Strengths and Weaknesses of Artificial Intelligence in Exploring Asbestos History and Regulations Across Countries

Abstract

Asbestos, consisting of six natural mineral fibrous silicate phases, was widely utilized in industrial development during the 20th century and has left a global legacy of health, environmental, and regulatory challenges. Its remarkable properties (e.g., heat resistance, sound absorption, and tensile strength) made it a useful material in numerous applications. However, scientific research revealed its serious health risks in the early 1900s, with growing evidence during the 1960s, and nowadays its role in the development of different diseases (e.g., respiratory diseases, such as lung cancer, mesothelioma, and asbestosis) is well defined. Mapping this complex history requires integrating heterogeneous and often inconsistent information from nearly 200 countries. In this study, we tested the use of generative artificial intelligence (AI) tools as exploratory and comparative instruments to support the collection of asbestos-related data worldwide. Using Google Gemini (version 2.5 flash) and OpenAI ChatGPT (GPT-4-turbo variant), we gathered historical, medical, and regulatory information and then systematically verified and contextualized it with expert analysis. This dual approach allowed us to assess both the global asbestos situation and the reliability, advantages, and limitations of AI-assisted research. Our results highlight how AI can accelerate data collection and provide useful first drafts while underscoring the necessity of human expertise for validation, interpretation, and critical integration. This study, therefore, contributes a dual perspective: a comprehensive overview of the asbestos legacy across countries and a methodological reflection on the opportunities and pitfalls of employing AI in geoscientific and environmental research.

1. Introduction

Asbestos, which consists of naturally occurring silicate mineral fibers belonging to the amphibole (actinolite, amosite, anthophyllite, crocidolite, and tremolite) and serpentine (chrysotile) groups [1], has been used for millennia due to its remarkable fire-resistant and insulating properties. Because of its widespread use, significant health concerns—such as respiratory diseases like lung cancer, mesothelioma, and asbestosis—have been observed [2,3], especially during the 20th century, though the first evidence of health risks was observed during the Roman period [4]. Moreover, other extra-respiratory diseases can be related to past asbestos exposures [5,6,7,8,9,10,11,12,13]; however, this topic needs to be further studied for clear evidence.

Asbestos national standards vary widely among countries, and it is difficult to find correct information on this topic. Moreover, rules relating to asbestos bans in different countries have evolved over the years, making it difficult to find correct information.

For this study, the research team used artificial intelligence (AI) as an exploratory and comparative tool to support the collection of information on asbestos history and regulatory status around the world. The aim was to analyze the strengths and weaknesses of AI generative tools before applying them to asbestos-related topics with more limited evidence (e.g., its role in gastrointestinal cancer development) [13].

The application of generative conversational models, such as Gemini (Google, 2024) and ChatGPT (OpenAI, 2025), allows for an extensive and detailed mapping of asbestos in nearly 200 countries, starting with simple but targeted textual requests (prompt engineering). Indeed, this workflow is useful because asbestos regulations vary widely among different countries, both in terms of ban dates and history of asbestos utilization and regulation. AI generative tools are powerful instruments that can be fundamental in future works regarding asbestos-related diseases. This initial study was designed to evaluate the strengths and limitations of AI application for well-documented topics before extending such methods to areas where evidence is still incomplete or debated. Thus, while asbestos was selected as a paradigmatic example due to its global relevance and extensive literature, the central contribution of this paper lies in critically assessing the potential and limits of AI in scientific research workflows.

The aim of this paper was to delve into the history of asbestos use and regulation in various countries, highlighting the similarities and differences in their experiences. This work began with a general prompt to the AI software Gemini (version 2.5 flash; 13 December 2024): “I want to write a scientific paper about asbestos”. Subsequent requests entailed more detailed prompts and an increase in the number of countries considered. After this first test, a deeper inquiry into asbestos regulations was carried out. ChatGPT (GPT-4-turbo variant) was chosen as the best tool, as it appeared to work better than the first software. Initially, the document created using Gemini was uploaded onto ChatGPT to produce a more narrative text (the first ChatGPT query was carried out on 17 April 2025). Afterward, the text produced by the AI tool was appropriately deepened for each country, with a brief introduction about asbestos history. In this work, the AI answers will be reported in italics, followed by a more in-depth analysis and fact-checking by the authors with the support of scientific publications. This work will prove the usefulness of AI in data collection, saving time (and operators) during data collection and comparison, which is particularly useful given the large number of data sources and considering the variability in national regulations on asbestos (e.g., the presence or absence of bans and the year in which asbestos use was prohibited in different countries). Free versions of AI tools were utilized to produce this work in order to make it more accessible to the general population and to people who do not have access to AI premium versions. Nevertheless, a comparison of the results obtained by free and premium versions of ChatGPT was carried out.

Finally, this work also reflects on ethical and legal considerations emerging from the use of generative AI in scientific contexts, including issues of transparency, copyright, and accountability.

2. Materials and Methods

The drafting of this work started, as mentioned above, with the following simple prompt posed to the Gemini AI tool (version 2.5 flash) on 13 December 2024: “I want to write a scientific paper about asbestos”. The authors decided to give minimal information to AI to avoid restricting its answers, reserving more precise prompts for subsequent, in-depth exploration of selected topics. These topics were chosen based on the possibility of unequivocal verification. AI answered by offering examples of potential research questions, such as long-term health effects, methods to improve environmental monitoring, economic and social impacts, and approaches to raising public awareness of asbestos hazards. Additionally, it outlined viable research methodologies (e.g., review, epidemiological study, laboratory investigation, field study), considerations (ethical issues, data quality, scientific rigor), and practical recommendations (e.g., crafting a strong introduction, developing a clear research design, rigorous data collection). The second prompt was “I want to write a history about asbestos in the whole world, comparing the different countries”. This case study was selected because it enables easier verification against established data, in contrast to contemporary research questions where evidence remains incomplete or debated (e.g., the possible link between asbestos exposure and colon cancer) [13]. The AI tool produced a concise document entitled “A Comparative History of Asbestos: A Global Perspective”; the information quality was adequate, but the format was a point-by-point report and was thus unsuitable for a research paper. Moreover, the document contained limited novel information and failed to inspire further inquiry by the authors. For these reasons, the operator chose another AI tool, ChatGPT (GPT-4-turbo variant), to enrich the document generated by Gemini. The subsequent prompt to ChatGPT was “I want to include asbestos history in the countries worldwide”. This AI system assembled information about 19 countries; two further prompts requested supplementary paragraphs detailing additional national bans. At the end of this initial sequence, the operator asked the AI to consolidate the information into a single document for use in developing this article. The operator also requested the generation of images illustrating a world map and timelines of asbestos bans across countries.

The following sections present and discuss a combined version of results retained from both AI tools, which have been critically analyzed by researchers. Italicized text denotes AI-generated content, followed by the authors’ in-depth analysis based on scientific literature.

Principal information about the AI tools used is reported in

Table 1. List of the AI tools and related information about their utilizations. N.A. = not applicable.

AI	Utilization Period	Information Precision	Total Time of Utilization (Estimation of Effective Use)	Figure Creation	Advantages	Disadvantages
Gemini (free version)	In the beginning of the work	Quite precise, too much schematic	2 h	N.A.	Good scheme creation, good information collection	Too much schematic
ChatGPT (free version)	About 5 days	It gives good information, but it needs a careful revision of the contents	8 h	Yes, but it needs a careful check; it is often less precise	Good writing text, quick information collection	Need a careful check for the information; image creation sometimes does not work well

.

