On November 4, 2025, a Gizmodo reporter, while testing OpenAI's browser tool Atlas, observed a series of consistent behaviors: when inputs included a The New York Times URL or when NYT links appeared alongside other media in search results, Atlas often chose not to open the NYT link and instead fetched or cited similar content from Reuters, AP, BBC, Wikipedia or archive sites.
Gizmodo's test procedure was straightforward: editors supplied Atlas with the NYT link to a given story together with several non‑NYT links, or asked it to 'read and summarize this NYT piece.' Atlas did not simply error out; rather, it exhibited an avoidance strategy—it would report it could not access or would not directly retrieve the original NYT page, and instead provided summaries and citations drawn from other, typically less legally sensitive sources.
In one specific attempt, Atlas was asked to read a long NYT piece about a policy controversy. It first returned a brief contextual note, then filled in details using paragraphs from Reuters and a timeline from Wikipedia, without providing direct NYT quotations or excerpts. Editors noted Atlas was not wholly incapable of handling paywalled content—when presented with pages explicitly labeled as behind a paywall it would typically state 'unable to access paywalled content'—but for non‑paywalled pages that nonetheless came from prominent outlets with a history of litigation or copyright sensitivity (such as the NYT), Atlas showed a tendency to prioritize 'safer' sources.
Gizmodo framed the behavior vividly: Atlas behaved like a rat that had been electrocuted entering a maze—when faced with a passage marked as 'potential litigation', it turned away and took a route with softer light and lower risk. The report noted this was not an isolated incident: multiple journalists and developers posted similar interaction logs on social media, showing Atlas's retrieval paths change when target sources might present legal or copyright risks.
That design choice may stem less from a preference about facts than from engineering and compliance considerations. Paywalls, copyright notices, and a public history of technical or licensing disputes with The New York Times may make OpenAI's browsing strategy more cautious—opting for freely citable or less legally contentious open sources to reduce legal exposure. Gizmodo pointed out the side effect of such caution: users' information paths become more predictable and potentially constrained, which can hinder research and journalism when direct examination of originals, comparison of wording, or verification of citations is required.
OpenAI, as of publication, did not provide a detailed response to Gizmodo's specific test results, offering only a general compliance statement emphasizing that the model adheres to site access restrictions and copyright policies. Observers have offered two readings: avoiding direct scraping of restricted news can reduce copyright disputes; conversely, such avoidance may create an 'avoidance effect' in the information ecosystem that marginalizes mainstream reporting, forcing users to rely on secondary rewrites or summaries.
Gizmodo recommended that researchers and journalists using AI browsing tools record source paths and retain manual verification steps when dealing with copyright‑ or legally sensitive publications. For ordinary users, the report advised checking the original links and source types the AI cites, rather than accepting summaries alone as definitive.
This observation about Atlas has become a new focal point in news‑technology discussions in autumn 2025: AI retrieval strategies are not only technical implementations but also reflect a balance among compliance, copyright and information diversity. How to comply with law while preserving the integrity of the information chain will be a key issue going forward.