ABSTRACT
Generative artificial intelligence (GenAI) has started transforming the tourism industry with wide research implications. While recognising its transformative power, tourism literature failed to identify the dark side of GenAI. Using advanced image analytics across 10 tourism destinations, this research investigates how GenAI-generated images reinforce tourism destination stereotypes. Our findings reveal that GenAI tends to generate highly homogenised images, which cannot fully capture the diversity of destinations, leading to stereotypes. This study advances extant tourism literature by providing critical insights into the complex relationships between generative artificial intelligence and tourism.
Introduction
Generative artificial intelligence (GenAI) has led to some of the most exciting moments for destination image projection in tourism destinations. GenAI-generated images are synthesised by neural networks that assimilate rules from copious image datasets (Lu et al., Citation2023). These AI systems generate vibrant and rich visual imagery in alignment with users’ envisaged concepts, using either image descriptions or stochastic noise as inputs (Miao & Yang, Citation2023). The generated images are highly customizable, creative and have a large amount of unique visual content compared to traditional images (Lyu et al., Citation2022), which can significantly increase sensory intensity and visual working memory capacity (Keogh & Pearson, Citation2014).
With the ability to generate impactful destination images and videos swiftly and extensively, GenAI has revolutionised tourism marketing, offering a myriad of possibilities. In tourism, GenAI image generation tools, such as Midjourney and Stable Diffusion empower Destination Management Organisations to realise their creative intents to generate tailored visual content (Miao & Yang, Citation2023). A notable example is Visit Finland, which incorporated GenAI in a campaign showcasing the authentic Finnish way of life (Gonzalo, Citation2023). While such excitement certainly fuels tourism researchers and practitioners and holds great promise in tourism marketing (Miao & Yang, Citation2023), GenAI-generated images have its own limitations. Hsu et al. (Citation2024) suggest that due to the bias of input data (e.g. relying on user-generated content), GenAI is unable to fully present the nuances of social-cultural diversity of a destination. These biases, unfortunately, may lead to tourism destination stereotypes.
Destination stereotypes, as an academic concept, refer to the general and oversimplified beliefs and judgments held by individuals or groups about a place (Karri & Dogra, Citation2023), formed based on a set of prominent and specific attributes of the destination (Shen et al., Citation2019). For instance, Paris is often seen as a city of romance, while Egypt is known for its pyramids and deserts. Through exposure to media, word-of-mouth and other sources of information, tourists often have preconceived notions and associations with a destination. These provide a shortcut for tourists in processing complex destination-related information (Karri & Dogra, Citation2023).
As GenAI is based on large amounts of existing destination image data, it is likely that the GenAI-generated destination image will include prevailing stereotypes about specific destinations, replicating or even intensifying these images. Against this backdrop, using image analytics, this research is the first to empirically investigate how GenAI-generated images reinforce the stereotypes of tourism destination image, providing critical insights into the complex relationship between GenAI and tourism.
Research design
AI-generated images were analysed using image analytics (Zhan et al., Citation2024). presents the analytics progress. Paris, London, Barcelona, Rome, Bangkok, Beijing, Tokyo, Sydney, New York and Istanbul have been selected as our destinations in this study because they are among the world's most popular tourism destinations. DeepAI (https://deepai.org/machine-learning-model/text2img) was used to generate 10 AI images each for the 10 tourism destinations. The rationale behind choosing DeepAI was its popularity and its ability to produce images that are more realistic and freer from excessive caricature-like elements than other platforms such as GPT4 and Midjourney (see ).
Figure 1. Research design.
Figure 2. Comparison of AI images generated by DeepAI, GPT4 and Midjourney.
Google Cloud Vision API was used to detect labels within these images by identifying ‘general objects, locations, activities, animal species, and products of a given picture’ (Google, Citation2023, p. 1). To assess the similarity between these images based on labels, we used the Doc2Vec model (Kim et al., Citation2019). The cosine similarity value (cosθ) shows that a perfect match between each photo label vector results in a cosθ of 1, indicating identicalness, while complete dissimilarity yields a cosθ of – 1. For training the Doc2Vec model, several parameters were pivotal: the size of the vectors, the window size and the number of iterations over the training corpus. After testing various configurations and assessing the effectiveness of each model in finding similar images, we determined the optimal settings to be a vector dimension of 20, a window size encompassing two words, and a total of 100 iterations. This particular configuration demonstrated superior performance compared to the alternatives.
Finally, by utilising the cosine similarity matrix, we calculated the similarity in semantic content across the images to indicate the tourism destination stereotypes.
Results
GenAI-generated images from DeepAI for the 10 tourism destinations present distinctive similarities, with the similarity score among the 10 images of each destination reaching 0.99 (with 1 being the maximum). For example, in the analysis of GenAI-generated images of Paris, the top 10 labels are: building (7), sky (7), vehicle (6), urban (5), window (5), architecture (5), façade (5), design (4), wheel (4), travel (4). The cosine similarity matrix of the labels between each image exhibited values greater than 0.99 (). This high level of similarity indicates close similarities between each image (see the Appendix for the cosine similarity matrix for 10 destinations).
