Deep Research is still more of a promise

The promise of deep research by AI tools like ChatGPT is simple: feed in a question, receive a nuanced, well-reasoned answer, complete with references and synthesis. And in some domains, it delivers admirably. When it comes to media trends, tech news, or cultural analysis, the tool works best. It sifts through the torrent of online commentary, news articles, blog posts, and social chatter to surface patterns and narratives. The real value here lies not just in the volume of data it processes, but in how the user frames the question. A clever, counterintuitive prompt can elicit insights that feel like genuine thought.

But the illusion shatters when the query turns academic. For scholarly literature reviews, this tool falters. It is not the fault of the software itself—there is no shortage of computational power or linguistic finesse. The limitation is upstream. Most academic journals sit behind expensive paywalls, historically inaccessible to large language models. The corpus they are trained on has lacked precisely the kind of data that matters most for rigorous research: peer-reviewed studies, meta-analyses, theoretical frameworks built over decades.

This, however, is beginning to change. In May 2024, Microsoft signed a $10 million deal with Taylor & Francis to license journal content for its Copilot AI. OpenAI, for its part, has secured a two-year agreement with The Associated Press and forged partnerships with European publishers like Axel Springer, Le Monde, and Prisa Media—giving ChatGPT a better grasp of reputable journalistic and scholarly content. Wiley joined the fray with a $23 million licensing deal to grant an unnamed AI developer access to its academic publishing portfolio. Even Elsevier, long a fortress of paywalled knowledge, is now channeling its scholarly data into AI collaborations.

These are significant moves. They mark a transition from aspiration to access. Once these agreements begin to reflect in AI performance, the quality of output will change markedly. A tool that can both identify the pivotal paper and incorporate it meaningfully into its reasoning would be a true research assistant—something closer to intellectual augmentation than just computational summarization.

It is still early days. Scite, for now, remains stronger at pointing users to the right academic sources, even if its analytical output lacks flair. ChatGPT and its peers, in contrast, excel at synthesis but stumble when the raw material is lacking. The ideal tool is still on the horizon.

There is an irony, nevertheless. AI, the most advanced information-processing technology ever built, has been running on the least rigorous parts of the internet. It has quoted tweets with ease but struggled to cite the peer-reviewed studies that ought to anchor serious inquiry. That is no longer a permanent condition. It is, increasingly, a transitional one.

The future of AI in research will be determined not solely by engineering breakthroughs, but by access negotiations. With major publishers now at the table, the landscape is poised for a shift. For the user, the best strategy remains what it has always been: ask sharp questions. But soon, the machine’s answers may finally rest on deeper knowledge.