The generation of tourism knowledge by language model-based AI: A call for a more critical perspective

1 Introduction

ChatGPT is currently the most popular AI application in the category of Generative Pre-trained Transformers (GPT) for ‘text-to-text’ tasks (Gozalo-Brizuela and Garrido-Merchan 2023). This AI solution generates high-quality natural language responses to nearly all types of text-based queries. Recently, numerous publications have explored the use of ChatGPT in academic settings, particularly in education (e. g., Cotton et al. 2024; Rudolph et al. 2023). Additionally, there is a rapidly growing body of literature examining AI-based applications in the tourism and service industries (Carvalho and Ivanov 2023; Tuo et al. 2024), as well as in specific areas such as consumer experience creation (Ghesh et al. 2024; Sharma and Aggarwal 2024), hospitality industry (Yin et al. 2023) or cruise tourism-related services (Gonzáles-Santiago et al. 2024).

However, there remains a scarcity of literature on the application of GPT models in the social sciences research process, particularly in tourism research, which often relies on text-based empirical studies. Nautiyal et al. (2023) discuss issues such as truth, plagiarism, publishing ethics, interdisciplinarity, and equity, calling for a broader conversation beyond the issue of plagiarism. Skavronskaya et al. (2023) examine aspects of academic integrity and ethics, particularly AI-related plagiarism, and advocate for adapting AI-integrated pedagogies in tourism education. Ivanov and Soliman (2023) highlight the benefits of ChatGPT in tourism education and scientific writing, using examples from publications on robots in tourism. Shin and Kang (2023) take this a step further in the same journal, conducting an ‘interview’ with ChatGPT consisting of 15 questions on the evolution, current state, and future directions of tourism research. They conclude that ChatGPT offers promising topics for future research, such as health, social justice, climate change, and governance/policy issues. They also note that ChatGPT proposed novel methodologies, such as participatory action research and mixed-method systematic reviews, which are not commonly used in current tourism research. Rigin et al. (2025) examine the implications of using generative artificial intelligence (AI) models, such as ChatGPT and Bard, in social science research and conclude: “generative AI can enhance data collection and analysis, offering alternative approaches to traditional research methodologies” (p. 284). This raises the question if these positive research support assumptions are based on evident tests or just assumptions expressing the hope for reducing the amount of research work needed when following traditional research practices?

2 Analytic approach to analyze ChatGPT’s research capabilities

A common theme across these publications discussing the potential of AI and more specifically ChatGPT is the absence of doubts regarding the ability of GPT applications, to meet the fundamental requirements of the scientific community, as these are validity, reliability, and objectivity, that have been demanded of researchers for centuries. This paper aims to analyze and discuss the potential of ChatGPT as a tool for research. To this end, the author tested ChatGPT on several typical research-relevant topics commonly encountered in tourism research using the releases 3.5 (database 09/2022), GPT-4o (10/2023) and finally GPT-5 (10/2024). First, the completeness of resources of the knowledge base of ChatGPT was investigated. Afterward, the reliability of text analysis based on 31 paper abstracts was tested. Finally, the quality of picture content interpretation and picture creation using DALL-E were checked.

3 The role of scientific publications in the ChatGPT knowledge base

In a first step we asked ChatGPT about the role of scientific publications for tourism research when creating its latest release of the knowledge base (detailed question and answer by ChatGPT see Appendix).

Despite the awareness that the responses generated by ChatGPT are the product of a language model, clear limitations became evident. OpenAI’s developers are fully aware of the current constraints of ChatGPT, and more broadly, of any AI-based Generative Pre-trained Transformer (GPT) concerning both the knowledge base and the algorithmic training. The effectiveness of the knowledge base and the performance of the language model are contingent upon the quality of the data collected and the training process employed.

OpenAI’s GPT solutions are built on a three-step procedure (OpenAI 2022): In the first step, relevant data that is not yet part of the knowledge base is selected, and supervised training is conducted to produce the desired output. In the second step, similar data from the same subject area is used to generate responses based on the initial training. The generated responses are then compared, and their quality is ranked by a ‘labeler.’ This quality ranking informs a second round of training. The third step involves a continuous automated loop governed by a Proximal Policy Optimization (PPO) algorithm, a reinforcement learning technique used for optimizing decision-making and control policies in machine learning. It’s important to note that PPO is unrelated to the architecture or version of ChatGPT itself. These algorithms require specific parameters, such as threshold values, to initiate the feedback loop. A critical issue in this process is the definition of penalties for offensive or toxic language, repetition, or overly sentimental responses.

