Many AI systems are now capable of generating detailed reports complete with citations, which simplifies the scholarship process significantly. This trend is gaining traction across various institutions, from universities to research labs. It’s hard not to wonder, though, what might be lost amid these conveniences.
Imagine conducting a search for, say, 5 to 30 minutes, gathering the findings into a structured report with footnotes. Google’s equivalent might require up to 80 search queries for a similar task. Meanwhile, these AI systems operate in the background while you focus on other things. In this setup, a lead agent can spawn multiple subagents that tackle different parts of a question. You’d be surprised to know that this arrangement has reportedly outperformed single-agent models by 90.2% during internal evaluations.
To respond to a single prompt, for instance, the AI might utilize 21 search queries and an enormous number of inference tokens—it’s something like 193,947. It then condenses this information into a readable format in just four minutes. But here’s the catch: what the system decides to treat as knowledge is not necessarily all-encompassing.
This quest for knowledge isn’t a new idea. Vannevar Bush envisioned it back in 1945, advocating for a more intelligent relationship between thinkers and human knowledge. Douglas Engelbart later conceptualized tools to enhance problem-solving through better processing and collaboration. Fast forward to today, and we have a system that integrates researchers as if they are just issuing prompts, leaving a sense of disconnect from the traditional research process where the researcher is actively engaged.
So, compression seems to be a cornerstone of this technology. Searches are narrowed down and ranked, producing a final report that, while comprehensive, reduces information into prose. We see a trend here: anything deemed worth condensing is accepted as knowledge. Anthropic’s system emphasizes that “the essence of search is compression.” This notion might give us insight into how we’ll perceive information in the near future.
However, mistakes do occur. OpenAI recognizes that its systems can fabricate facts, draw faulty conclusions, and misjudge reliable sources versus unreliable rumors. In initial testing, early agents were found to lean heavily toward search-optimized content rather than trustworthy sources. Google also warns about the risks of encountering misleading information from dubious websites.
While there exists a benchmark of short-answer tasks with varying accuracy levels—OpenAI’s Deep Research coming in at 51.5%, for example—it raises questions. Short answers are easier to evaluate, but it’s unclear how well this translates to complex, freeform tasks in real life. Just because a machine passes a software test doesn’t mean it’s ready for practical applications.
One significant challenge lies in the fact that valuable knowledge often isn’t available through standard searches. Many critical insights remain tacit, local, or embedded in practices that scholars, labs, or courts have internalized over the years. Automated AI research can handle portions like summarizing methodologies or comparing papers, but it doesn’t always grasp the unspoken context—nuances that researchers might hesitate to articulate.
There are even reports of AI-generated papers being passed for evaluation at conferences, but concerns about misrepresentations and reproducibility linger. The pace of synthesis exceeds that of sound judgment, suggesting a disconnection between what AI can produce and what constitutes valid scientific inquiry.
As automated AI tools become integral to research, they could disrupt the existing business models of knowledge-sharing platforms. A lawsuit recently highlighted this issue, suggesting potential problems with copyright and distribution as research agents become primary conduits of information.
Interestingly, a Pew survey indicated that only 16% of American workers believe AI helps them in their jobs, and many who use chatbots find them more useful for speeding up tasks than enhancing quality. Furthermore, while 58% reported seeing AI-generated summaries on Google, just 13% had used AI tools recently. Essentially, AI-driven research is becoming a backdrop in our lives more than an explicit tool—people may not even recognize its presence in research activities.
In the end, we’re seeing the industrialization of certain forms of intellectual labor—like extensive searching and drafting—in ways that could alter how we think and what web publishers must adapt to. However, this shift has yet to provide a full substitute for the exploration that fosters creativity: the unexpected discoveries in conversations, the debates in the halls, and that serendipitous reading that happens in a moment of quiet. For now, while the system is adept at compressing the world, what truly gets lost in that process remains an open question.
