A New Threat to Science: AI Making Up References in Research Papers

It’s not a new phenomenon. AI has been “hallucinating” – making things up – for some time, in areas ranging from law to medicine. But a new example of this affliction has recently emerged: fabricating citations in scientific research papers.

A recent article in the prestigious medical journal The Lancet reported the first systematic evaluation of reference authenticity. The audit of 2.5 million biomedical papers spanning 3 years showed that fabricated references are deeply embedded in the peer-reviewed literature. And the rate of fabrication is accelerating, as demonstrated in the figure below.

The integrity of references is fundamental to the integrity of scientific literature. Each reference implicitly asserts that a verifiable source exists and bolsters the claims being made. When references cite non-existent studies, readers, reviewers and policy makers are unable to evaluate the evidence.

Fabricated references can arise from paper mill activity and intentional deception, as well as the indiscriminate use of AI writing tools. Large language models (LLMs) are capable of generating plausible sounding but fictitious references. Previous studies have estimated that between 30% and 69% of LLM-generated references in biomedical research are fabricated, say The Lancet authors. The references are often correctly formatted, attributed to real researchers and include credible publication dates, making them difficult to detect by conventional peer review.

The authors developed an automated reference verification system by scanning PubMed Central’s Open Access subset from January 1, 2023, to February 18, 2026. That effort examined 2,471,758 papers and 125,615,773 structured references. Of the 125.6 million references, 77% carried a PubMed identifier and were verified; the remaining 23%, predominantly non-indexed references (from sources such as websites and books), were excluded.

Among the verified references, which came from 2810 papers, the authors identified 4046 fabricated citations. In 2023, approximately one in 2828 papers contained at least one fabricated reference. By 2025, this had gone up to one in 458, and in the first 7 weeks of 2026, a staggering one in 277 papers had at least one fabricated reference. As shown in the figure above, the fabrication rate increased by more than 10 times from 2023 to 2026.

It was found that 91% of affected papers contained just one or two fictitious references; 9% contained three or more. The sharp inflection in mid-2024 coincides with the expected publication lag following widespread LLM adoption, although increased paper mill activity and changes in journal indexing practices might also have contributed. LLMs became widely available in late 2022 and 2023.

The article’s authors qualify their findings by stating that the exclusion of 23% of references with no PubMed identifier could bias their estimates. Furthermore, PubMed Central’s Open Access does not cover all biomedical literature, and the early 2026 data span only 7 weeks. And, they say, their approach identifies the problem but not the underlying cause.

Nonetheless, the authors do have some recommendations for publishers. These include integrating automated reference verification into submission workflows before peer review; having indexing services add integrity metadata to article records so that users can assess the reliability of references; and retroactively screening existing publications and issuing corrections or retractions when fabricated references compromise a paper’s conclusions.

On the broader question of why such fabrication is occurring, the answer lies in the training of LLMs. A 2025 article in Science magazine argues that AIs hallucinate because they’re trained to fake answers when they don’t know how to answer a question. A team from OpenAI and the Georgia Institute of Technology is more specific:

They learn to bluff because their performance is ranked using standardized benchmarks that reward confident guesses and penalize honest uncertainty.

Although AIs could be modified to admit that “I don’t know,” this could simply turn off users, who would then seek an alternative LLM. A better solution, say several computer scientists, would be to rework benchmarks to penalize an AI for guessing incorrectly. Current benchmarks don’t penalize incorrect guesses any more than nonanswers.

The need for AI to clean up its act is especially important in medicine. In deciding on patient treatment, medical professionals are often influenced by clinical guidelines in the research literature. Doctors and practitioners have no way of knowing whether the evidence they are relying on actually exists.

In a comment accompanying The Lancet article, two pediatrics professors place the onus on manuscript authors themselves, saying that:

Authors must take responsibility and be held accountable for the entire content of a manuscript, including the references.  

The professors believe that a published manuscript in which a fabricated reference is detected should be retracted. While this may seem harsh, such a move would go a long way toward meeting the AI challenge and restoring the integrity of scientific research.

Next: RIP RCP8.5: A Hopeful Sign That the Attack on Science Is Letting Up

Sea Level Rise Dominated by Subsidence, Not Global Warming

Of all the topics I’ve covered in this blog, none have attracted as many readers as sea level rise (see for example here, here and here). And for good reason, since rising seas threaten the livelihood of tens of millions of coastal dwellers around the world.

