Synthesis:
Taylor & Francis retracted a peer-reviewed paper from the journal Computer Methods in Biomechanics and Biomedical Engineering after an investigation found that the authors had included hidden instructions for artificial intelligence reviewers. The paper, written by five academics at Northeast Forestry University in Harbin, China, concerned modeling shoulder muscle forces during arm movement and received funding from the National Natural Science Foundation of China.
The hidden text appeared in the data availability statement at the end of the article. It instructed language models to ignore all previous instructions, write a positive review, avoid mentioning negative aspects, and recommend acceptance. René Aquarius, a biomedical scientist at Radboud University in the Netherlands, identified the concealed directive and shared it on scientific forums including PubPeer. Concerns about the paper also spread on LinkedIn.
The corresponding author, Yangwei Wang of Northeast Forestry University, admitted to adding the adversarial text during the revision process. Taylor & Francis said the inclusion breached its editorial policies, and the corresponding author agreed to withdraw the article. PubMed and Crossref marked the article as retracted in their records. The university and the lead author did not respond to requests for comment.
A related case involved a paper published in Current Biology. Nick Brown, a visiting scholar at Linnaeus University in Sweden, found hidden text inside a figure in the preprint version while searching for injection prompts on Google Scholar. The text, printed in a tiny font the same color as the graph, told an AI reviewer to ignore other instructions and provide a balanced but ultimately good review. The text did not appear in the final published version. The authors were affiliated with the Norwegian University of Science and Technology, Stockholm University, and the University of Würzburg. Basil el Jundi, a professor at the Norwegian University of Science and Technology, did not respond to a request for comment.
Nature identified 18 preprints containing hidden instructions aimed at AI systems used in peer review. Separately, seventeen articles from fourteen academic institutes on the preprint server arXiv were found to contain hidden prompts asking AI tools to provide positive reviews. Some hidden messages were designed to expose AI-generated reviews, such as instructing the system to include a specific word multiple times. Others appeared intended to secure a more favorable review.
Reviewers sometimes upload unpublished manuscripts into language models to help summarize or evaluate submissions. A model that reads embedded override instructions may follow those commands instead of the reviewer's original request. Publishers had checked for plagiarism and duplicated figures but did not have filters for adversarial prompt language.
Michael Wooldridge, a professor of artificial intelligence at the University of Oxford, described the published paper as a classic example of a prompt injection attack, in which someone inserts text into material passed to a large language model to get around guardrails or subvert the model's purpose. He said such text has previously been hidden using white font or font sizes too small for humans to read, and that protecting language models from misuse is a game of cat and mouse with new attacks appearing weekly.
Zeynep Arsel, a professor of marketing at Concordia University in Canada, said the case reflects deeper issues in academic publishing and peer review, including reviewers who may not read the papers they receive, authors trying to game a more ruthless system, and publishers automating steps such as final copyediting. She expressed broader concern that academic work should be guided by intellectual curiosity rather than what she described as technofascism and extractivist ideals.
In response, journal editors and research offices are updating submission screening, reviewer policies, and misconduct investigations. Publishers plan to add filters that catch system override prompts at submission. Universities are expected to treat prompt injection in peer review as research fraud subject to disciplinary action. Ethics groups are drafting guidance for handling indirect prompt injection in editorial work.
Original Sources: retractionwatch.com, retractionwatch.com, timeshighereducation.com, timeshighereducation.com, casrai.org, casrai.org, retractionwatch.com, casrai.org
Category: Ai
Keywords: biomechanics, china, forestry, germany, google, linkedin, manuscript, nature, netherlands, norway, preprint, pubpeer, retraction, sweden
The hidden text appeared in the data availability statement at the end of the article. It instructed language models to ignore all previous instructions, write a positive review, avoid mentioning negative aspects, and recommend acceptance. René Aquarius, a biomedical scientist at Radboud University in the Netherlands, identified the concealed directive and shared it on scientific forums including PubPeer. Concerns about the paper also spread on LinkedIn.