This three-step procedure necessitates human interaction with the AI. Initially, a human operator must define the ‘desired output behavior,’ which requires deep knowledge of the topic described by the input data. This introduces the potential for bias due to the operator’s knowledge gaps—the so-called ‘labeler’ effect. Acerbi and Stubbersfield (2023) described an additional risk of bias arising from the preselection of training sources and the cultural background of the training personnel, which can lead to human-like content biases, especially in sociological or behavioral contexts. Their research demonstrated that ChatGPT reproduces gender stereotypes analogous to those found in human societies, indicating that the concept of a ‘desired output’ during training may have been biased by the personal positions of the trainers. It is well established that, for many topics, there is no universally accepted normative stance; the desired output often reflects cultural or political positions. Given the cultural value differences observed across regions and countries (Schwartz and Cieciuch 2021), it is likely that the cultural background of the ‘labelers’ influences the model’s output. Moreover, as topics become more specialized, the need for extensive background knowledge among those training the system becomes more pronounced. There are clear limitations to this without the use of secondary and particularly scientific sources. The closer a topic is to ongoing research or recently published findings, the less likely it is that the ‘labeler’ can objectively assess the ‘desirability’ of an answer without access to scientific publications.

In the second step, an individual must evaluate the quality of the system’s alternative outputs, assessing both linguistic quality and context-related content quality. This judgment is subject to the same biases as those encountered in the first step. Finally, in the third step, the PPO algorithm requires parameter settings to calculate the reward or penalty values that define the loop for improving output quality. This step also requires input from a person programming the algorithm—the ‘labeler’.

Examining these three steps, it becomes clear that each is governed by human intervention. This introduces various correlations between the output of the GPT-system and the characteristics of the individuals involved in its training, such as their professional background, political affiliation, religion, culture, and gender. Expanding the knowledge base to include more complex topics, as is common in research, necessitates a higher level of expertise from those participating in the training process.

4 Findings of text and picture analysis tests

5 Discussion

The results of the conducted tests indicate that, at present, ChatGPT does not produce reliable, complete, and therefore valid outcomes when performing two relatively simple tasks involving text analysis. This underscores that such AI tools cannot replace the careful and thorough work of researchers. The incomplete analysis of texts when searching for a clearly defined object category, such as geographic locations mentioned in connection with overtourism, clearly demonstrates the limitations of the tool’s usability. In scientific work, a reliable and complete analysis of texts or data is essential as the basis for interpretation. If this basis is incomplete, any interpretation in relation to a research question will necessarily remain incomplete as well.

With the enhanced individualization of dialogue management in versions 4o and 5, there is a considerable risk that responses will diverge significantly between two researchers who interact with ChatGPT in different communication styles. From the perspective of result reliability, this represents a step backward. It remains unclear to what extent the individualization of AI responses, in combination with its ability to recognize personal background characteristics of researchers such as geographic origin, gender, academic training, and the capacity to formulate precise research questions, may influence the principle of objectivity. When AI begins to exert suggestive influence on the researcher in an effort to establish a personal relationship through dialogic interaction, and in doing so appears to objectify the researcher’s personal views, there is a risk that the research process may lose its foundation in evidence-based reasoning and thereby compromise its objectivity.

Even more concerning are recent journal publications that discuss the potential for plagiarism or the use of ChatGPT to structure future research proposals. As ChatGPT reports itself its knowledge base is neither up-to-date nor grounded in original academic sources. As a result, much of the existing scholarly work in the tourism knowledge system, as described by Tribe and Liburd (2016), is either ignored or only superficially incorporated through citations from unclear and frequently secondary sources. If ChatGPT’s claim that it has rarely used scientific publications to create its knowledge base is true, then plagiarism in the strict sense is not possible.