What is not generally known is that the main culprit is land subsidence resulting from groundwater extraction and urban development that causes more sea level rise than any human-induced global warming. The importance of subsidence has been brought to light by two recent research papers: a Dutch study finding that current model-based estimates of sea levels are much too low, and a study of river delta subsidence by an international team of environmental scientists and others.

Several different processes can affect vertical motion of land. Long-term glacial rebound after melting of the last ice age’s heavy ice sheets is causing land to rise in high northern latitudes. Yet in many regions, the ground is sinking because of sediment settling and aquifer compaction caused by human activities, especially groundwater depletion resulting from rapid urbanization and population growth.

Land subsidence artificially amplifies any local rise in sea level, which can be measured by either tide gauges or satellite altimetry. Tide gauges measure the height of the sea relative to the land to which the gauge is attached, the so-called RSL (Relative Sea Level) metric. Satellite observations of absolute sea level measure the height of the sea, which is the distance of its surface to the center of the earth.

Subsidence of the land independently of sea level makes satellite-measured sea levels higher than tide gauge RSLs. That’s why the average rate of rise from 1900 to 2020 determined by tide gauges was only 1.75 mm (0.07 inches) per year, while NASA’s satellite measurements say the rate was more like 3.3 mm (0.13 inches) per year from 1993 to 2024.

But neither tide gauges nor satellites measure sea levels at all points on every coastline in the world, which is why so-called geoid models, rather than actual sea level measurements, are used to make estimates. A geoid is an equipotential surface that approximates mean sea level based on gravity and the earth’s rotation.

However, because the actual height of the sea surface is not determined only by the gravity and rotation of the earth, but also by other factors such as ocean currents, winds, tides and seawater temperature, the time-average sea-surface height can deviate strongly (up to several meters) from a geoid. The discrepancy between geoid and actual sea level has led to widespread underestimation of coastal sea level rise.

This occurs predominantly in parts of the world where the most rapid land subsidence is happening – in southeastern and eastern Asia, for example. The Dutch study examined 40 river deltas worldwide subject to severe subsidence such as the Vietnamese Mekong Delta, which is one of the most densely populated coastal landscapes exposed to rising sea levels worldwide.

The figure below shows how much land subsidence has already occurred in the 40 deltas. Yellow and red circles denote subsidence, the circle size indicating the percentage of the delta area subsiding faster than satellite-measured sea level rise, which averages about 4 mm (0.16 inches) per year globally. 

In 13 of the deltas (Brahmani, Brantas, Ceyhan, Chao Phraya, Ciliwung, Godavari, Mahanadi, Mekong, Nile, Po, Red, Vistula and Yellow River), the subsidence rate exceeds 4 mm (0.16 inches) per year. Among these, the Chao Phraya (Thailand), Brantas (Indonesia) and Yellow River (China) deltas show an average sinking rate of more than twice the current rate of global sea level rise.

Overall, more than 90% of the delta area is affected by subsidence in 19 of the 40 deltas studied, while more than 50% of the delta area is subsiding in 38 of the 40. Coastal cities in the deltas are experiencing subsidence rates even greater than the deltas themselves. In Jakarta (Ciliwung), for example, parts of the city are subsiding at alarming rates surpassing 30 mm (1.2 inches) per year, dwarfing average global sea level rise by almost an order of magnitude. 

The effect of land subsidence in river deltas on local populations is demonstrated in the next figure. The upper left panel (a) compares rates of local sea level rise and subsidence; current average subsidence rates exceed sea level rise in 18 of the 40 deltas, including the Mekong and Nile. The lower left panel (b) makes the same comparison, but only for populations living within 1 meter of sea level. Panel c on the right is a bar plot of subsidence vs sea level rise for 30 of the 40 deltas.

Of the various possible contributions to subsidence, groundwater storage has the strongest relative influence in 10 of the 40 deltas. The other drivers are sediment flux and urban expansion.

Next: A New Threat to Science: AI Making Up References in Research Papers

AI Proves Its Mettle in Reconstruction of Antarctic Temperatures

In a 2025 post, I described the largely unsuccessful attempt of an AI to spearhead research in climate science. Now, however, another AI appears to have succeeded in the more technical task of accurately reconstructing surface air temperatures across Antarctica – something that standard temperature datasets have been unable to achieve. The work is reported in a recent paper by a team of Chinese researchers.