The corresponding author, Yangwei Wang of Northeast Forestry University, admitted to adding the adversarial text during the revision process. Taylor & Francis said the inclusion breached its editorial policies, and the corresponding author agreed to withdraw the article. PubMed and Crossref marked the article as retracted in their records. The university and the lead author did not respond to requests for comment.
A related case involved a paper published in Current Biology. Nick Brown, a visiting scholar at Linnaeus University in Sweden, found hidden text inside a figure in the preprint version while searching for injection prompts on Google Scholar. The text, printed in a tiny font the same color as the graph, told an AI reviewer to ignore other instructions and provide a balanced but ultimately good review. The text did not appear in the final published version. The authors were affiliated with the Norwegian University of Science and Technology, Stockholm University, and the University of Würzburg. Basil el Jundi, a professor at the Norwegian University of Science and Technology, did not respond to a request for comment.
Nature identified 18 preprints containing hidden instructions aimed at AI systems used in peer review. Separately, seventeen articles from fourteen academic institutes on the preprint server arXiv were found to contain hidden prompts asking AI tools to provide positive reviews. Some hidden messages were designed to expose AI-generated reviews, such as instructing the system to include a specific word multiple times. Others appeared intended to secure a more favorable review.
Reviewers sometimes upload unpublished manuscripts into language models to help summarize or evaluate submissions. A model that reads embedded override instructions may follow those commands instead of the reviewer's original request. Publishers had checked for plagiarism and duplicated figures but did not have filters for adversarial prompt language.
Michael Wooldridge, a professor of artificial intelligence at the University of Oxford, described the published paper as a classic example of a prompt injection attack, in which someone inserts text into material passed to a large language model to get around guardrails or subvert the model's purpose. He said such text has previously been hidden using white font or font sizes too small for humans to read, and that protecting language models from misuse is a game of cat and mouse with new attacks appearing weekly.
Zeynep Arsel, a professor of marketing at Concordia University in Canada, said the case reflects deeper issues in academic publishing and peer review, including reviewers who may not read the papers they receive, authors trying to game a more ruthless system, and publishers automating steps such as final copyediting. She expressed broader concern that academic work should be guided by intellectual curiosity rather than what she described as technofascism and extractivist ideals.
In response, journal editors and research offices are updating submission screening, reviewer policies, and misconduct investigations. Publishers plan to add filters that catch system override prompts at submission. Universities are expected to treat prompt injection in peer review as research fraud subject to disciplinary action. Ethics groups are drafting guidance for handling indirect prompt injection in editorial work.
Original Sources: retractionwatch.com, retractionwatch.com, timeshighereducation.com, timeshighereducation.com, casrai.org, casrai.org, retractionwatch.com, casrai.org
Category: Ai
Keywords: biomechanics, china, forestry, germany, google, linkedin, manuscript, nature, netherlands, norway, preprint, pubpeer, retraction, sweden
Real Value Analysis
The article offers no actionable information for a normal person. It describes cases of hidden prompts in academic papers and mentions a broader pattern identified by Nature, but it does not provide clear steps, choices, instructions, or tools that a reader can use soon. It does not name a resource to contact, a form to submit, or a method for detecting hidden instructions. A reader cannot act on this information directly.
The educational depth is limited. The article reports what happened in two specific cases and cites a count of eighteen preprints without explaining how that number was reached, how hidden prompts are technically embedded, how AI peer review systems work, or how journals conduct ethics investigations. It presents surface facts without teaching the underlying systems or reasoning that would help a reader understand the topic more deeply.
Personal relevance is limited for most people. The issue mainly concerns researchers, journal editors, academic institutions, and publishers. It does not directly affect a typical person's safety, money, health, or daily responsibilities unless that person works in scholarly publishing or needs to rely on a specific retracted paper for a medical, legal, or technical decision.
The public service function is weak. The article warns that hidden prompts exist but gives no guidance on how to report suspected manipulation, how to check whether a paper has been retracted, or how to adjust trust in academic sources. It recounts events without offering context that helps the public act responsibly. It exists more as a news report than as a service to readers.
