Ever since generative AI has become all the rage, the media has been flooded with success stories: Breakthroughs in science, improved results in education, a more efficient health system, increased productivity in enterprise.
Some of these stories have later been retracted. But there is a multidimensional gap between the reported success and retraction: A gap in time, a discrepancy in attention and the influence both have over decisions: the “retraction gap”.
For the industry, there is a lot at stake. As is for many governments, not only because of financial but also political entanglement: Many have firmly tied their hope for economic growth to the success of AI, even justifying their reluctance to regulate.
The narrative of inevitability of AI in all aspects of society, as “revolutionary” new industrial revolution, demands success stories especially in areas that are perceived to need “new solutions”. Some of those success stories have later been questions, or outright found to be misleading or false. The reporting of those successes is not necessarily deliberate deception, but part of the natural media cycle.
What is important is the gap between reported success and retraction, which can measure years, and the discrepancy between the amount of publicity the success story has received vs the retraction can be significant. Often ignored is the fact that decisions have already been made based on alleged success that will not get reversed even if the underlying story is discredited.
While not unique to the AI industry, this phenomenon is more pronounced here because of the immense “fear of missing out” pressure.
First we should look at five examples selected from very diverse fields where the overly optimistic assessments of AI have been retracted, and why neither academia nor politics is immune from the issue.
One of the most prominent studies that has been recently retracted by a major scientific journal was a study that assessed “the effectiveness of ChatGPT in improving students’ learning performance”. Originally published in May 2025, it was retracted by Nature a year later.
With 549,000 accesses and 420 citations the paper received considerable attention, likely because it appeared to offer a solution in a field where people were especially concerned about the impact of AI.
The paper was interestingly specific in its assessment, solely focusing on ChatGPT as a tool in education:
“In summary, the meta-analysis results of this study confirm the positive impacts of ChatGPT on learning performance, learning perception, and higher-order thinking […] ChatGPT could be actively incorporated into STEM and related courses” (Wang, Fan 2025). The paper ends in a glowing endorsement and recommendations:
“(2) the broad use of ChatGPT at various grade levels and in different types of courses should be encouraged to support diverse learning needs; […] 5) ChatGPT should be flexibly integrated into teaching as an intelligent tutor, learning partner, and educational tool.” (Wang, Fan 2025)
The lack of critical distance should likely have raised eyebrows, especially because evaluating a chatbot represent a specific snapshot in time.
The journal decided to retract this article, in its own words, due to concerns relating to “discrepancies in the meta‑analysis”, sadly without giving specific details.“Taken together, the identified issues undermine the Editor’s confidence in the validity of the analysis and the conclusions drawn from it.” (Nature 2026)
The distance between publication, ravish reviews and retraction was about a year, but a significant year when it comes to adoption of AI in education.
Meanwhile, a large study by the OECD recently came to a very different conclusion: AI can be useful in education, but only if it is used in a way that challenges the student: “At first glance, the PISA data aren’t positive: Students who are using AI chatbots for specific tasks such as summarising texts get lower scores in science, on average, than those who don’t.” and concludes that there is little exceptional about AI: “Where technology enables or enhances that cognitive struggle, students will advance. Where technology short-circuits the productive struggle of learning, it will undercut students’ development.” (OECD 2026)
A paper published at MIT in 2024 bundled two interesting claims together: That “AI-assisted scientist discovered 44% more materials” while AI is causing job dissatisfaction for those scientists. We covered this paper and its retraction in details in the article “The Curious Case Of That Withdrawn MIT AI Paper” in 2025.
For journalists it was also about breaching the gap of praising our brilliant new AI future and concerns of their readers, spurred by expectations of a “white collar job wipe out”. The paper fit this narrative perfectly, at the right time. (Misaliged 2025)
The retraction notice was curiously reported in the Wall Street Journal, and for good reason: The Journal had prominently featured the paper, and most prominently the author in December 2024.
When asked about the potential benefits of generative AI, medical science is often cited as a prime example: The promise that AI will literally “cure cancer” is as much as a meme as it is a frequent argument by AI influencers.
Medical studies often take years to conduct, and more time to be verified independently. The study “Bridging the gap: explainable AI for autism diagnosis and parental support with TabPFNMix and SHAP”, published in Nature in 2025 appeared to report on promising results:
“The findings of this research reinforced the reliability of the proposed framework and its potential applicability in clinical settings, offering clinicians and caregivers a transparent and accurate AI-assisted diagnostic tool.” (Jiang 2025)
For once, not only did the retraction come only two months after publication, it also did not hold back:
“Post-publication investigation confirmed that the paper does not contain an adequate description of the framework it proposes, uses an AI-generated figure containing nonsense to demonstrate the framework, and does not include code underlying the model developed.” (Nature 2026)
The retraction here shines a light on the complexity of peer review in the age of AI. Even under traditional circumstances research data is hard to judge from the outside, but even more so if it involves claims related to generative AI.
Whenever a new technology arrives that targets enterprise and governments, the consultant companies play an essential role in its adoption.
In 2026, it emerged that a KPMG report on how AI is being used by businesses across the world was found to have exaggerated the adoption of the technology with bogus case studies that ironically appear to have been based on AI hallucinations. The report was found to have made numerous false claims, including about the adoption of AI at UBS, the UK’s National Health Service, Swiss Federal Railways and Transport for London (Bratton, Foley 2026).
