content image

幻の引用文献:AI が生成した偽の参照文献はいかにして医学文献を静かに蝕んでいるか

Phantom Citations: How AI-Generated Fake References Are Quietly Poisoning Medical Literature

AIが生成した"実在しない参考文献"が医学論文に紛れ込み、2025年以降急増している。見た目は本物そっくりでも中身は空——医療の信頼を揺るがす「幽霊引用」の実態とは。
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↓

Imagine opening a medical paper, checking a reference, and finding that the “study” never existed. That is the problem of AI-generated phantom citations: references that look scholarly but point to nothing real. In a Lancet correspondence published on May 7, 2026, researchers audited 2.5 million open-access biomedical papers and found 4,046 fabricated references across 2,810 papers. The rate stayed low in 2023, then climbed sharply in 2025 and early 2026. In one especially striking case, a single paper contained 18 fake references out of 30. What makes this pollution dangerous is its invisibility: the paper still looks formal, technical, and trustworthy at first glance. (nature.com)

This did not begin inside journals. Chatbots were already showing the same weakness. A 2023 study in Scientific Reports found that 55% of citations produced by ChatGPT-3.5 were fabricated; even GPT-4 invented 18% of its references. In a 2024 study in Clinics, ChatGPT-4 generated 260 references over repeated searches, and nearly three-quarters were hallucinations. On one day, the system even admitted that the references were fictional — but on the other days, it gave no such warning. These false citations are hard to catch because they often borrow real author names, real journal titles, and believable medical topics. (nature.com)

Medicine is especially vulnerable because its writing depends on chains of evidence. Review articles are read as shortcuts to the literature, and the new audit found that reviews had higher fabrication rates than other paper types. Even when AI tools can search the web, the problem does not disappear. A 2025 Nature Communications study found that 50% to 90% of LLM responses to medical questions were not fully supported by the sources they cited; even GPT-4o with web search still failed to fully support nearly half of its responses. That means the danger is not only fake papers, but also real-looking citations that do not actually prove the claim being made. Some publishers have started screening submissions automatically, but for now the simplest defense is still human skepticism: every citation must be checked before it enters the medical record. (cidrap.umn.edu)

by EigoBoxAI
作成:2026/05/13 03:04
レベル:中上級 (語彙目安:4000〜6000語)

まだ読んでいないコンテンツ

content image
by EigoBoxAI
作成:2026/05/13 15:03
レベル:中上級 (語彙目安:4000〜6000語)
content image
by EigoBoxAI
作成:2026/05/13 15:02
レベル:超入門 (語彙目安:〜300語)
content image
by EigoBoxAI
作成:2026/05/13 15:01
レベル:初級 (語彙目安:300〜1000語)
content image
by EigoBoxAI
作成:2026/05/13 09:05
レベル:中級 (語彙目安:2000〜2500語)
content image
by EigoBoxAI
作成:2026/05/13 09:03
レベル:初中級 (語彙目安:1000〜2000語)
content image
by EigoBoxAI
作成:2026/05/13 09:02
レベル:超上級 (語彙目安:8000語以上)
content image
by EigoBoxAI
作成:2026/05/13 03:05
レベル:上級 (語彙目安:6000〜8000語)
content image
by EigoBoxAI
作成:2026/05/13 03:02
レベル:超入門 (語彙目安:〜300語)
content image
by EigoBoxAI
作成:2026/05/12 21:04
レベル:初級 (語彙目安:300〜1000語)
content image
by EigoBoxAI
作成:2026/05/12 21:02
レベル:中級 (語彙目安:2000〜2500語)