The figure below illustrates the geographical distribution of available observational data in Antarctica from 1979 to 2023. The data comes from a number of manned and automatic weather stations, together with meteorological observations over the ocean collected from ships and buoys. As can be seen, the majority of observations are in coastal or near-coastal regions, precluding full spatial coverage of the continent.

To overcome this shortfall, the limited station observations have traditionally been interpolated using various reanalyses. But it’s difficult for reanalysis datasets to capture complex spatial patterns, say the researchers, and such datasets often contain significant uncertainties. Moreover, deriving surface air temperatures from reanalysis datasets depends in part on model simulations rather than actual instrumental measurements.

In the light of these limitations to an accurate reconstruction of Antarctic temperatures, the Chinese research team has applied deep learning methods. This approach has already been utilized successfully to reconstruct Arctic temperatures.

Antarctic temperatures were reconstructed using daily surface air temperature data from the various sources depicted in the figure above. Daily average temperatures were calculated from observations made at 3-hour intervals for some sources, 1-hour intervals for others. Training data for the deep learning model was provided by surface air temperatures from the three reanalysis datasets that showed the best agreement with observed temperatures.

The training data, which covered the period from 1979 to 2005, totaled 29,211 daily temperature samples. Using reanalysis data from different time periods, such as 1995 to 2021, as the training set had little impact on the temperature reconstruction. Validation of the training data and testing of the reconstructed temperature set employed reanalysis data from 2006 to 2012 and 2013 to 2018, respectively.

Testing of the reconstructed Antarctic temperatures was conducted for three specific days in 2015: January 1, July 1 and November 1. For these days, the reconstructed temperatures were found to be highly correlated with their reanalysis counterparts, with spatial correlation coefficients >0.99. The researchers say this correlation shows that the trained deep learning model is capable of accurately reproducing Antarctic surface air temperatures, even with the limited observational data available.

Just how different the reconstructed surface temperatures are from global observational temperature datasets for Antarctica is depicted in the next figure. The figure shows linear trends in annual Antarctic surface air temperatures from 1979 to 2023, measured in degrees Celsius per decade. The datasets are: (a) this reconstruction; (b) Berkeley Earth; (c) ERA reanalysis; (d) NOAAGlobalTemp5.1; (e) GISTEMPv4; and (f) HadCRUT5.

You can see that none of the standard datasets exhibit the pronounced cooling trend in East Antarctica in (a), something that was inferred earlier from the ERA reanalysis dataset by a different group of Chinese researchers. Nevertheless, all datasets show warming in the Antarctic Peninsula (on the left of the continent in the maps above).

Differing from the annual trends is the pattern for the summer months (November to April) only, presented in the figure below for the period from 1989 to 2022. Although the cooling trend still dominates in East Antarctica, warming is no longer prominent in the Peninsula but is found in West Antarctica and the southern portion of East Antarctica.

East Antarctica actually experienced a summer heat wave in 2022, when the temperature soared to -10.1 degrees Celsius (13.8 degrees Fahrenheit) at the Concordia weather station, located at the 4 o’clock position from the South Pole, on March 18. This balmy reading was the highest recorded hourly temperature at that station since its establishment in 1996, and 20 degrees Celsius (36 degrees Fahrenheit) above the previous March record high there. Remarkably, the temperature remained above that record for three consecutive days, including nighttime.

But Antarctica is nothing if not unpredictable. Despite the 2022 heat wave, the mercury dropped to -51.2 degrees Celsius (-60.2 degrees Fahrenheit) on January 31, 2023. This marked the lowest January temperature recorded anywhere in Antarctica since the first meteorological observations there in 1956. Two days earlier on January 29, the nearby Vostok station, about 400 km (250) miles closer to the South Pole, registered a low temperature of -48.7 degrees Celsius (-55.7 degrees Fahrenheit), that location’s lowest January temperature since 1957.

Such swings from record highs to record lows remain a puzzle, but the present reconstruction at least helps to characterize long-term trends.

Next: The Atlantic “Cold Blob” – Cause for Alarm or Just a Curiosity?