Another report, published by EY (formerly known as Earnest & Young), had claimed in October 2025 that its AI-related revenue had grown 30 per cent in the previous year and 15,000 staff had worked on client projects “ranging from delivering enterprise-wide transformations to AI governance frameworks that help drive the responsible implementation of AI” (Foley 2026). The report turned out to have contained “included apparent AI hallucinations and fake footnotes”.
As researchers from GPTZero point out: “When the report includes fake information […] it can ‘poison the well’ by misleading future researchers, especially if the report is published by a well-known consulting firm and hosted on a high-traffic website.” (Foley 2026)
The sway consultant companies have over governments and enterprise can hardly be overstated. The studies will have been cited at business conferences and undoubtedly made their way into internal business presentation.
Consultants play another important role: They help companies to “outsource trust”: If it is a success, the credit will go to company management. If it goes wrong, or emerges to be based on false data, the blame can at least be put on the consultants. A mechanism that does not inspire a critical approach to consultant studies.
The involvement of American company Palantir in the British health system has been controversial from the start, with concerns ranging from international politics, the ideology of the company’s founders to data privacy concerns.
NHS England originally suggested that Palantir’s federated data platform has contributed to 110,000 additional operations being carried out. Those figures were widely cited in the media and left little doubt about the causation: Palantir allegedly helps to reduce the NHS waiting list.
Anyone familiar with the discourse around the NHS will know that the issue of waiting lists is a hotly debated one, often presented as the crucial data point to judge the effectiveness of the British health system. Extremely emotionally charged, it and has been a focus of the political fight between government and opposition for more than a decade.
Government ministers — and Palantir — have long promoted benefits from the Federated Data Platform (FDP), software provided by the US tech giant. A page on NHS England’s website titled: ‘NHS Federated Data Platform uptake and benefits’ includes a number of claims, such as “111,589 — Total number of additional patients undergoing procedures in theatres”. (Foxglove 2026)
It has recently been suggested that the software’s effectiveness is questionable, something the NHS has somewhat quietly admitted:
“This is an observational before/after comparison showing changes to this metric following the introduction of the NHS FDP Inpatient theatre module. We cannot therefore draw conclusions about cause and effect as other variables have not been controlled for.” (Foxglove 2026)
In fact, it looks like a classical case of mixing up causation and correlation:
“A number of improvements in NHS performance attributed to Palantir have been driven by just a handful of hospitals, according to data that shows some trusts have actually delivered fewer operations since adopting the technology.” […] “Where trusts have reported significant improvements, it is not clear how much this reflects FDP adoption as the figures are not adjusted for other factors, such as trusts investing in more staff or weekend clinics.” (Financial Times 2025)
The NHS/Palantir example shows how the burden of proof has been shifted to the critics once a system has been introduced, even if questions emerge and claims are “quietly” retracted.
The paper “Analysis of Retracted Publications on Artificial Intelligence: Trends, Ethical Concerns, and Scientific Integrity” published in November 2025 examined “a total of 764 retracted AI-related papers”.
The paper concludes that “for scientific literature integrity, academic institutions, publishers, and researchers should stress transparency, ethics, and rigorous post-publication inspection.” (Kocyigit et al. 2025). The researchers also found that “the median duration to retraction, 510 days, reveals a gap in detecting and fixing inaccurate papers, allowing potentially misleading findings to remain in the scientific environment.”
510 days is a long time in public memory, and in the current environment will have impact way beyond the academic environment, enough to have already resulted in real-life decisions.
As we recently showed in “AI Everywhere, Regulation Nowhere”, an overly optimistic stance towards a new technology is not unique to industry or science, and can have direct impact on decision-making and people’s lives to the extent where a governments decision not to regulate AI heavily relies on those optimist assessments.
While academia is fairly good at keeping track of retracted papers and scientific claims and has developed databases to do so (such as Retraction Watch), this can not be said for policy papers or industry studies.
We need to develop a methodology to measure the impact on public perception, possibly offering tools for decision makers in government and industry.
Success stories find there way from academia into mainstream media, into blogs and to social media, and nowadays also into the training material of AI chatbots. In times of a hype we have to be more attentive than usual.
Academic studies play a special part in the media hyper cycle. They can build the basis of a claim despite most journalists having read or understood them, nor are readers expected to read the papers themselves, stressing the importance of peer review. A mysterious “research has shown” is regularly sufficient to bolster for mainstream media claims.
And as long as the hype cycle continues, we should take success stories not just with a grain, but with a whole shovel of salt.
RETRACTED Paper: The effect of ChatGPT on students’ learning performance, learning perception, and higher-order thinking: insights from a meta-analysis, Nature
RETRACTED Paper: Bridging the gap: explainable AI for autism diagnosis and parental support with TabPFNMix and SHAP. Nature
Elizabeth Bratton, Stephen Foley(2026) KPMG report contained AI hallucinations on benefits of . . . AI. Financial Times
Stephen Foley (2026) EY retracts study after researchers discover AI hallucinations. Financial Times
Laura Hughes, Amy Borrett (2026). Claims of Palantir improvement in NHS driven by a handful of hospitals. Financial Times
Foxglove (2026). UK Government and Palantir must correct public record after NHS admits published data does not back up FDP benefit claims
PISA 2025 Results (Volume I), OECD [PDF]
Kocyigit BF, Okyay RA, Seiil B, Qumar AB, Sumbul HE. Analysis of Retracted Publications on Artificial Intelligence: Trends, Ethical Concerns, and Scientific Integrity. J Korean Med Sci. 2025 Nov 17;40(44):e280. doi: 10.3346/jkms.2025.40.e280. PMID: 41250649; PMCID: PMC12624210