3. Results

3.1. Asbestos History Collection

3.1.1. Prehistoric Period

AI reports evidence of asbestos usage during the prehistoric period, particularly in Northern Europe (Finland, Karelia—Russia, and parts of Scandinavia), dating back approximately 2500 BC. The primary reported applications were in pottery and in the creation of asbestos ceramics for high-temperature cooking and storage [14,15,16,17]. Documentation also suggests asbestos was utilized in textile and burial garments, even as a symbolic value. Furthermore, AI reports that asbestos use occurred in Canada, though evidenced only to a limited extent. Indigenous groups reportedly employed it for its decorative and insulating properties. However, the AI provided a reference to a scientific paper that could not be verified upon review; only a few non-academic websites were found, offering generic information about this topic.

3.1.2. Ancient Civilizations

The AI-generated text began with the following statements: “The earliest known use of asbestos dates back to ancient civilizations like the Greeks and Romans, who valued its fire-resistant properties. They used it in various applications, including wicks for lamps, theater curtains, and even burial shrouds”. It is important to note that the authors might have accepted this information as accurate without posing specific inquiries regarding prehistoric asbestos use, potentially leading to uncritical acceptance of the AI initial response. According to [18], ancient Greek philosophers used two terms to denote asbestos minerals in their writings: αμιαντος (“amiantos”, synonymous with the English word “asbestos”) and ασβεστος (“asvestos”, the precise transliteration rather than “asbestos”). Both terms functioned as nouns or adjectives. In the first case, “amiantos” is synonymous with “asbestos”, while adjectivally it means “pure” or “undefiled”. Conversely, “asvestos” as a noun means lime, quicklime, or unslaked lime (CaO); as an adjective, it means inextinguishable, unquenchable, or not quenched. During the Latin period, Pliny the Elder interpreted another Greek word, ασβεστινον (“asbestinon”), to mean incombustible material (specifically, incombustible linen). Besides “asbestinon”, Pliny employed two other terms: “amiantus”, referring to fire-resistant alumen, and “asbestos”, which indicated iron-colored material.

3.1.3. Middle Ages and Renaissance

AI did not consider these two periods; nevertheless, significant events in asbestos history occurred during these centuries. For example, during the Middle Ages, Charlemagne (748–814 AD) used an asbestos tablecloth. After meals, he would throw it into the fire to astonish his guests. This legend led to one hypothesis that Charlemagne’s death could have been caused by mesothelioma [17]. Another notable example comes from Marco Polo (1254–1324 AD), who described in his manuscript “The Travels of Marco Polo” (Italian title: “Il Milione”, lit. “The Million”) the origin of asbestos minerals in Chapter XXXIX, specifically referring to the City of Chinchitalas (likely located in Eastern Turkestan). Marco Polo detailed how asbestos was used in textile production to make clothing. Additional evidence of asbestos use is found in several manuscripts from other regions of Asia [17], emphasizing the high financial and social status associated with the ownership of asbestos-containing goods. In the realm of art, asbestos was employed in the production of composite materials for Byzantine wall paintings [19].

During the Renaissance, minimal evidence of asbestos uses has been reported, and no scientific papers on this topic were identified by our research group. The most frequently cited use of asbestos minerals from this period refers to the naturalist physician Boethius, who reportedly included asbestos as a component in medicinal formulations during the 17th century [20].

3.1.4. Modern Period

AI generated the following text: “The Industrial Revolution marked a significant increase in asbestos production and use. Its versatility and affordability made it a popular material in construction, manufacturing, and shipbuilding”. Commercial asbestos mining began in Russia in the early 18th century [18]. In the latter half of the 19th century, the asbestos market expanded, leading to a rapid increase in mineral applications. During this era, Italian mines reopened, and substantial chrysotile deposits in Canada (Québec) began to be exploited, notably from 1878 onward. The global spotlight intensified in 1878, when asbestos products were publicly presented at the Paris Universal Exposition [18]. In Russia, a major chrysotile asbestos deposit was discovered in 1884, near Asbest City, with mining commencing two years later, in 1886 [18].

Regarding crocidolite asbestos, extraction started in 1893 in Northern Cape Province, South Africa (near Koegas), followed by expansion in the Pomfret area in 1926. Amosite asbestos production later began near Penge, Transvaal Province, South Africa, in 1916 [18].

Historical statistics from USGS indicate global asbestos production was approximately 31,500 tons in 1900, peaking at 4,790,000 tons in 1977. By 2022, total global output decreased to 1,330,000 tons (USGS data, accessed on 4 July 2025) [21].

Between 1900 and 2022, cumulative asbestos consumption exceeded 213 million metric tons (data from USGS [21], accessed on 4 July 2025). Fluctuations in production occurred, with notable declines during the Great Depression (a drop from 400,000 metric tons in 1929 to 205,000 metric tons in 1932), as well as during shorter recessions (1949, 1978, 1981, and 1986), preceding modern regulatory bans. These downturns coincided with global or regional economic crises, including the 1949 recession (November 1948–October 1949) and the crises of the late 1970s and early 1980s (the 1970s energy crisis, the Indian crisis in 1979–1980, and the Latin American debt crisis in the late 1970s and early 1980s). The lowest output in 1986 followed a downward trend beginning in 1983, during the early-1980s recession, significantly affecting Canada’s mining sector—the world’s second largest producer—with average output per worker decreasing by about 1% [22]. Although direct relationships are not fully documented, the temporal overlap suggests that global economic slowdowns influenced asbestos demand.

3.2. Asbestos-Related Diseases History Collection

AI was prompted to provide an overview of asbestos-related diseases (ARDs) throughout history and generated the following summary: “While the clinical recognition and nomenclature of asbestos-related diseases—namely asbestosis, mesothelioma, and asbestos-induced lung cancer—emerged only in the early 20th century, various forms of respiratory morbidity associated with asbestos exposure were descriptively acknowledged in premodern historical records. This retrospective review draws from classical texts, archaeological findings, and proto-industrial medical observations to trace the early footprints of asbestos-related pathology, long before the material’s toxicological properties were scientifically understood”. The authors subsequently cross-referenced this preliminary reply with the relevant scientific literature, expanding it through a comprehensive examination of ARDs across different historical periods. The results are detailed in the following sections.

3.2.1. Prehistoric Period

As with the history of asbestos uses, the AI provided no information for the prehistoric period, instead starting its overview with classical antiquity. When prompted for more details about prehistoric times, AI answered that there is no direct evidence of ARDs from that era due to the absence of medical documentation, the long latency period of such diseases, and the likely low level of exposure. Additionally, it noted that populations from these epochs were unaware of the health risks associated with asbestos use. It can therefore be hypothesized that ancient populations employing asbestos—for example, in Finland, Russia, or Canada—may have been exposed to fiber concentrations sufficient to induce ARDs. However, no dedicated scientific papers on ARDs in the prehistoric period were found by the authors. The absence of direct evidence highlights a key limitation of generative AI in historical scientific research: when explicit sources are lacking, AI systems tend to either extrapolate or omit information, reinforcing the necessity for rigorous human validation and expert interpretation.

3.2.2. Classical Antiquity

The earliest references “To the health hazards posed by asbestos are rooted in Greco-Roman antiquity, where the material was both admired for its fire-resistant properties and noted for its adverse effects on those who processed it”. This point stimulated the authors to explore more deeply the history of ARDs during ancient times, shifting the research focus beyond commercial or legal aspects. “Among the most salient sources is Pliny the Elder’s Naturalis Historia (about AD 77), which explicitly refers to lung disease in slaves who wove asbestos cloths (verified by the authors, [4]). Pliny’s warning that these workers developed “disease of the lungs” from inhaling fine dust resonates with modern understandings of asbestosis, a chronic interstitial lung condition resulting from prolonged inhalation of asbestos fibers”. This description is consistent with what we know as asbestosis, formally characterized in the 1920s by the physician Cooke [23].

