2026年5月13日 / 美国东部时间下午5:09 / 哥伦比亚广播公司新闻（CBS News）
作者：梅根·塞鲁洛（Megan Cerullo）
最新研究结果显示，人工智能正在凭空捏造不存在的医学研究参考文献。
《柳叶刀》（The Lancet）上的一篇文章指出，近期一项审计在数百万份生物医学论文中发现，超过4000篇论文引用了根本不存在的研究。该研究的主要作者、哥伦比亚大学护理学院副教授马克西姆·托帕兹（Maxim Topaz）表示，这类伪造的参考文献可能会破坏医疗专业人员在诊疗过程中所依赖的临床指南。
研究人员在近期发表于《柳叶刀》的文章中称，对数百万份生物医学论文的审计共发现了超过4000篇引用虚假研究的文献。
马克西姆·托帕兹在接受哥伦比亚广播公司新闻采访时表示，伪造的参考文献十分危险，因为它们会影响临床指南的制定，而临床指南正是基于医疗专业人员在诊疗中所遵循的公开研究成果。
他说：“当这些虚假参考文献被纳入学术文献后，最终会出现在临床指南中，而医生正是据此决定如何为患者提供治疗。”“你的医生可能会基于从未存在过的研究来制定治疗方案。”
日益严重的问题
托帕兹及其团队发现的错误无一得到纠正或撤回，仍可能在影响患者护理，这一点同样令人不安，他说道。
托帕兹补充道：“发表在医学文献中的虚假参考文献占比正在不断上升。”他指出，过去三年间，这类错误引用的数量增长了12倍。这些伪造的参考文献涉及近3000篇学术论文。
托帕兹的个人经历促使他开展这项调查。他告诉哥伦比亚广播公司新闻，自己曾使用一款人工智能应用来润色一篇科学论文，结果该应用插入了一条伪造的参考文献。这条假引用在经过多层同行评审后，才被一名眼光敏锐的编辑发现。
“我当时感到非常羞愧，因为我已经研究人工智能15年了，所以如果这种事会发生在我身上，那任何人都可能遇到，”他说。
托帕兹解释道，当作者提出一项事实陈述并要求人工智能提供引用依据时，就会出现这类问题。“在某些情况下，人工智能会在无意间插入这些虚假引用，”他说。“你本以为这些事实是准确的，但如果它们的支撑依据是伪造的参考文献，你就无从得知这些‘事实’是否真实。”
托帕兹表示，在某些情况下，人工智能工具会在引用真实作者的同时，编造研究内容并将其归到该作者名下。还有些时候，参考文献完全是凭空捏造的。
“这还只是冰山一角，”他说，并指出其他领域的研究也可能存在同样的问题。
与此同时，人工智能生成的虚假科学参考文献“看起来可能完全真实”，托帕兹补充道，他强调研究人员必须严格核查自己的研究成果。
本文编辑：阿兰·谢尔特（Alain Sherter）
AI is fabricating citations in biomedical studies, researchers find
May 13, 2026 / 5:09 PM EDT / CBS News
By Megan Cerullo
Artificial intelligence is fabricating references to medical research that does not exist, according to recent findings.
A recent audit found that, among millions of biomedical papers, more than 4,000 contained citations to non-existent research, according to an article in The Lancet. Such fabricated citations can undermine the clinical guidelines that health care professionals rely on to provide care, said Maxim Topaz, an associate professor at the Columbia School of Nursing and the study’s lead author.
An audit of millions of biomedical papers found more than 4,000 citations to bogus studies, the researchers said in a recent article published in The Lancet.
Fabricated citations are dangerous because they influence clinical guidelines, which are based on public research that health care professionals follow in providing care, Maxim Topaz, an associate professor at the Columbia School of Nursing and the study’s lead author, told CBS News.
“When those fake references are making it into the literature, they will end up in those guidelines, and that’s how doctors decide how to provide care for you,” he said. “Your doctor could be making decisions around treatment based on studies that never existed.”
Growing problem
Also troubling is that none of the mistakes Topaz and his team identified have been corrected or retracted, and could still be influencing patient care, he said.
“The rate of fake references showing up in published medical literature is growing,” Topaz added, noting that the number of such erroneous citations has grown 12-fold over the last three years. The fabricated references spanned nearly 3,000 academic papers.
Topaz’s own experience spurred him to investigate the issue. An AI app he was using to help polish one of his own scientific papers inserted a fake citation, he told CBS News. It then slipped through several layers of peer reviews before one sharp-eyed editor caught the phony reference.
“I was mortified, because I’ve been studying AI for the past 15 years, so if it can happen to me, it can happen to anyone,” he said.
Such mishaps arise when an author asserts a statement of fact and asks AI for a citation, Topaz explained. “In some cases, AI would slip those in, inadvertently,” he said. “You would hope the facts are accurate, but if they are supported by fabricated citations, you don’t know if the ‘facts’ are accurate.”
In some cases, an AI tool will also cite a real author while inventing research and attributing it to that person. Other times, citations were completely fabricated, Topaz said.
“This is just the tip of the iceberg,” he said, noting that research across other fields could also be subject to the same issues.
Meanwhile, faux AI-generated scientific citations can “look perfectly real,” Topaz added, who emphasized the importance of researchers rigorously fact-checking their work.
Edited by Alain Sherter
发表回复