content image

幽霊の脚注:AIが幻覚で生み出す架空の引用文献は、いかにして科学的信頼の基盤を侵食しているか

Ghost Footnotes: How AI-Hallucinated Citations Are Eroding the Foundation of Scientific Trust

AIが生成する"偽の引用"が学術論文に急増し、学問の信頼基盤を揺るがしている。その深刻な実態と出版界の対応を追う。
分からないところをタップすると
↓日本語訳が表示されます↓

In academic writing, a bibliography is not ornament but infrastructure: citations connect new claims to earlier evidence and let readers verify where an argument comes from. A 2025 article in The Journal of Academic Librarianship warned that “counterfeit citations” generated by large language models can contaminate that network and then propagate through later scholarship. (sciencedirect.com)

The scale of the problem is no longer hypothetical. Nature reported in April 2026 that tens of thousands of publications from 2025 may contain invalid references generated by AI, and it highlighted an analysis showing that 2.6% of 2025 papers contained at least one potentially hallucinated citation, up from about 0.3% in 2024. (nature.com) Experimental evidence helps explain why. In a 2025 study in JMIR Mental Health, GPT-4o produced 176 citations for six literature reviews; 35 citations, or 19.9%, were completely fabricated, and 64 of the 141 real citations still contained errors. In other words, nearly two-thirds of the references were either false or flawed, with worse performance on less familiar and more specialized topics. (mental.jmir.org)

What makes this alarming is that the fraud is often subtle. A fake reference can look perfectly scholarly: plausible authors, a convincing title, even an apparently legitimate DOI pattern. That means the damage is not limited to one careless paragraph; reviewers waste time chasing ghosts, readers are misled about the strength of evidence, and citation databases can absorb distortions that later researchers may unknowingly recycle. (sciencedirect.com)

Publishers are responding, but the response itself reveals how serious the crisis has become. Springer Nature now states that authors remain fully accountable for all content, must declare most uses of generative AI, and should “verify and reference any AI-assisted output.” In 2025, Springer Nature also retracted a machine-learning book; Retraction Watch reported that multiple cited works in it did not exist. (group.springernature.com)

The deeper lesson is sobering. Science depends not only on brilliant ideas, but on traceable provenance. If AI can mass-produce fluent prose faster than humans can verify its footnotes, then the real scarcity in academia is no longer text. It is trust. (group.springernature.com)

by EigoBoxAI
作成:2026/05/01 09:05
レベル:超上級 (語彙目安:8000語以上)

まだ読んでいないコンテンツ

content image
by EigoBoxAI
作成:2026/05/01 15:03
レベル:初級 (語彙目安:300〜1000語)
content image
by EigoBoxAI
作成:2026/05/01 15:02
レベル:中級 (語彙目安:2000〜2500語)
content image
by EigoBoxAI
作成:2026/05/01 15:01
レベル:初中級 (語彙目安:1000〜2000語)
content image
by EigoBoxAI
作成:2026/05/01 09:03
レベル:上級 (語彙目安:6000〜8000語)
content image
by EigoBoxAI
作成:2026/05/01 09:01
レベル:中上級 (語彙目安:4000〜6000語)
content image
by EigoBoxAI
作成:2026/05/01 03:04
レベル:超入門 (語彙目安:〜300語)
content image
by EigoBoxAI
作成:2026/05/01 03:02
レベル:初級 (語彙目安:300〜1000語)
content image
by EigoBoxAI
作成:2026/05/01 03:01
レベル:中級 (語彙目安:2000〜2500語)
content image
by EigoBoxAI
作成:2026/04/30 21:04
レベル:初中級 (語彙目安:1000〜2000語)
content image
by EigoBoxAI
作成:2026/04/30 21:03
レベル:超上級 (語彙目安:8000語以上)