A second relevant source often mentioned is Strabo, “A Greek geographer active during the reign of Augustus [18], who recorded that workers in the asbestos mines of Cappadocia exhibited rapid respiratory decline and premature mortality”. Indeed, this author referred to a mine where slaves, who had been sold cheaply due to their crimes, were employed [24]. AI also pointed out that “While neither author employed the diagnostic precision of contemporary medicine, their observations demonstrate a clear empirical link between occupational asbestos exposure and deleterious respiratory outcomes”. This point is particularly noteworthy because it suggests that the adverse effects of asbestos on human health were recognized, at least anecdotally, since ancient times, despite the continued use of these minerals during those periods. Finally, the AI concluded the section on ancient times with the following sentence: “The absence of formal anatomical terminology does not diminish the validity of these descriptions, which likely reflected the early manifestations of occupational pneumoconiosis—a concept only medically codified many centuries later”.

3.2.3. Middle Ages and Renaissance

For the Middle-Age period, AI answered as follows: “During the medieval period, Europe experienced a decline in both the use of asbestos and the systematic recording of occupational disease. Asbestos, having become a rare and mystified commodity, was often imbued with magical or religious significance rather than studied for its physiological effects”. Indeed, it has been reported that asbestos was used in various applications during this period, including warfare, cremation rituals, and by the wealthy for decorative and practical purposes [25]. “Consequently, there is a notable absence of direct medical commentary on asbestos-related illness during this era. Nevertheless, the broader recognition of miners’ diseases persisted”.

For the Renaissance, however, AI found some evidence and reported interesting information.

“Physicians such as Avicenna and later Georgius Agricola (in De Re Metallica, 1556) alluded to “ulcers of the lungs” and progressive respiratory impairment among those engaged in mining activities”. Indeed, Agricola became a physician in a mining town after his studies in medicine and natural sciences. His opera was divided into twelve sections, and the last one described both accidents and diseases affecting miners [26]. “Although these references did not explicitly mention asbestos, they underscore an emerging conceptual framework for environmental and occupational etiology of lung disease that is a foundational step toward modern occupational medicine. It is plausible that individuals exposed to asbestos-like fibers, even if incidentally, were subject to the same respiratory degradation noted in these early accounts”.

3.2.4. Industrial Precursors: Proto-Clinical Recognition in the 19th Century

As concerns the modern era, AI provided more detailed information, reflecting the expansion of industrial uses of asbestos and advances in medical research on related diseases. The following section presents the texts obtained from AI:

“The nineteenth century marked a transitional period in which asbestos usage expanded significantly across Europe and North America, driven by industrial demand for fireproofing, insulation, and durable textiles. With this expansion came a discernible rise in pulmonary morbidity among factory and mine workers. However, the precise causal relationship between asbestos exposure and disease remained elusive due to both diagnostic limitations and the long latency of asbestos-induced illnesses.

One of the earliest institutional acknowledgments of asbestos-related health risks was articulated in 1898 by Lucy Deane, a factory inspector for the British Home Office”. Here, AI reported a document by Lady Deane, dated 1899 and titled “Report on the effects of dust inhalation in asbestos factories”, published by Her Majesty’s Stationery Office. However, the original report could not be located, although secondary sources referencing this report were found [27]. Interestingly, the second proposed citation was accurate despite lacking any mention of Lucy Deane by name [28]. “In her official report, Deane highlighted the “sharp, glass-like particles” of asbestos dust and warned of their insidious effect on workers’ lungs. Her observations were not derived from histological or radiographic data but from direct exposure assessment and case pattern recognition, lending them considerable historical significance. This report serves as an antecedent to formal medical descriptions of asbestosis, which would only be delineated in the 1920s”. Additional research on Lucy Deane has revealed that she was one of the first female inspectors working within the asbestos industries. Specifically, she collaborated with Adelaide Anderson, and together they were the first to document the adverse effects of asbestos exposure on women employed at the Turner brothers’ factory in Rochdale (Yorkshire, UK). Their findings were published in the 1898 Annual Report of the Lady Inspectors of Factories. Notably, this report was the first to emphasize the significance of asbestos fiber shape and dimensions in relation to health risk [29].

“Additionally, scattered physician reports from the 1830s through the 1870s detail cases of chronic bronchitis, dyspnea, and unexplained fibrosis among workers in textile mills and asbestos processing plants, particularly in Italy and the United Kingdom. These clinical features, when analyzed through a modern lens, are highly suggestive of interstitial lung disease and likely include early cases of undiagnosed asbestosis or pleural fibrosis”. During the fact-checking of this second part of the section, it was not possible to conclusively confirm or refute the information reported using the methods employed.

3.2.5. Methodological Limitations and Retrospective Epidemiology (AI Opinion)

“The primary obstacle in identifying premodern asbestos diseases with certainty lies in the absence of diagnostic technologies such as chest radiography, pulmonary function testing, or histopathology. Furthermore, the symptoms overlapping with other prevalent conditions of the time—most notably tuberculosis—complicated clinical differentiation. Nevertheless, historical epidemiologists and occupational health scholars have employed retrospective analysis of medical texts, factory records, and death registries to reconstruct plausible case series consistent with asbestos-related morbidity.

Moreover, the latency period of asbestos-related illnesses—often spanning decades—meant that causality was not readily apparent. Many early deaths were likely attributed to generalized “consumption” or “lung rot,” terms that obscured specific etiologies and delayed recognition of industrial pathology. The lack of longitudinal surveillance and the fragmentary nature of archival data further hinder definitive epidemiological reconstructions. Still, the recurrence of specific symptom complexes among exposed populations provides compelling circumstantial evidence of premodern asbestos toxicity”.

3.2.6. The 20th Century

AI introduced this section by referring once again to Lucy Deane’s work as a precursor of the research conducted during the twentieth century. Nevertheless, it concluded this introduction with the statement: “despite her (Lucy Deane’s) prescient observations, the burgeoning demand for asbestos as a “miracle material” with fireproof and insulating properties led industries to ignore the associated health risks [30]. The material became integral to shipbuilding, construction, and manufacturing, embedding itself deeply into global economic expansion [31].

The human cost of this industrial enthusiasm began to surface more starkly in the early 20th century. In 1924, the death of Nellie Kershaw, a 33-year-old textile worker, marked a turning point. Kershaw suffered from severe pulmonary fibrosis, and her autopsy revealed the characteristic lung damage now recognized as asbestosis [2,32,33]”. It should be noted that this case was described by Dr. William Edward Cooke, who first used the term “asbestosis”. The earliest presentation of pulmonary fibrosis in an asbestos worker was by Sir Montague Murray in 1906 at the Samuel Committee, although he did not employ the term “asbestosis” [29]. “Her case led to the first formal diagnosis of an asbestos-induced disease. Nevertheless, it would take decades before regulatory bodies seriously considered asbestos a major occupational hazard [34]. As reports accumulated, researchers in the mid-20th century uncovered further connections between asbestos exposure and a spectrum of deadly conditions, including lung cancer and pleural thickening. Among these, mesothelioma—a rare, aggressive cancer affecting the lining of the lungs and abdomen—emerged as the most fearsome consequence [35].

Pathologist Chris Wagner’s landmark study in the 1960s offered irrefutable evidence linking asbestos exposure directly to mesothelioma. Conducted among South African miners, Wagner’s research highlighted not only the occupational risks but also the environmental dangers, demonstrating that even low-level exposure could be fatal [32]. This study was pivotal in shifting public health perspectives and eventually led to tighter regulations in many developed countries [36].