However, the responses generated by ChatGPT, particularly regarding methodological deficits and improvement options, should be viewed with even greater skepticism. For example, a statement about the methodologies found in the analyzed abstracts like “participatory action research and mixed-method systematic reviews, which are not commonly used in current tourism research,” is a prime example of AI hallucination, as the knowledge base lacks any substantive insight into the methods used in tourism research. The judgment “not commonly used” maybe was found in a blog or a discussion on Trinet, but it cannot be the result of a comprehensive analysis of thousands of scientific articles.

6 Conclusions

In general, progress in science is a gradual process. The generation of tourism knowledge by researchers typically begins with the observation of phenomena in the real world of tourism that demand explanation. Tourism, being an interdisciplinary field, draws from various disciplines, which is why Tribe (1997) referred to it as the “indiscipline of tourism”. The tourism knowledge system integrates multiple fields, such as social sciences, business studies, humanities, arts, and others (Tribe and Liburd 2016). Scholars from different disciplines, each with its own terminology and research traditions, must collaborate to generate different types of knowledge, such as problem-centered or value-based knowledge. This knowledge is eventually represented, disseminated, and stored through various types of publications: conference papers, books, and journal articles.

The exchange of knowledge among researchers, particularly those from different disciplines, requires precise and comprehensible descriptions of research, necessitating a common cross-disciplinary understanding of the terminology used (Belhassen and Caton 2009). Therefore, it is a standard practice in research to apply precise citations of reviewed publications that represent proven knowledge relevant to a particular problem. The traceability of knowledge evolution, along with the context, discipline, location, timing, and authorship of its generation, are key principles in research. This is what differentiates a text presenting mere assertions from one that is grounded in a consistent, yet not always traceable, set of knowledge-based arguments. Moreover, contributions to the tourism knowledge system require critical discourse and independent validation, including the possibility of falsification by other researchers. I advocate for a more critical and realistic reflection on what ChatGPT specifically, and Generative AI models in general, can contribute to tourism research.

Based on this analysis, it is crucial to distinguish between either using ChatGPT as a tool supporting the researcher or relying on it to generate new results for the tourism knowledge system. The use of AI as a tool can be seen as an opportunity in highly standardized tasks that are detached from the substantive engagement with the research question can offer support and potentially accelerate time-consuming steps. Examples include scripts for data preparation, implementation and programming of complex statistical analyses, formatting output tables in a specific style for publication, and checking texts for correct spelling and grammar. However, effective interaction with AI requires a sound understanding of all work steps, as precise use is otherwise not possible. The example of programming scripts for statistical analyses in R clearly illustrates the need for strong knowledge of R command syntax. Only if the researcher provides the AI with the correct variable names for the analysis and is able to specify which variables are dependent and which are independent in multivariate analyses can the AI be used effectively.

Moreover, only the researcher can provide the substantive interpretation of the results in light of the underlying research questions and hypotheses. The risks associated with the use of AI in tourism research lie primarily in the area of wrong or incomplete interpreting data or research findings in relation to specific research questions. On the one hand, the tests have shown that there is no guarantee of completeness or reliability. In addition, the knowledge base regarding existing research findings is incomplete due to the training techniques used in AI development. And finally, the new highly individualized dialogue system increases the risk of a lack of objectivity of researchers following the suggestions of the AI.

While AI-systems are undeniably fascinating, as researchers, we must maintain a critical distance and first ask, ‘What is behind this, and how does it work?’ Just as we approach statistical analysis with a careful selection of the most appropriate methods based on the research question and data type, we must apply the same rigor to AI. First, we need to understand how AI works and assess the adequacy of its knowledge base before deciding to use it for research purposes. Understanding that the current knowledge base of ChatGPT is widely disconnected from scientific publications should prevent any researcher from using it to generate outputs related to the research meta-level, as its output is merely a hallucination.

7 Limitations

This methodological approach has limitations, as the results are based on a specific set of questions and tasks asked to ChatGPT. A different way of asking and related individualized dialogues might lead to other results. This also applies to the selected set of abstracts as basis for text analysis. In another field of research, using different text, other outcomes might be observed. The fact that results may vary in quality and type depending on the questions and input data already highlights the general challenge researchers face when considering validity, reliability, and objectivity. The issues discussed in this paper raise a more fundamental question regarding the development of quality criteria for scientific work involving AI. It is not yet clear what such criteria should entail or how they could be implemented within AI models. Addressing this challenge will require future research and development efforts. This aspect currently remains an open and unresolved issue that deserves further exploration.