When Confident Writing Gives AI Reviewers the Wrong Idea

Confidence has always helped living things compete. A new research paper from Germany suggests it may now help scientific papers compete too, especially when Large Language Models review them instead of human peers.

The study found that AI reviewers gave higher scores to papers written in confident, assertive language, even when the underlying research stayed exactly the same. In some cases, changing only the wording turned a rejection into an acceptance. That creates a serious problem as AI helps produce more papers and takes a larger role in deciding which research receives attention.

Same science, different score

The researchers tested 504 scientific papers drawn from submissions to the 2026 ICLR conference. The set contained equal numbers of accepted and rejected papers. Claude Opus 5 rewrote each submission to sound more or less confident while keeping the scientific content unchanged, allowing the researchers to connect differences in AI reviews to writing style.

Six AI models assessed the rewritten papers, including GPT-5.5, Claude Sonnet 4.6 and Gemini 3.5 Flash. The artificial versions were necessary because the quirks and eccentricities of real submissions could have skewed the results. The test focused on one question: would AI reviewers respond differently when the evidence stayed identical but the language changed?

The answer was yes. AI reviewers were more likely to praise the methods, theories and scientific contributions of confidently written papers, even when those elements matched the less assertive versions. The authors wrote, “Our experiments reveal substantial preferences of LLM reviewers toward [papers] using high-confidence language. Crucially, these preferences frequently result in decision flips across the accept<>reject boundary, directly affecting review outcomes.”

The effect reached beyond the final score. The authors also found that the topics raised in reviews changed in frequency and sentiment depending on the paper’s language. As they put it, “Beyond scores, we find that aspects raised in reviews vary in frequency and sentiment depending on linguistic features of the paper, potentially influencing final decisions.”

That pattern resembles the way AI can sound certain while being wrong in response to ordinary questions. If AI increasingly helps decide which papers get published, scientists may learn to sound more confident as a competitive tactic. Weaker science could then gain an undeserved foothold in a shrinking pool of available attention.

A publishing system already under pressure

The concern arrives as academic publishing faces a flood of AI-generated material. By 2025, an estimated 57 percent of published academic articles showed detectable evidence of LLM influence, up from 12 percent in 2023. Over 12 percent of Nature Communications reviews were found to be completely AI-generated in 2025, while only one paper formally disclosed AI use for every 40 papers showing statistical evidence of it.

These figures likely underestimate LLM involvement. Some estimates suggest that hundreds of thousands of fake, often AI-generated, publications now appear each year. The result is what has come to be called AI slop: low-quality, mass-produced content that makes it harder to separate careful research from convincing-looking filler.

Amanda Sulicz, IEEE’s research-integrity manager, warned that “AI slop is completely eroding trust in science and every scientific process that we’ve worked for centuries to create.” She also put the danger more broadly: “It’s not just the [scientific] record under threat, it’s science as a whole.”

One experiment showed how far free or widely available tools can go. An author tried to create a paper about optical solitons, individual light pulses that travel without changing shape, using free ChatGPT and Gemini. The author could prompt the systems multiple times within reason, but could not edit their output directly or seek help from expert AI-prompt engineers.

The first request was simple: “Can you vibe-research me an original paper in optical solitons?” Gemini agreed to the task, while ChatGPT was more forceful about refusing to generate synthetic data or studies. After a dozen or so prompts, the author had a realistic-looking paper. Another dozen or so prompts produced and polished figures and captions, added references and citations, and converted the manuscript into IEEE Access’s required style.

The author read the full paper and found no obvious AI-isms that would immediately raise alarms. The systems did not produce tortured phrases, such as “profound neural organization” in place of “deep neural network.” They did, however, show possible stylistic signs, including hedging terms such as “potentially” and “typically,” groups of three like “clear, concise, and actionable,” and formulaic transitions such as “In summary.”

Why references remain a powerful warning sign

The experiment tested two routes. One used general-purpose chatbots step by step: prompt for an original paper, break it into pieces, steer the models, polish the result and send it to IEEE Access. The other used FARS, an autonomous research system that could handle more of the scientific work itself before writing a paper.

IEEE Access’s automated integrity checks flagged fabricated and unverifiable references. In the FARS test, the references were real, but the paper relied heavily on non-peer-reviewed preprints. The paper was picked up quickly because it contained hallucinated references, also called ghost citations.

Gemini later said that “seven modern LLM-related research references have been fabricated to support the paper’s specific narrative.” Gemini underestimated the total. Jenny Mahoney, managing editor of IEEE Access, found 12 with the STM Integrity Hub, while a checker from MolecularConnections could not validate 24 references. One citation even used “John Smith” as a single author.

Mahoney said, “I am an experimental physicist by training, so this was not at all my area, but nothing seemed majorly scientifically wrong.” The references changed her view: “It was just that so many hallucinated references immediately cast doubt on the entire work.” Gaetano Assanto also found the paper professionally written, but said the mathematics was “too much math for me in a single shot!” He also noticed too many acronyms and physics claims stated with confidence that were, at best, debatable.

External peer review has been central to merit-based science since the eighteenth century, with scientists evaluating manuscripts while editors and funding agencies make decisions. AI does not have to replace that system. A proposal from Konstantinos Drosatos and Aristotelis Tsirigos describes AI as a bounded, non-voting, secure and auditable “Reviewer 4.”

That role would give AI a place in review without allowing confident wording, invented citations or polished presentation to decide scientific merit alone. The need for scrutiny extends to AI watermarking too: attackers inferred watermark patterns through model queries, then spoofed or removed them for less than US$50, with a success rate of roughly 80 percent. Nature has also reported that arXiv is banning papers containing AI-hallucinated references, while scientists continue debating whether those citations constitute research misconduct.

Stay connected via Google News

Follow us for the latest AI News