However, the legacy of asbestos remains troublingly persistent. Today, asbestos exposure is estimated by the World Health Organization (WHO) to cause more than 200,000 deaths annually [37,38]”. In this case, AI overestimated the mortality figure reported in [37]; the referenced study actually indicates a minimum of 107,000 deaths. Regarding [38], this work provides a comprehensive review of toxicological and epidemiological literature supporting the evidence of an association between chrysotile exposure and increased risk of lung cancer and mesothelioma. “Even as many industrialized nations have imposed bans or severe restrictions on asbestos use, the material continues to be mined, exported, and utilized in several developing countries where occupational health regulations are weak or poorly enforced [39]. Underreporting of asbestos-related diseases remains significant, often due to inadequate health surveillance systems, lack of awareness, and the long latency period of diseases like mesothelioma, which can take decades to manifest [40]. Consequently, asbestos continues to claim lives silently, highlighting an urgent need for global cooperation to eliminate its use and to strengthen surveillance and compensation systems for affected workers worldwide”.

This extensive AI-generated response provided important information about ARDs during the twentieth century. However, when we queried ChatGPT to verify whether this summary was comprehensive, the tool affirmed its completeness. Nevertheless, experts in the field of ARDs identified a significant omission (Dr. Irving J. Selikoff).

Upon being alerted to this, the AI acknowledged that excluding Dr. Selikoff represented a major gap in any thorough review of ARDs. Notably, Selikoff’s 1964 study, published in the Journal of the American Medical Association (JAMA), documented a dramatic increase in mortality from asbestosis, lung cancer, and mesothelioma among asbestos-exposed workers in the United States. This seminal work played a critical role in accelerating research on ARDs [41].

The data reported, coupled with the operators’ fact-checking efforts, underscore the necessity for comprehensive validation of the AI-generated content through multiple queries and strategic questioning, particularly on topics well known to subject experts.

Building on these historical insights, the subsequent section aims to provide a concise overview of current asbestos use and regulatory frameworks worldwide, supported by AI-generated data. We will briefly address the specific histories of various countries.

3.3. Country-Specific Timelines

A list of nations was generated by querying Chat-GPT four subsequent times about the global status of asbestos use. This strategy, involving repeated questioning, aimed to identify errors or inconsistencies in the responses while expanding the dataset to include a larger number of countries. The compiled results are presented in the Supporting Materials, listing countries alphabetically. All AI-generated answers were subsequently verified against official websites via Google search for each country. In Table S1, countries are listed alphabetically along with their asbestos ban status, categorized simply as “ban”, “partial ban”, or “no ban”. During the authors’ verification process, discrepancies were found between nations listed in Table S1 and those mentioned in the AI-generated text, as detailed in the Supporting Material. This highlights the necessity for meticulous validation of AI-generated content, as inaccuracies can occur. Specifically, five mistakes were identified in Table S1, all attributable to AI misinterpretations—for example, instances where a formal ban exists but is not effectively enforced. In one case, Macedonia was initially omitted by AI, and it was subsequently added by the operator. Additionally, Serbia was initially reported by AI as lacking asbestos ban legislation. Corrections made by the authors, based on website information [42], are indicated in bold in Table S1.

Moreover, AI was tasked with generating an image depicting a timeline of asbestos bans (Figure 1). At first glance, the data appeared consistent with the actual asbestos ban statuses across countries; however, a thorough review by the authors identified some minor inaccuracies. While the bans implemented during the 1980s were accurately represented, numerous discrepancies were found in the subsequent decades. The verified national bans, as confirmed by the authors, are presented in Figure 2. Notably, the errors in the ban dates differed from those found in the text reported in the Supporting Materials, underscoring the essential role of operator verification in correcting AI-generated inaccuracies.

Table 2. Countries with an active asbestos ban and the relative year when it was emanated.

Country	Year	Country	Year	Country	Year	Country	Year
Algeria	2009	Chile	2001	Finland	1994	Iran	2012
Argentina	2003	Colombia	2021	France	1997	Iraq	2016
Australia	2003	Croatia	2006	Gabon	2004	Ireland	1999
Austria	1990	Cuba	2001	Germany	1993	Israel	2003
Bahrain	1996	Cyprus	2005	Gibraltar	2005	Italy	1992
Belgium	2001	Czech Republic	1999	Greece	2005	Ivory Coast	1996
Brazil	2017	Denmark	1986	Greenland	2010	Japan	2006
Brunei	1994	Djibouti	1999	Honduras	2004	Jordan	2006
Bulgaria	2005	Egypt	2005	Hungary	2005	Kuwait	1995
Canada	2018	Estonia	2001	Iceland	1983	Latvia	2001
Liechtenstein	2005	New Caledonia	2007	Saudi Arabia	1998	Switzerland	1990
Lithuania	2005	New Zealand	2016	Serbia	2015	Taiwan (PRC)	2018
Luxembourg	2001	Nicaragua	2001	Seychelles	2009	Turkey	2010
Macedonia	2014	Norway	1984	Slovakia	2005	United Kingdom	1999
Malta	2005	Oman	2017	Slovenia	1996	United States	2024
Mauritius	2004	Poland	1997	South Africa	2008	Uruguay	2003
Monaco	2005	Portugal	2005	South Korea	2009
Mozambique	2010	Qatar	2010	Spain	2002
Netherlands	1993	Romania	2007	Sweden	1982

presents the list of countries with an active asbestos ban, along with the respective year in which each restriction was enacted, to provide clarity for readers. These dates were meticulously verified by the authors.

Regarding AI responses, we highlight two examples of differing answers or interpretations. During the creation of Table S1, issues arose in interpreting the operators’ queries posed to AI: in some instances, certain countries were omitted, or incomplete tables were generated. This situation led us to observe that ChatGPT provided two conflicting pieces of information about Ukraine and Liechtenstein on two different occasions. When queried about these discrepancies, the AI responded that both sets of information were technically correct. The explanations for these differences are detailed in the following two sections.

3.3.1. Ukraine (First Answer “Ban in 2022”, Second Answer “No Ban”)

Regarding Ukraine, AI clearly explained the reason for its inconsistent responses during the two different days of data collection. The explanation is as follows.

“In 2017, Ukraine introduced a full ban on asbestos use through Cabinet of Ministers Resolution No. 339, prohibiting the use of all forms of asbestos, including chrysotile. However, in 2019, the ban was annulled due to strong lobbying from industrial sectors, reinstating chrysotile asbestos in national construction standards.

Following Ukraine’s accelerated integration with the European Union—especially after signing the EU-Ukraine Association Agreement—public health advocates and international organizations increased pressure for realignment with EU chemical safety laws.

In 2022, the Ukrainian government, through the Ministry of Health, reaffirmed its commitment to a national asbestos ban, aiming to bring its policies closer to European norms. Despite these policy announcements, actual implementation remains inconsistent due to disruptions caused by the ongoing war and administrative instability [43].

Summary: Ukraine officially reintroduced its commitment to banning asbestos in 2022, but full enforcement and on-the-ground application are still incomplete”.

Considering this explanation, it is clear that Ukraine has a formal ban on asbestos; however, it is not yet effective due to the ongoing war situation. Consequently, the authors did not regard the asbestos ban in Ukraine as active, and it is omitted from both Figure 2 and Table S1.

3.3.2. Liechtenstein (First Answer “Ban in 2005”, Second Answer “No Ban”)

Again, based on the interpretation of the situation, it is possible to understand why the AI generated two different replies. The explanation provided by the AI is reported below.

“Liechtenstein does not have an independent, explicit asbestos ban within its domestic legislation. Instead, it is a member of the European Economic Area (EEA) and participates in a customs union with Switzerland.

