92%. That's the share of AI chatbot responses that contained zero source attribution, even when the answers clearly drew on news reporting. A new study from McGill University's Center for Media, Technology and Democracy tested ChatGPT, Claude, Gemini, and Grok on their knowledge of Canadian current events and found that all four models freely reproduce journalism without telling you where it came from.
The research team, led by Taylor Owen and Aengus Bridgman, ran two tracks. In the first, they tested 2,267 Canadian news stories with web search disabled. The models demonstrated knowledge of the underlying reporting in 74% of responses, but nearly all of those answers came without any mention of the original outlet.
The second track tested 140 specific articles with web search turned on, which is how most people actually use these tools. The results split into two telling numbers: how much of the original reporting the model reproduced (coverage), and how often it named the source (attribution).
The Attribution Gap by Model
- Claude: 72% coverage, 16% attribution
- Gemini: 81% coverage, 6% attribution
- Grok: 59% coverage, 7% attribution
- ChatGPT: 54% coverage, 1% attribution
Claude attributed sources roughly 16 times more often than ChatGPT, though 16% is still a failing grade by any journalism standard. Gemini reproduced the most original material of any model while crediting sources just 6% of the time.
When researchers explicitly asked the models to cite their sources, the numbers jumped dramatically. Claude named the outlet 97% of the time, Gemini 95%, ChatGPT 86%, and Grok 74%. This gap between default and prompted behavior suggests the models have the information available but are not designed to surface it unprompted.
Free Outlets Get the Visibility, Paywalled Papers Get Nothing
The study also found a visibility bias: freely accessible outlets like CBC, CTV, and Global News received disproportionate mentions, while paywalled publications and regional newspapers were largely invisible. This is a problem for local journalism. If a regional paper breaks a story and an AI chatbot summarizes it without attribution, the paper gets no traffic, no credit, and no subscriber conversion from its own work.
What This Means for People Using These Tools
If you use AI chatbots for research or staying current on news, the takeaway is straightforward: the answers you get are built on journalism, but the tools are not telling you that. You are reading summaries of reporting that someone paid to produce, stripped of any link back to the original.
For anyone who cares about the accuracy of what these models tell them, this is also a verification problem. Without knowing where a claim originated, you cannot check it. Adding "cite your sources" to your prompts is a partial fix, but it should not be the user's job to ask for basic attribution that the product should provide by default.
Anthropic's Claude performs least badly here, but "least bad" is doing a lot of heavy lifting when your best score is 16%.