why It Matters

Without transparent attribution, AI‑generated news summaries erode journalistic revenue and public trust, prompting urgent policy and design interventions.

Key Takeaways

• AI models cover 74% of stories, cite 8% sources.
• ChatGPT attributes only 1% without explicit citation request.
• Claude cites 16% by default, 97% when asked.
• Paywalled and regional outlets receive minimal AI visibility.
• Explicit citation prompts raise attribution above 50% for all models.

Pulse Analysis

The recent AI News Audit from McGill’s Center for Media, Technology and Democracy highlights a paradox in large language models: they absorb vast quantities of Canadian journalism yet rarely disclose where specific facts originate. By testing both baseline queries and web‑search‑enabled prompts, researchers showed that while 74% of responses captured distinctive reporting, only 8% offered any source credit. ChatGPT, the most widely used model, fell dramatically short, attributing sources in just 1% of default answers, whereas Claude and Gemini performed modestly better under the same conditions.

These attribution gaps have concrete consequences for the news ecosystem. Media organisations rely on traffic and licensing fees that are undermined when AI delivers near‑complete summaries without linking back to the original outlet. The study also notes a bias toward freely accessible, national broadcasters—CBC, CTV and Global—while paywalled or regional papers such as the Toronto Star or Postmedia’s local titles receive scant AI visibility. This disparity amplifies existing revenue pressures and fuels legal actions, exemplified by the ongoing OpenAI copyright lawsuit in Canada. Transparency deficits risk eroding public confidence in both AI tools and the journalism they repurpose.

Technical fixes are within reach. The audit found that explicit citation requests push attribution rates above 50% for all models, with Claude reaching 97% when prompted. Designers can embed automatic source tagging or hyperlinking as default behavior, reducing reliance on user‑initiated prompts. Policymakers may consider standards that mandate citation metadata for AI‑generated content, balancing innovation with the need to protect the integrity and sustainability of news production. As AI continues to shape information consumption, clear attribution will become a cornerstone of responsible digital journalism.