Switzerland implemented a comprehensive asbestos ban in 1990, and the European Union enforces asbestos bans under multiple directives [44]. Consequently, Liechtenstein automatically applies these bans through legal harmonization, despite the absence of a national-level legislative text.

In practical terms, asbestos is entirely banned in Liechtenstein’s territory, enforced through Swiss and EU compliance mechanisms”. In this sentence, AI gave two references that were inaccurate; however, the information regarding the asbestos ban in Liechtenstein is correct (see, for example, [42]).

“In summary, although Liechtenstein has no direct national ban, it fully prohibits asbestos through automatic adoption of Swiss and EU regulations”.

Moreover, during the final verification of this information, the authors identified an error in the date of the asbestos ban in Switzerland. Initially, 2002 was considered the year the ban came into effect. However, the actual ban was implemented in 1990. The 2002 date corresponds instead to a reaffirmation of the ban, aligning Swiss regulations more closely with European and international regulatory frameworks.

The two examples illustrate the importance of careful oversight by operators when using AI for data collection. It is essential to verify all information step by step. Accordingly, the nation-by-nation data collected in the Supplementary Materials (Text S1) were reviewed individually to ensure accuracy and reliability.

AI proved to be a valuable tool for gathering extensive information on approximately 200 countries, substantially reducing the time and effort required during the initial phase of analysis. This feature represents a key advantage in writing a scientific article such as this, enabling the creation of a first draft containing core information that can subsequently be verified and enriched by the authors. It is, however, crucial that authors do not rely solely on AI to generate text without careful review and assume responsibility for refining the language and verifying the content. In response to the operator’s query, “Thanks, so in how many countries is there the ban, and in how many nations is there not?”

Table 3. Number of countries where an asbestos ban is active and where there are no or partial regulations, created by AI as opposed to data obtained by the authors. 195 = 193 UN Member States + Vatican City + Kosovo (both included for completeness).

Category	Number of Countries (AI)	Number of Countries (Human)	Notes
Banned	69	73	Full, legal prohibition of asbestos use
No ban/Partial ban/No formal ban	126	122	No national ban, partial ban, or just follows external rules

provides a simplified summary of countries with active asbestos bans, those with partial bans, and those without bans.

Nevertheless, this number was derived from a baseline that does not fully correspond to the actual number of countries with an asbestos ban. Careful cross-referencing of AI-generated data revealed contradictions between different sources. For example, while asbestos ban dates for Austria and Belgium are reported in Figure 1, these two countries are absent from Table S1. This discrepancy arises from the different datasets compiled by the authors, which are reported in Table 2.

Furthermore, the operators commissioned AI to generate an image summarizing global asbestos consumption among the leading asbestos-producing countries, aiming to evaluate AI’s capability in scientific image production. The initial image generated by AI contained incorrect data. During a subsequent revision in July 2025, AI was asked to review the April 2025 image (Figure 3A). AI acknowledged errors in the original values and supplied alternative data. When queried about the data source, AI cited the USGS website. However, upon direct consultation of the USGS report [45], operators found that the second dataset suggested by AI was also inaccurate (Figure 3B). This case exemplifies the risks of delegating scientific image production solely to AI tools. The authors therefore generated a corrected image using accurate production values (Figure 3C).

The obtained results demonstrate that AI is a highly valuable tool in data collection, particularly when operators face large volumes of data or significant variability of information. Regarding asbestos regulations and production/consumption, wide variability exists among countries worldwide, and approaching these topics is prone to errors, especially when the subject matter is not well known. Moving forward, it may be beneficial to systematically collect the status of asbestos use and regulations on a year-by-year basis. This would facilitate access to accurate information for both researchers and the general public, particularly if AI models are specifically trained in this domain.

A comparative analysis of the AI-generated asbestos regulations is presented in the following section.

3.4. Asbestos Ban Status Worldwide

AI reports that, “As of 2025, a total of 76 countries (based on the data previously reported, this number is not correct, and it might be better to consider approximately 70 countries; moreover, this value differs from the 69 countries reported in Table 3) have formally enacted comprehensive bans on the production, importation, and use of asbestos and asbestos-containing materials”. Thus, counties with regulation “Represent approximately 39% of the world’s sovereign states (37% considering the 73 nations considered by the authors). Conversely, 119 countries (61% of nations globally) have not established a full prohibition (for the authors, these nations are 122, equal to 63% of the total). In fact, in these nations, asbestos continues to be extracted and/or produced and/or utilized, either legally with minimal restrictions or under partial regulatory frameworks. Despite decades of robust scientific evidence unequivocally linking asbestos exposure to a spectrum of severe diseases, global regulatory measures remain inconsistent. This situation underscores the pressing need for international collaboration and policy harmonization to eradicate asbestos-related diseases and to promote equitable protection of public health worldwide”.

Following this section, AI was tasked with producing a global map of asbestos bans. Despite more than ten attempts to correct the map, it was not possible to generate an accurate image distinguishing countries with bans from those without regulation (Supplementary Materials, Figure S1).

3.5. Comparative Analysis of Global Asbestos Histories

Analyzing the AI-generated results, it is evident that the overall output enabled the collection of a substantial and impressive volume of information within a relatively short timeframe (approximately 2 h using AI Gemini and an estimated total time of about 8 h employing ChatGPT to generate the list of countries as well as associated figures and tables). Nevertheless, careful verification of the results was essential, including instances where consecutive responses appeared consistent, as legal aspects proved particularly complex to accurately gather. Countries where data collection was notably challenging tended to be those experiencing ongoing conflicts or industrial and legal pressures. It should be emphasized that the information reported primarily concerned the word “ban”, without consistently distinguishing between partial or total bans, which may lead to reader confusion.

Based on the collected results, a general overview can be produced by comparing the asbestos regulatory histories of various countries. Accordingly, when prompted to generate a comparative analysis of asbestos histories worldwide, AI produced an initial article including a chapter devoted to this topic. Following this, a request for a more detailed section was made, resulting in the generation of the subsequent section.

“Despite varying national timelines, the global history of asbestos reveals strikingly parallel patterns shaped by industrial interests, regulatory failures, civil society activism, and persistent economic dependencies”.

3.5.1. Early Adoption and Industrial Enthusiasm

“In the early 20th century, many countries enthusiastically embraced asbestos for its exceptional fire resistance, tensile strength, and insulating properties. Early industrial leaders such as the United States, Canada, and the United Kingdom drove mass consumption without fully acknowledging or understanding the material’s latent health hazards [35,46]. Asbestos became integral to a wide range of industries—e.g., construction, shipbuilding, and automotive manufacturing—fostering a widespread perception of it as a “miracle mineral” that outweighed emerging health concerns”.

During the 20th century, the use of asbestos in buildings, transportation, and a wide range of objects (including everyday use) increased, reaching a peak in 1977 (data from [21]). Subsequently, its use declined due to mounting health concerns and growing awareness among workers and the general public regarding asbestos-related risks. In the region where our research group operates (Piedmont, Northern Italy), a poignant symbol of ARDs is Casale Monferrato (Alessandria Province). This town hosted a large Eternit factory (the greatest in Europe) operational from 1907 to 1986 [47], and the adverse effects on the local population continue to persist.