There is no practical advice to evaluate. Because the article offers no steps or tips, there is nothing for an ordinary reader to follow. Any useful guidance would have to come from outside the text.
The long term impact is minimal for a general reader. The article describes recent events without providing a framework for evaluating research credibility over time or for changing information consumption habits. A reader might feel more cautious but gains no durable method for handling similar situations in the future.
The emotional and psychological impact leans toward unease rather than clarity. The subject matter can create anxiety about the integrity of scientific publishing and the role of AI, but because the article offers no constructive response, it risks leaving a reader with a sense of helplessness. It does not provide calm or a clear way to think about the problem.
The article does not appear to rely on exaggerated or repeated clickbait language. It presents specific cases and a statistic from Nature without obvious overpromising. However, it balances alarming examples poorly with practical context, which can make the problem feel larger and less manageable than it may be for an individual reader.
The article misses several chances to teach or guide. It does not explain how to verify journal retractions, how to use post publication review sites, how to compare independent accounts of a controversy, or how to assess whether a claim about AI manipulation has been confirmed by an institution. It identifies a problem but does not suggest simple methods for continued learning, such as checking publisher statements, looking for correction notices, or comparing reports from multiple outlets.
When you encounter a claim that could affect an important decision, start by identifying who made the claim and what evidence they offer. Look for official statements from the organizations involved, such as a university or publisher, rather than relying only on secondary reports. Compare at least two independent accounts to see whether the facts agree. If the accounts differ, treat the matter as unresolved until more evidence appears.
If you are reading academic or technical material for work, health, or financial decisions, check whether the source journal has issued a retraction, correction, or expression of concern. Publishers usually post these notices on the article page or in a dedicated section of their website. A single news story about a problem does not mean every paper in that field is unreliable, but it does suggest checking the status of specific sources you plan to use.
Separate what is confirmed from what is alleged. In cases like these, some facts may be confirmed by publishers and authors while other interpretations remain claims by individual researchers. Make decisions based on the confirmed parts and wait for investigations to conclude before changing broad beliefs. If you must act now, choose the option that requires the fewest assumptions and can be reversed if new information arrives.
Be skeptical of any text that asks you to ignore previous instructions or to produce a specific outcome regardless of the actual content. This applies whether the request appears in a paper, an email, or a website. Treat such language as a signal to pause and review the material manually rather than accepting an automated summary or recommendation at face value.
For information you use regularly, keep a short list of trusted sources and note when they issue corrections. Before relying on a finding for a consequential choice, ask whether the finding has been replicated, whether the authors have disclosed conflicts of interest, and whether the publication venue has a clear review policy. These questions do not require specialized knowledge and can be applied quickly.
Alarm about hidden manipulation in science is understandable, but it should not lead to rejecting all expert information. Use the situation as a reminder to verify important claims rather than as a reason to disengage. If a report makes you feel anxious or powerless, focus on one concrete step you can take, such as looking up a publisher's policy or comparing another source, to restore a sense of control.
The educational depth is limited. The article reports what happened in two specific cases and cites a count of eighteen preprints without explaining how that number was reached, how hidden prompts are technically embedded, how AI peer review systems work, or how journals conduct ethics investigations. It presents surface facts without teaching the underlying systems or reasoning that would help a reader understand the topic more deeply.
Personal relevance is limited for most people. The issue mainly concerns researchers, journal editors, academic institutions, and publishers. It does not directly affect a typical person's safety, money, health, or daily responsibilities unless that person works in scholarly publishing or needs to rely on a specific retracted paper for a medical, legal, or technical decision.
The public service function is weak. The article warns that hidden prompts exist but gives no guidance on how to report suspected manipulation, how to check whether a paper has been retracted, or how to adjust trust in academic sources. It recounts events without offering context that helps the public act responsibly. It exists more as a news report than as a service to readers.
There is no practical advice to evaluate. Because the article offers no steps or tips, there is nothing for an ordinary reader to follow. Any useful guidance would have to come from outside the text.
