Battle of the AIs: Can Bing and Bard Beat ChatGPT at Research?

The most practical takeaway from the comparison is that no single AI tool dominates research writing end-to-end: ChatGPT (with GPT-4) tends to win when the task is turning a paper’s content into polished language, while Microsoft Bing is strongest when the job is finding starting points—especially references—and navigating the early stages of a literature search.

All three systems were tested on common research workflows: locating papers, summarizing PDFs, and transforming academic material into outputs such as bullet-point summaries, press releases, and blog posts. For paper-finding, the baseline reality is that ChatGPT lacked internet access during the tests, so it couldn’t reliably surface the newest literature. That limitation showed up immediately when asked for the latest papers on organic photovoltaic materials. Bing, with internet access, produced usable links and references, and the results improved compared with earlier attempts using tools without strong retrieval. Still, even Bing’s paper suggestions sometimes required verification and follow-up expansion using external tools like Connected Papers, Litmaps, or ResearchRabbit.

When the workflow shifted from retrieval to reading and summarizing, ChatGPT’s performance stood out. The transcript describes a “text splitter” approach to feed paper text into ChatGPT, after which it produced structured five-bullet summaries that included more of the paper’s concrete details (and felt more trustworthy to the tester) than Bard’s higher-level, figure-light summaries. Bing also summarized, but it didn’t consistently follow the requested format (for example, not delivering exactly five bullets) and tended to be less detailed.

The same pattern repeated for rewriting tasks. Asked to convert a paper into a press release, ChatGPT produced a more faithful, publication-ready structure and captured the paper’s essence more accurately. Bard generated a quick draft but introduced factual errors—such as attributing findings to the wrong journal—while Bing’s output was less aligned with strict press-release conventions, though it still offered subheadings and a usable narrative.

For blog-style writing, ChatGPT again matched the brief more closely, producing language appropriate for a science publication audience. Bard’s output skewed toward a more formal, exploratory tone that wouldn’t easily pass editorial standards, and Bing’s responses leaned toward generic “how to write a science blog” guidance rather than generating an actual publishable draft.

Finally, when starting from scratch—like drafting an introduction for a literature review on transparent electrode materials—Bing looked best for providing seed references and field orientation, even though the transcript cautions that reference accuracy still needs checking. Bard also performed reasonably on mapping the topic (e.g., noting indium tin oxide and alternatives like carbon-based materials), but Bing’s reference scaffolding made it the better launchpad.

The conclusion is a division of labor: use Bing for the hardest parts of research discovery (references, initial exploration, and PDF interaction), and use ChatGPT for language-heavy transformation once the source material is in hand. Bard is described as comparatively weak for research-specific nuance and referencing reliability in these tests.