3.5.2. Industry Influence and Regulatory Delay

AI also attempted to create a concluding section by compiling data that explain the delays in asbestos regulations across different countries. Specifically, it stated the following: “A critical feature of global asbestos histories is the role of the asbestos industry in delaying regulatory responses. Corporations systematically downplayed and obscured the mounting evidence of asbestos-related health risks through extensive lobbying, scientific manipulation, and misinformation campaigns [34]. In both the United States and Australia, internal company documents later revealed that industry leaders had known about the material’s fatal risks, including asbestosis and mesothelioma, decades before acknowledging them publicly [46,48]. This deliberate suppression of knowledge contributed to prolonged periods of insufficient regulation, enabling ongoing exposure for workers and the public”. This interpretation, although generated by AI, is accurate upon fact-checking. Considering the Italian experience—particularly the case of the Casale Monferrato population—this interpretation reflects the historical trajectory of asbestos use and related health impacts associated with the Eternit factory [49].

3.5.3. Role of Litigation and Civil Society

AI further extends its analysis by addressing the social and legal implications related to asbestos use, stating that “Litigation has served as a powerful tool for exposing the asbestos industry’s malpractices and compelling regulatory change. In countries with robust legal frameworks, notably the United States, United Kingdom, and Australia, legal actions against asbestos manufacturers revealed suppressed evidence and spurred more stringent occupational health regulations [50,51]. Civil society organizations, labor unions, and health advocates played pivotal roles in amplifying affected workers’ voices, pushing for public health protection, and eventually advocating for national bans on asbestos use [52]”. In this case, during fact-checking by the authors, AI cited a nonexistent research paper; therefore, the correct citation was provided. “Their activism highlights the importance of bottom-up pressure in countering industry resistance”. In this case, AI did not consider Italy, despite Casale Monferrato representing a pioneering example in asbestos regulation. Although Italy implemented a national asbestos ban in 1992 [53], the municipality of Casale Monferrato issued a local decree prohibiting the asbestos use and production in 1987, following the closure of the Eternit factory in 1986 [47].

The social implications in this area have been and remain profound, encompassing significant public health burdens and the moral consequences for families advocating for justice—the Eternit trial is still ongoing. The exclusion of this notable case by AI may stem from its algorithmic emphasis on English-speaking countries, as the queries posed in this study were formulated in English. Supporting this hypothesis, when asked in Italian, “What are examples of states where lawsuits against manufacturers of asbestos-containing materials have been more effective?” AI included Italy alongside the United States, the United Kingdom, and Australia (see Supplementary Materials, Text S2).

3.5.4. Economic Dependency and Continued Use

Following this analysis, AI continued its overview by addressing countries where asbestos extraction and/or manufacturing is still ongoing, stating the following. “Despite the overwhelming scientific consensus on asbestos hazards, several countries continue to mine and export asbestos, primarily chrysotile, due to economic dependency. Resource-abundant nations such as Russia, China, and Kazakhstan maintain asbestos production as a significant economic activity, often framing it as essential for employment and regional development [37,54]. These countries often contest international efforts to regulate asbestos strictly, arguing for “controlled use” despite evidence showing that no safe exposure level exists”. Moreover, further investigation revealed instances in which the purported safer properties of chrysotile asbestos are defended (e.g., [55]). This example reflects broader trends, as various associations and governmental institutions continue to advocate chrysotile use, particularly in countries where asbestos remains economically significant.

3.5.5. Regional Coordination and Global Disparities

In this section, AI generated a conclusion comparing functional regulatory frameworks, such as those in the EU, with countries that continue to produce asbestos. The AI-generated conclusion section is presented below.

“The European Union stands as a notable example of successful regional coordination, implementing comprehensive asbestos bans across its member states and integrating asbestos control into broader occupational health policies [56]. However, global regulatory efforts have faced substantial challenges. International agreements like the Rotterdam Convention, which aims to regulate the international trade of hazardous substances, have struggled to reach consensus on chrysotile asbestos. Producer countries, leveraging their economic interests, have persistently blocked attempts to list chrysotile under the Convention’s prior informed consent procedure, thereby perpetuating global regulatory disparities [52].

The comparative trajectories of asbestos across different nations underscore a shared narrative: the initial industrial zeal, compounded by corporate obfuscation and delayed regulatory action, led to widespread human suffering. While litigation and civil activism have achieved significant victories in some regions, entrenched economic dependencies and political resistance continue to undermine global efforts to eliminate asbestos-related diseases. Future success will likely depend on strengthening international cooperation and prioritizing public health over short-term economic gains”.

4. Discussion

In this work, the authors evaluated the use of AI in collecting information within the complex field of asbestos regulations worldwide. Initially, this research aimed to create a comprehensive asbestos review using the AI Gemini software; however, the results obtained were overly schematic and insufficient to support the development of a scientific paper. Although the information was detailed and provided an excellent foundation for further inquiry, it did not inspire the creation of new content. Subsequently, the authors employed another AI tool—ChatGPT—which enabled the generation of a more discursive text and inspired the compilation of a global overview of asbestos ban regulations. This tool proved valuable for rapidly obtaining information about different countries, facilitating the alphabetical ordering of nations, data schematization, and image generation. Nonetheless, the role of the operators in verifying the quality and accuracy of AI outputs was critical. Without a deep understanding of the subject matter, there was a high risk of incorporating false or misleading information. The legal landscape concerning asbestos is characterized by a wide array of regulatory frameworks, necessitating careful control and interpretation of collected data. Moreover, specific challenges were encountered during image generation. For instance, the AI struggled to accurately assign correct values to countries on the world map depicting asbestos ban statuses (Figure S1). Another example involved errors in the scale bars of asbestos consumption charts (Figure 3). These cases underscore the importance of meticulous source verification and human oversight when using AI-assisted tools in scientific writing. Overall, while AI can serve as an invaluable source of inspiration and support, it should not replace the scholarly rigor and critical judgment of human researchers.

In fact, conducting multiple queries to AI models over time, with variations in question formulations, enabled verification of response consistency and identification of contradiction. This approach facilitated the identification of both strengths and weaknesses inherent to AI. In the first case, the identified strengths included rapid response time, the capacity to aggregate dispersed knowledge, and data synthesis. Nevertheless, weaknesses were observed in the variability of responses when sources were lacking, the absence of real-time legislative updates, and ambiguities in regulatory classification, such as distinctions between total and partial bans. A representative example involves the information regarding Ukraine and Liechtenstein, where ChatGPT produced inconsistent outputs at different times. This underscores the necessity of complementing AI-generated content with rigorous validation by human researchers, utilizing official sources, regulatory documents, and institutional sites (e.g., WHO, ILO, and EU legislation). Such discrepancies align with literature findings, which emphasized that large language models (LLMs) are not expert systems but probabilistic text generators based on patterns learned during the training process, lacking true semantic understanding or communicative intentionality [57].

Overall, AI enabled the execution of a highly time-consuming task—the collection of asbestos regulations in nearly 200 countries—in a relatively short period. Furthermore, it facilitated the creation of an initial comprehensive draft, which the authors augmented with additional information on history, diseases, and law. This approach significantly reduced the time compared to conducting traditional research using classic search engines.

The analysis conducted through the application of generative AI tools revealed variable performance across the historical periods examined. For classical antiquity, AI-generated information was predominantly descriptive and general, lacking precise bibliographical references. This necessitated critical verification by the authors to differentiate unsupported extrapolations from historically documented evidence. During the Middle Ages, the output was even more limited, reflecting the inherent scarcity of available sources for this era. In this context, integrating archeological and historical materials was essential, highlighting that exclusive reliance on AI is insufficient to guarantee scientific robustness in understudied domains. In the modern age and throughout the Industrial Revolution, AI-generated content was richer but often characterized by oversimplifications and the absence of primary citations. Therefore, the researchers’ work of verification and contextualization efforts was crucial to reconstruct an accurate and coherent historical account. In the contemporary era, AI demonstrated improved information retrieval capabilities; however, inconsistencies and inaccuracies remained. Collectively, these findings confirm that, while AI can expedite the identification of key themes and serve as a valuable preliminary tool, expert critical appraisal remains indispensable to ensure the development of scientifically reliable narratives.