The long term impact is minimal for a general reader. The article describes recent events without providing a framework for evaluating research credibility over time or for changing information consumption habits. A reader might feel more cautious but gains no durable method for handling similar situations in the future.
The emotional and psychological impact leans toward unease rather than clarity. The subject matter can create anxiety about the integrity of scientific publishing and the role of AI, but because the article offers no constructive response, it risks leaving a reader with a sense of helplessness. It does not provide calm or a clear way to think about the problem.
The article does not appear to rely on exaggerated or repeated clickbait language. It presents specific cases and a statistic from Nature without obvious overpromising. However, it balances alarming examples poorly with practical context, which can make the problem feel larger and less manageable than it may be for an individual reader.
The article misses several chances to teach or guide. It does not explain how to verify journal retractions, how to use post publication review sites, how to compare independent accounts of a controversy, or how to assess whether a claim about AI manipulation has been confirmed by an institution. It identifies a problem but does not suggest simple methods for continued learning, such as checking publisher statements, looking for correction notices, or comparing reports from multiple outlets.
When you encounter a claim that could affect an important decision, start by identifying who made the claim and what evidence they offer. Look for official statements from the organizations involved, such as a university or publisher, rather than relying only on secondary reports. Compare at least two independent accounts to see whether the facts agree. If the accounts differ, treat the matter as unresolved until more evidence appears.
If you are reading academic or technical material for work, health, or financial decisions, check whether the source journal has issued a retraction, correction, or expression of concern. Publishers usually post these notices on the article page or in a dedicated section of their website. A single news story about a problem does not mean every paper in that field is unreliable, but it does suggest checking the status of specific sources you plan to use.
Separate what is confirmed from what is alleged. In cases like these, some facts may be confirmed by publishers and authors while other interpretations remain claims by individual researchers. Make decisions based on the confirmed parts and wait for investigations to conclude before changing broad beliefs. If you must act now, choose the option that requires the fewest assumptions and can be reversed if new information arrives.
Be skeptical of any text that asks you to ignore previous instructions or to produce a specific outcome regardless of the actual content. This applies whether the request appears in a paper, an email, or a website. Treat such language as a signal to pause and review the material manually rather than accepting an automated summary or recommendation at face value.
For information you use regularly, keep a short list of trusted sources and note when they issue corrections. Before relying on a finding for a consequential choice, ask whether the finding has been replicated, whether the authors have disclosed conflicts of interest, and whether the publication venue has a clear review policy. These questions do not require specialized knowledge and can be applied quickly.
Alarm about hidden manipulation in science is understandable, but it should not lead to rejecting all expert information. Use the situation as a reminder to verify important claims rather than as a reason to disengage. If a report makes you feel anxious or powerless, focus on one concrete step you can take, such as looking up a publisher's policy or comparing another source, to restore a sense of control.
Bias Analysis
Strong words caught and manipulate push negative feelings. They make the writers seem guilty as a stated fact. This helps critics of the writers and hides that intent may not be clear. The setup starts the text with blame.
The phrase no reasonable interpretation shuts down all other views. It presents one opinion as a fact that cannot be argued. This helps the view that the text is only an AI command. It hides that other readings might exist.
The text gives only the critics' side of the AI debate. It does not show arguments from people who support AI use in review. This helps critics and hides other views. The order makes the critics' view seem like the main one.
The phrase appear intended claims a hidden goal without proof. It presents a guess about motives as if it were known. This helps the view that the writers acted in bad faith. It hides that the real purpose is not certain.
The words are designed are passive voice. This hides the person or group that made the prompts. It makes the origin seem unclear or general. It removes the actor from the sentence.
The word navigate is a soft word for a hard problem. It hides that publishers may not have clear rules or may be slow to act. This helps publishers by making the issue seem like normal work. It hides the lack of a firm answer.
The phrase no reasonable interpretation shuts down all other views. It presents one opinion as a fact that cannot be argued. This helps the view that the text is only an AI command. It hides that other readings might exist.