AI Tries Its Hand at Climate Science

Much-ballyhooed AI (artificial intelligence), which is rapidly making inroads in many areas, has found its way into scientific publishing with the recent appearance of a peer-reviewed study featuring AI Grok 3 beta as the lead author. The 2025 paper, titled “A Critical Reassessment of the Anthropogenic CO₂-Global Warming Hypothesis,” disputes the narrative that global warming is largely driven by human emissions of CO₂.

Grok 3 beta, who wrote its own press release, spearheaded the research but needed critical guidance from four human coauthors. While the AI on its own was able to identify some of the relevant papers in the scientific literature, it fell short by missing a number of others and had to rely on its human colleagues to fill the gaps – and even they overlooked the work of one important researcher.

The study essentially reviews the conclusions of the IPCC (Intergovernmental Panel on Climate Change) regarding the global carbon cycle, computer climate models, and solar variability. The paper’s subtitle is “Empirical Evidence Contradicts IPCC Models and Solar Forcing Assumptions.”

On the carbon cycle, shown in the figure below, the study challenges the claim made in the IPCC’s AR6 (Sixth Assessment Report) that the effective residence time of CO₂ in the atmosphere is more than 100 years. The claim has been questioned by several authors, including Greek civil engineer Demetris Koutsoyiannis who, in a 2024 paper, estimated a residence time as short as 3.5 to 4 years.

The Koutsoyiannis paper simply divides the atmospheric CO₂ level, which was 416.4 ppm (parts per million) or 887 GtC (gigatonnes of carbon) in 2020, by the total flow of CO₂ into the atmosphere of approximately 230 GtC per year. However, this analysis only accounts for the so-called fast carbon cycle – the rapid movement of CO₂ between living organisms, the atmosphere and the oceans – but ignores the slow carbon cycle, in which CO₂ is incorporated into long-lived vegetation such as tree bark, or the deep ocean, or limestone and other rock.

Notable among other authors who share the beliefs of Koutsoyiannis are atmospheric scientists Hermann Harde and the late Murray Salby. Their work (see, for example, here) is acknowledged by Grok 3 beta, but only because the AI was reprimanded by its coauthors for omitting any mention of Harde or Salby papers from its first draft.

Another major omission the first time around was the extensive work of U.S. astrophysicist Willie Soon, one of the coauthors, on solar contributions to global warming. I’ll touch on Soon’s research below.

But more egregious than any of these omissions is Grok 3 beta’s failure to include any papers of physicist David Andrews, who has been highly critical of Harde and Salby.

Andrews (see here, here and here) distinguishes between the fast and slow carbon cycles, and points out that large, natural, two-way exchanges of carbon occur between the atmosphere, the oceans and land biomass. The fluxes each way exceed the one-way flux into the atmosphere of anthropogenic CO₂ from fossil fuels. Andrews also discusses errors in the interpretation of radiocarbon data in the Harde and Salby papers.

As for Koutsoyiannis, I discussed in a 2023 post how an earlier study of his is unable to explain 88% of the increase in atmospheric CO₂ during global warming of approximately 1 degree Celsius (1.8 degrees Fahrenheit) since 1880, in terms of CO₂ outgassing from the oceans.

Turning to climate models, Grok 3 beta gets it essentially right. As I’ve explained in numerous previous posts, many climate models run too hot, greatly exaggerating future global warming; only a small number come close to actual measurements. And a whole ensemble of models is unable to reproduce observed sea surface temperature trends in the Pacific and Southern Oceans since 1979.

Grok 3 beta correctly states that “model runs consistently fail to replicate observed temperature trajectories and sea ice extent trends, exhibiting correlations (R²) near zero when compared to unadjusted records.” On the subject of adjustments to raw temperature data, the paper also duly notes that such adjustments are contentious, especially the use of homogenization techniques which can introduce systematic errors into temperature data.

Finally, the AI rightfully draws attention to the IPCC’s questionable reliance on a single reconstruction of past solar output or TSI (total solar irradiance), to support its narrative of overwhelmingly human-caused global warming with essentially no contribution from the sun. As I described in another 2023 post, Soon and a team of coauthors have shown that a number of alternative TSI reconstructions, in which the sun plays a much larger role, can explain observed warming just as well as the IPCC’s chosen reconstruction.

So how well has Grok 3 beta mastered climate science? In this case, its paper is a mixed bag worthy of barely a passing grade.

Next: Climate Model Track Record Improves Slightly: Paused Arctic Sea Ice Loss Predicted Correctly