As it concerns the ethical considerations on the use of AI in scientific research, we can propose some important considerations. Indeed, the integration of AI into scientific research processes represents a methodological frontier of growing significance. However, the adoption of generative tools based on LLMs raises important epistemological and ethical issues. The experience gained in the present work, in which two generative models—Gemini and ChatGPT—were used to support the historical and normative investigation of asbestos globally, has allowed us to highlight four critical areas that deserve careful consideration. In the following Section 4.1, Section 4.2, Section 4.3 and Section 4.4, some important considerations are reported.

4.1. Epistemic Accountability and Content Validation

Although AI models demonstrate an extraordinary ability to aggregate and synthesize information, they lack epistemic autonomy or cognitive intentionality. Their responses are generated based on statistical correlations of linguistic patterns learned during training, without genuine semantic understanding or verification of source accuracy. Consequently, the responsibility for ensuring the accuracy, consistency, and scientific validity of the generated content rests with human researchers, who must engage in critical validation and contextualization of the results. This principle aligns with the concept of epistemic accountability, which mandates that all scientific content be traceable, verifiable, and subject to expert scrutiny [57].

This accountability extends to the design phase of AI-based tools. The inherent lack of accountability within language models implies that any misuse—including misinterpretation, systemic biases, or the dissemination of harmful information—is ultimately the responsibility of the developers, implementers, or users of these technologies. Indeed, the illusion of consistency and competence that language models generate—what [57] refer to as “stochastic parrots”—can obscure the absence of cognitive grounding, leading human users to overestimate the reliability of the generated content. This situation calls for the establishment of an ethics of epistemic curation, demanding that researchers exercise not only scientific rigor but also critical awareness of the inherent limitations of such technologies [58].

Finally, the epistemic dimension is closely linked to cognitive justice. Models trained predominantly on English-language data from mainstream Western sources tend to reproduce hegemonic worldviews, thereby marginalizing minority perspectives, knowledge, and languages [59,60]. Ethically responsible use of AI in scientific research should therefore incorporate transparent data documentation practices, robust bias assessment methodologies, and a concrete commitment to epistemic inclusion [61].

4.2. Methodological Transparency and Reproducibility of Results

Reproducibility is a fundamental epistemological pillar of contemporary science, serving as a necessary condition for the empirical validation of the generated knowledge. In the context of generative AI, this principle faces significant challenges due to critical structural issues. LLMs, such as ChatGPT, exhibit inherent randomness in their generative processes, resulting in variable responses even with identical inputs. This variability arises from factors including stochastic sampling methods (such as top-k or nucleus sampling), continuous model updates, and the lack of determinism within the computational framework [62,63].

In this context, methodological transparency is a prerequisite for establishing the scientific reliability of interactions with generative systems. It entails detailed and rigorous documentation of the conditions under which the outputs were generated, including the exact prompts used, the model version, the date and context of generation, any modifications applied to the output, and the criteria for selecting and interpreting responses [64]. Proper reporting of these elements enables traceability of the processes undertaken and controlled reproducibility of results, despite the inherently probabilistic nature of the tool.

Where such documentation is absent or incomplete, AI-generated content risks losing any epistemic value, becoming equivalent to results that cannot be verified or replicated. Consequently, the use of generative AI in scientific research must be accompanied by an explicit and shared methodological framework that allows the integration of AI-generated data into validated knowledge systems [65,66].

This requirement aligns with emerging recommendations on the responsible use of AI, which stress the importance of open standards, auditability protocols, and algorithmic accountability tools. These measures are supported not only for internal verification but also for peer review and replication within the scientific community [67,68].

4.3. Systemic Biases and Information Inequalities

The use of generative AI in scientific research presents a tangible risk of reproducing and amplifying pre-existing biases embedded within training data, with profound implications for epistemic fairness and the inclusiveness of knowledge production. LLMs are trained on vast text corpora that inherently reflect existing power structures, cultural hierarchies, and geopolitical imbalances. Consequently, there is a predominance of English-speaking sources, which are academically more visible, alongside a marginalization of local or peripheral experiences, which are often underrepresented in the global information landscape [57,69].

Content analysis generated by LLMs reveals the structural presence of linguistic, geographic, and cultural biases. These biases are not incidental but rather arise from the statistical learning process applied to text corpora that embody existing power hierarchies, geopolitical imbalances, and inequalities in knowledge production and dissemination. Notably, Anglophone and institutional sources are overrepresented, whereas local, non-Anglo-Saxon, or peripheral experiences tend to be underrepresented or entirely excluded [57,69,70], as exemplified by the Casale Monferrato case within the Italian context (see Section 3.5.3).

The use of generative AI in scientific research thus entails a tangible risk of reproducing—and in some cases amplifying—existing cognitive inequalities. This phenomenon has direct implications for epistemic equity, defined as the capacity of all contexts to generate and disseminate knowledge with equal status and visibility within the global scientific discourse [58,71]. Consequently, a dual effect emerges: the marginalization of non-hegemonic perspectives and the standardization of scientific narratives that tend to privilege dominant cultural models.

Addressing this issue necessitates the implementation of data and model auditing strategies, alongside intentional data diversification practices that incorporate sources heterogeneous in geographic, linguistic, and cultural origins. Additionally, it is essential to foster critical reflection on the epistemic injustice that may arise from the unregulated use of AI, making visible the systemic silences and structural exclusions present in the source data [72].

4.4. Cognitive Delegation and the Role of the Human Researcher

The integration of AI into scientific production processes raises important questions concerning cognitive delegation—that is, the partial or total transfer of complex intellectual tasks, such as critical analysis, data interpretation, and hypothesis formulation, to algorithmic systems. While AI-based generative models offer considerable heuristic potential, their deployment must be accompanied by epistemically active supervision from human researchers [73].

The central issue concerns maintaining an epistemologically responsible position, whereby AI is regarded as an extended cognitive tool [74] rather than a substitute for human critical thinking. In other words, while AI can serve as a catalyst for a researcher’s analytical and synthetic capabilities, it cannot assume—or authentically simulate—cognitive intentionality, ethical responsibility, or the ability to attribute meaning to the generated data [75,76].

From this perspective, an ethically sustainable use of AI in research requires conscious integration, wherein the researcher retains a regulatory role within the cognitive process, maintaining ongoing control over the validity of inferences and the appropriateness of interpretations. As observed by [77], the absence of such epistemic vigilance risks generating a form of opaque automation, thereby undermining transparency and accountability in knowledge production.

Finally, it is important to emphasize that over-reliance on intelligent systems may lead to the de-skilling of researchers’ critical faculties, thereby undermining the development of autonomous and reflective knowledge [78]. Consequently, the use of AI in scientific research should be complemented by training practices aimed at enhancing researchers’ capabilities to engage with algorithmic tools in a critical, responsible, and reflective manner.

4.5. Comparing Free and Premium Versions of AI Generative Tools

This work was initially aimed at evaluating the usefulness of AI tools in data collection and the generation of initial document drafts, using free software versions. However, an additional important test conducted by our research group involved comparing the free and premium versions of OpenAI ChatGPT. This test followed the same protocol applied to the free version. In this section, the main differences between the two accounts are presented and summarized in

Table 4. Comparison between ChatGPT Free and Premium versions.