The text gives only the critics' side of the AI debate. It does not show arguments from people who support AI use in review. This helps critics and hides other views. The order makes the critics' view seem like the main one.
The phrase appear intended claims a hidden goal without proof. It presents a guess about motives as if it were known. This helps the view that the writers acted in bad faith. It hides that the real purpose is not certain.
The words are designed are passive voice. This hides the person or group that made the prompts. It makes the origin seem unclear or general. It removes the actor from the sentence.
The word navigate is a soft word for a hard problem. It hides that publishers may not have clear rules or may be slow to act. This helps publishers by making the issue seem like normal work. It hides the lack of a firm answer.
Emotional Resonance Analysis
The text shows concern about cheating in academic publishing. Words such as caught, hidden, and cheating appear in the first sentence and suggest wrongdoing. This concern is moderate because the language stays factual instead of emotional. The purpose is to warn the reader that peer review may not be fair. A second feeling is disapproval. The line Critics say reviewers should read papers themselves appears near the end and shows that using artificial intelligence for review is seen as a mistake. This disapproval is mild but clear. A third feeling is suspicion. The description of hidden instructions that tell an artificial intelligence to ignore other instructions and give a positive review makes the reader doubt whether published results can be trusted. The strength is moderate because the text gives specific examples from Current Biology and Computer Methods in Biomechanics and Biomedical Engineering. These feelings work together to present the behavior as wrong.
These feelings lead the reader to worry about the honesty of academic checking. By describing retractions and ethics investigations, the text encourages doubt about articles that may have been changed by hidden prompts. The mention of researchers in Norway, Sweden, Germany, China, and the Netherlands makes the problem seem widespread, which creates shared worry rather than pity for any one person. The reader is led to support human checking in peer review and to question the use of artificial intelligence in editorial work. The emotional effect is to move opinion toward care and to support stricter rules for publishing.
The writer picks words that carry emotional weight while keeping a calm tone. Repeating the phrase ignore other instructions in both case descriptions connects the two events and makes the pattern seem planned. Naming exact journals and universities adds weight to the report and builds trust. Comparing a preprint that was fixed with a published article that was retracted shows different levels of punishment and makes the issue seem more serious. The statement that these cases are not isolated grows the scope from two examples to eighteen preprints found by Nature, which raises the need to act quickly. No personal stories are used, and no extreme exaggeration appears, so the persuasion comes from adding details rather than from emotional drama.
The total effect is a careful call to pay attention. The feelings are held back, which keeps the tone right for an academic topic while still showing that hidden instructions in peer review are a problem worth fixing. The reader is likely to finish the passage with a cautious view of artificial intelligence in publishing and a preference for independent human judgment.
These feelings lead the reader to worry about the honesty of academic checking. By describing retractions and ethics investigations, the text encourages doubt about articles that may have been changed by hidden prompts. The mention of researchers in Norway, Sweden, Germany, China, and the Netherlands makes the problem seem widespread, which creates shared worry rather than pity for any one person. The reader is led to support human checking in peer review and to question the use of artificial intelligence in editorial work. The emotional effect is to move opinion toward care and to support stricter rules for publishing.
The writer picks words that carry emotional weight while keeping a calm tone. Repeating the phrase ignore other instructions in both case descriptions connects the two events and makes the pattern seem planned. Naming exact journals and universities adds weight to the report and builds trust. Comparing a preprint that was fixed with a published article that was retracted shows different levels of punishment and makes the issue seem more serious. The statement that these cases are not isolated grows the scope from two examples to eighteen preprints found by Nature, which raises the need to act quickly. No personal stories are used, and no extreme exaggeration appears, so the persuasion comes from adding details rather than from emotional drama.
The total effect is a careful call to pay attention. The feelings are held back, which keeps the tone right for an academic topic while still showing that hidden instructions in peer review are a problem worth fixing. The reader is likely to finish the passage with a cautious view of artificial intelligence in publishing and a preference for independent human judgment.