I Used DeepSeek R1 for Research – Here's What Happened!

DeepSeek R1 lands as a free, open-source AI model that can meaningfully assist academic work—especially for drafting structure and finding peer-reviewed papers—though it still lags behind top competitors on figure understanding and polished writing quality. In practical tests, it produced a usable literature-review outline for carbon nanotubes, generated a draft introduction section with citations, and returned 32 “recent” peer-reviewed papers (2023–2025) that the user could click through to publisher pages. That combination—starting academic writing and supplying clickable sources—makes it a credible research sidekick, even if it isn’t the strongest option for high-stakes, publication-ready output.

For literature review tasks, DeepSeek R1 performed best as a scaffold rather than a replacement for established writing tools. It generated a structured outline quickly, then followed up by drafting an “anxiety-free” introduction section. The draft was judged as a solid starting point (roughly a C+ / “pass”), with citations included—such as a Nature paper on mechanical strength—yet it didn’t match the depth and layout guidance produced by tools like ChatGPT in the user’s experience. The model’s “thinking” process was visible in real time, which the user found informative, but the resulting prose still required human judgment and refinement.

Where DeepSeek R1 showed clear utility was reference discovery. Using an internet-enabled search prompt for “recent peer-reviewed papers on nanoparticles for OPV devices,” it returned 32 results and linked out to major outlets and databases (including ScienceDirect, Nature, PubMed, and a Hong Kong Polytechnic University page). The user flagged one mismatch: the first item’s displayed title didn’t align with what opened on click, suggesting occasional citation/title confusion. Even so, the workflow—generate a list of recent, relevant papers and then click through to verify—was considered genuinely helpful, particularly if run periodically alongside other research tools.

DeepSeek R1’s weaker area was visual reasoning from complex figures. When given a demanding set of microscopy-related visuals (SEM and AFM images, plus a height profile) for non-planar vs planar electrode comparisons involving transparent electrodes, single carbon nanotubes, and silver nanowires, the model struggled to correctly “mesh” the components. It identified parts of the figure but missed how the elements connected, and the user repeatedly found ChatGPT better at extracting figure text and translating it into accurate, integrated conclusions. DeepSeek improved somewhat with additional prompting, but the gap remained.

On higher-level synthesis—turning multiple figures into a compelling research story—the model produced a plausible ordering and narrative arc, though again it was judged inferior to ChatGPT’s stronger ability to integrate figure relationships. For editing, it offered improvement suggestions to an abstract, but the overall verdict was that other mainstream tools (Claude and ChatGPT) can do similar work without the need to switch specifically to DeepSeek.

Overall, DeepSeek R1 earns its keep for free academic assistance: outlines, draft starters, and paper-finding workflows. The tradeoff is reliability and polish—especially for figure interpretation—where established competitors still lead for publication-grade writing and analysis.