Analyzed Aspects	Free Version	Premium Version
Amount and format of information	Fewer details, but presented in a more discursive and narrative style	Larger amount of information, often presented in bullet-point format
Integration with SCISPACE	Suggested and integrated	Not suggested
Initial setup	Same basic structure as Premium	Same basic structure as Free
Draft options	No choice available	Possibility to choose between a concise or extended draft
Identification of countries	Immediately provides 19 countries	Initially fewer countries, expandable up to 14 or put in groups
Content details	Split into two main areas (health case studies and environmental contamination) via SCISPACE	More detailed, but obtainable only through multiple targeted prompts
Single document (word count)	No choice available	Two options: 3000 or 7000 words
Single document generation (loss of information)	Loss of many countries and more synthetic content	Loss of some information, but preserved when generated step by step
Materials and methods	Approximate, requiring revisions	Approximate, requiring revisions
Comparative data analysis	Similar to Premium	Similar to Free
Bibliography	Reduced from over 100 references to ~30; includes non-existent articles or mismatched journal attributions	Same reduction, but with fewer errors compared to Free
Image generation	Same as Premium	Same as Free
Accuracy of consumption data	Errors for Spain and South Korea; Japan is also inaccurate	Correct for Spain and South Korea; Japan is still inaccurate

.

The comparative analysis between the free and premium versions of ChatGPT reveals significant differences in both the quantity and quality of the returned information. The premium version offers immediate access to a larger and more detailed volume of content, which is often presented schematically, primarily through bullet-point lists. In contrast, the free version tends to provide shorter, more limited texts from an informational perspective but articulated in a more discursive and linear manner.

A distinctive feature of the free version is its integration with the SCISPACE platform, which divides content into two primary strands specifically dedicated to case studies in healthcare and environmental contamination. This functionality is absent in the premium version, which delivers more in-depth content but necessitates subsequent targeted queries to reach a comparable level of detail. Moreover, since the premium format does incorporate SCISPACE usage, researchers intentionally opted not to utilize it, enabling assessment of the premium version’s efficiency without external tool influence. Both versions initiate from a similar foundation structure for work organization but diverge in customization capabilities. The premium version, notably, allows selection between synthetic or extended drafts, a feature unavailable in the free version.

Another notable difference arises in the identification of nations included in the analysis. The free version provides a relatively large number of countries (19 countries), whereas the premium version starts with fewer countries, which can be expanded upon request up to 14 or by generating grouped sets. Distinct approaches are also observed in document management: the premium version offers selection of texts with variable length—equal to 3000 or 7000 words—but carries the risk of information loss if outputs are not produced incrementally; the free version generates single, comprehensive documents, resulting in fewer countries analyzed and more concise content.

Both versions exhibit some common critical issues. The section dedicated to materials and methods is approximate in both cases and requires careful operator review. Similarly, comparative data analyses yield comparable results, while the bibliography experiences a drastic reduction, with approximately thirty references retained from over one hundred initially. Beyond this quantitative loss, inaccuracies are present in citation quality. Both versions include incorrect references, sometimes citing nonexistent articles or real articles attributed to journals and editions differing from the originals. Nevertheless, the premium version contains fewer errors than the free one, demonstrating slightly greater reliability.

Regarding image generation, no substantial differences emerged between the two formats. However, it is noteworthy that the premium version provided the correct image of asbestos consumption in the three main producing countries upon the first request (Figure S2). Some discrepancies were observed in the reported years of asbestos bans for certain countries: the free version produced errors for Spain and South Korea, whereas these values were corrected in the premium version. Nonetheless, both versions presented inaccurate information for Japan.

Overall, the comparison highlights that the premium version is a richer, more flexible, and more reliable tool, capable of offering greater detail and customization options but requiring a finishing process and incremental use to preserve data completeness. The free version, while simpler and more linear, carries a higher risk of information loss, reduced bibliographic accuracy, and a tendency to produce overly synthetic content in the final synthesis stages.

This analysis underscores the essential role of supervision by an experienced operator in content definition and verification, a fundamental aspect that complements the important opportunity AI offers in assisting and accelerating information processing and synthesis.

4.6. Copyright and Use of AI in Science

The role of generative AI tools in scientific content production presents significant challenges related to copyright and other forms of intellectual property protection. When AI models “read” (i.e., develop concepts and ideas) texts, tables, or images based on knowledge derived from other authors, studies, or data, there is a risk of recycling such materials without proper attribution or reference to the original sources. This situation raises issues that are not only ethical and academic but also legal in nature.

However, the lack of an autonomous style and authorial intentionality implies that, under current copyright law, AI-generated content cannot be considered original works eligible for protection. Nonetheless, whenever AI uses—directly or indirectly—materials already protected by copyright, the responsibility for the proper source management and citation falls on the users of these tools. Therefore, verification by scientists is essential to ensure compliance with the norms of the scientific community and international law.

In this context, the European Regulation on Artificial Intelligence (EU Artificial Intelligence Act) establishes fundamental principles for the responsible use of AI, which are particularly relevant in the scientific domain [79]. The AI Act mandates that users be informed whenever content is created or modified by a generative system. For high-risk applications, the regulation also requires traceability between the data and the process, including explicit indication of the sources used. This has direct implications for scientific production: users employing AI tools for writing or analysis must ensure compliance with third-party rights by properly citing and contextualizing the original information.

The inclusion of AI in research activities must comply with three fundamental principles: transparency, requiring clear disclosure of AI tool usage; accuracy, entailing critical verification and integration of content with reliable, correctly cited primary sources; and legal and ethical responsibility, which remains with the research group, solely accountable for the published results and content.

The authors have made every effort to report transparently all sources used in drafting this document, fully recognizing the intellectual property of original authors. Content produced with AI assistance was verified to ensure compliance with copyright and responsible AI usage regulations. Graphic elements such as tables and images were created directly by the authors using only available support systems.

5. Conclusions

Regarding the history of asbestos, AI generated the following conclusions: “The global history of asbestos is a cautionary tale of industrial advancement outpacing public health protections. Although its extraordinary properties fostered widespread use across civilizations and industries, asbestos ultimately left a devastating legacy of disease, litigation, and socio-political conflict. National experiences varied, influenced by economic dependency, regulatory courage, public pressure, and legal frameworks.

Addressing these gaps requires sustained international collaboration, stronger occupational health policies, and active support for affected communities. Learning from past mistakes, the international community must continue to promote the total eradication of asbestos-related risks, ensuring that economic interests never again outweigh fundamental human rights to health and safety. Only by acknowledging this complex history can future generations fully break free from the enduring shadow of asbestos”.

In addition to the AI-generated conclusions, it can be asserted that despite its limitations, the application of AI models significantly reduced the time required for data collection and preliminary phases, functioning as a heuristic tool to identify critical nodes and facilitate a global comparative analysis of asbestos history and national regulations. In this sense, AI provides valuable support in the early research stages if employed methodically and critically, enhancing rather than replacing human judgment and improving the ability to navigate a vast amount of scientific information [80].

In this work, we deliberately focused on historical and regulatory data to test the strengths and weaknesses of free AI generative tools, with the aim of developing a method applicable to unresolved questions related to asbestos-associated diseases. These include the biological activity of asbestiform non-regulated fibers (e.g., fibrous antigorite, fluoro-edenite, and erionite), the role of shorter fibers (<5 μm) in carcinogenicity, and the impact of ingested asbestos in extrapulmonary diseases [81,82,83,84]. Moreover, our approach can be extended through collaboration with experts focusing on other environmental pollutants, including microplastics, per- and poly-fluoroalkyl substances, glyphosates, heat waves, and electromagnetic pollution, to address similarly complex interdisciplinary challenges. This approach would enable the exploration of possible synergistic and antagonistic effects of multiple pollutants on human health, contributing to the advancement of the so-called “pollutome” methodology [85]. Such applications could particularly benefit from more academically oriented AI generative tools, such as Scite and Consensus, including their premium versions.