Table 1. An example of the cosine similarity matrix of GenAI-generated images for Paris.
Moreover, the landmarks depicted in the images of each destination align with popular locations or iconic sites. gives two examples of GenAI-generated images for 10 destinations, showing similar images along with their corresponding similar labels. These examples demonstrate that GenAI-generated images for each destination exhibit a high degree of similarity, which confirms the presence of destination stereotypes. For instance, Paris is represented by the Seine River and the Eiffel Tower, while London features red buses and European-style architecture. Barcelona is characterised by its unique modernist architecture. Rome's images focus on ancient Roman structures, Bangkok on temples, Beijing on the architectural grandeur of the Forbidden City, Tokyo around the Tokyo Tower, Sydney revolves around the iconic Sydney Opera House, New York is epitomised by Times Square, and Istanbul showcases its distinctive Turkish architectural landmarks.
Table 2. Two examples of each AI-generated destination images generated by DeepAI.
Conclusion
Addressing the call of Hsu et al. (Citation2024) for practical applications of GenAI in tourism, this study examines how GenAI contributes to the reinforcement of tourism destination stereotypes. While studies have highlighted the positive and creative role of GenAI in bypassing creative workers and increasing tourists’ sensory strength and visual memory, this study reveals that GenAI tends to generate highly homogenised images, which cannot fully capture the diversity of destinations and subsequently lead to tourism destination stereotypes.
This research highlights the need for ongoing discussions surrounding the deployment of GenAI in shaping destination images. Previous research has highlighted that bias in training data may lead to unjust, unfair and prejudicial results (Akter et al., Citation2021). This research provided further evidence, calling for more inclusive approaches in the training data to have local voices and perspectives reflect the diversity of tourism destinations. This research also emphasises transparency and accountability to be aware of the sources of AI-generated content and the potential biases involved. As GenAI continues to play a prominent role in influencing societal narratives, it becomes imperative to critically assess and address the implications of AI-generated content on the perpetuation or challenge of tourism destination images. Here we present a small step to invite future research and debate on unpacking the complex relationships between GenAI and tourism.
Supplemental Material
Download MS Excel (28.2 KB)Disclosure statement
No potential conflict of interest was reported by the author(s).
References
- Akter, S., McCarthy, G., Sajib, S., Michael, K., Dwivedi, Y. K., D’Ambra, J., & Shen, K. N. (2021). Algorithmic bias in data-driven innovation in the age of AI. International Journal of Information Management, 60, 102387. https://doi.org/10.1016/j.ijinfomgt.2021.102387
- Gonzalo, F. (2023). How generative AI is transforming travel marketing. https://fredericgonzalo.com/en/how-generative-ai-is-transforming-travel-marketing/
- Google. (2023). https://cloud.google.com/vision/docs/labels
- Hsu, C. H. C., Tan, G., & Stantic, B. (2024). A fine-tuned tourism-specific generative AI concept. Annals of Tourism Research, 103723. https://doi.org/10.1016/j.annals.2023.103723
- Karri, V. R. S., & Dogra, J. (2023). Destination stereotypes: A phenomenon of destination image. Journal of Hospitality and Tourism Insights, 6(3), 1290–1308. https://doi.org/10.1108/JHTI-03-2022-0111
- Keogh, R., & Pearson, J. (2014). The sensory strength of voluntary visual imagery predicts visual working memory capacity. Journal of Vision, 14(12), 7–7. https://doi.org/10.1167/14.12.7
- Kim, D., Seo, D., Cho, S., & Kang, P. (2019). Multi-co-training for document classification using various document representations: TF–IDF, LDA, and Doc2Vec. Information Sciences, 477, 15–29. https://doi.org/10.1016/j.ins.2018.10.006
- Lu, Z., Huang, D., Bai, L., Liu, X., Qu, J., & Ouyang, W. (2023). Seeing is not always believing: A quantitative study on human perception of ai-generated images. arXiv preprint arXiv:2304.13023.
- Lyu, Y., Wang, X., Lin, R., & Wu, J. (2022). Communication in human–AI co-creation: Perceptual analysis of paintings generated by text-to-image system. Applied Sciences, 12(22), 11312. https://doi.org/10.3390/app122211312
- Miao, L., & Yang, F. X. (2023). Text-to-image AI tools and tourism experiences. Annals of Tourism Research, 102, 103642. https://doi.org/10.1016/j.annals.2023.103642
- Shen, X., Lv, X., Lin, S., & Li, C. (2019). Application of the stereotype content model to destination image: Evidence from residents of mainland China. Journal of Destination Marketing & Management, 14, 100375. https://doi.org/10.1016/j.jdmm.2019.100375
- Zhan, L., Cheng, M., & Zhu, J. (2024). Progress on image analytics: Implications for tourism and hospitality research. Tourism Management, 100, 104798. https://doi.org/10.1016/j.tourman.2023.104798