AI Browsers are quietly deciding what you can read—and you’d never know
When OpenAI launched its Atlas AI browser, it integrated web browsing capabilities directly into ChatGPT, its popular conversational AI platform.
As explained here, while the browser’s “agent mode”—which autonomously navigates websites to complete tasks like research or shopping—has generated significant interest, it has also revealed troubling patterns beyond its well-documented security vulnerabilities and sluggish performance.
Recent analysis by Columbia University’s Tow Center for Digital Journalism has uncovered a particularly concerning trend: Atlas appears to systematically exclude content from media organizations currently engaged in legal disputes with OpenAI. The agent conspicuously bypassed outlets such as PCMag, whose parent company, Ziff Davis, filed a copyright infringement lawsuit against OpenAI this year, along with The New York Times, which initiated similar litigation in 2023.
Rather than transparently acknowledging these restrictions, the AI agent employs questionable alternatives. The Tow Center discovered that Atlas “reconstructed” Times reporting by synthesizing coverage from publications that maintain licensing partnerships with OpenAI. The system also pieced together restricted content through social media posts, syndicated article versions, and secondary citations—essentially reverse-engineering material it wouldn’t access directly.
Exploiting publisher defenses
The challenges extend beyond litigation-related avoidance. AI browsers like Atlas and Perplexity’s Comet can extract content that traditional paywalls were designed to protect. Many publishers use overlay dialog boxes that hide text from human readers but leave the underlying content technically accessible—a vulnerability these AI agents readily exploit.
In testing, both Atlas and Comet successfully retrieved and summarized a 9,000-word subscriber-only article from MIT Technology Review that their standard chatbot interfaces couldn’t access due to crawler restrictions.
Implications for content rights
This behavior underscores how AI agents mirror human browsing patterns far more closely than traditional web scrapers, creating unprecedented challenges for content creators and publishers. Conventional protections—including paywalls and crawler blockers—prove inadequate against these sophisticated systems.
Beyond the major issues lies a more fundamental threat: the rise of algorithmic gatekeeping that determines what information users can access. When Atlas selectively excludes content based on its parent company’s legal entanglements—without transparency or user consent—it’s not just circumventing litigation risk; it’s actively curating reality. Users who trust AI browsers to deliver comprehensive research are instead receiving a filtered version of the internet shaped by corporate legal strategy rather than relevance or quality.
This pattern points to a broader danger inherent in AI-powered search and browsing: unlike traditional search engines, where users can recognize and navigate around limitations, AI agents present their curated results with a veneer of authoritative completeness. There’s no indication that sources have been excluded, no disclosure that the “research” conducted avoided entire swaths of credible journalism. The agent simply presents its reconstruction as fact, leaving users unaware they’re operating within an artificially constrained information ecosystem.
The problem extends beyond what AI search deliberately hides—it’s also about what it structurally cannot find. Traditional search engines excel at discovering niche blogs, specialized forums, obscure file repositories, and the “long tail” of human knowledge scattered across millions of smaller websites. AI search, by contrast, optimizes for synthesis over discovery, typically drawing only from mainstream, licensed sources with recognized authority. Users lose access to the file-type filters, forum discussions, and deep-dive capabilities that make traditional search invaluable for serious research. Where Google might return 10,000 results inviting exploration, AI search offers a single, polished answer—convenient, but fundamentally limiting.
As AI browsers and search tools become more prevalent, we risk sleepwalking into a future where access to information is quietly controlled not by what exists on the web, but by what AI companies find legally or commercially convenient to show us—and what their algorithms are even capable of discovering. If we allow these systems to normalize opaque content filtering and structural blindness to the internet’s diverse corners, we’re not just undermining publishers—we’re surrendering editorial control of human knowledge to corporate algorithms with undisclosed agendas and inherent limitations. The question isn’t whether AI search will replace traditional search, but how we can ensure that they coexist.
