What can AI see? The image of the ‘migrant’ in the era of AI post-visualization

ABSTRACT

Over the past decade, artificial intelligence (AI) has become omnipresent in migration control and mobility surveillance, with AI systems now deployed across all aspects of migration management. Critics of such trends typically examine questions of ethics and rights from the vantage point of regulatory mechanisms and the limited venues for the redress of grievances. But if legal frameworks are as of yet forthcoming and do not necessarily apply to migrants, are there alternative mechanisms to critique algorithmic decision making? To explore this and related questions, this paper engages one such alternative by taking a ‘visual turn.’ In asking ‘what can AI see’ the paper interrogates the role of images in constructing AI's capacity to both understand migration and make appropriate decisions about migrants. In addition, a visual turn allows for exploration of an emergent age of post-visualization: a phenomenon whereby the values and meanings of what we see will be increasingly imparted to us by AI systems. The paper examines what AI sees with the help of an experiment, prompting an AI generative platform to draw distinctions between migrants, refugees, and people.

Acknowledgement

The author is grateful to Natasha Saunders, Alex Sager and the journal’s editors for their review and comments on earlier drafts. My deepest gratitude to Akanksha Narode and Iva Gumnishka for many spirited conversations about AI.

Notes

1 The DALL-E dataset was compiled between August 2022–August 2023. The complete set is available in the Supplemental Materials section appended to this paper.

2 For the steps in machine learning see (Theobald 2017).

3 The experiment is also limited because there are differences between generative AI systems (such as DALL-E) and discriminative AIs, typically deployed in migration management. However, because the paper’s main focus is on how AI conceptualizes objects and people without the benefit of context and the complex systems of values and beliefs typically associated with images, the images generated by DALL-E suffice in showcasing how and where AIs might be in the process of learning.

4 https://laion.ai/blog/laion-5b/

5 For the purposes of this paper, I only review work generated on DALL-E. Initial comparisons between images of ‘migrants’ produced by DALL-E, Craiyon, and Midjourney were indistinguishable.

6 Personal interviews with AI experts.

7 Prompts for ‘aliens’ only produce green figures hence not discussed here.

8 On the complexities of text-to-image mapping, see also (Rassin, Ravfogel, and Goldberg 2022).

9 See (Christian 2020) for further explication of the issue.

10 Observation from a social experiment conducted in Singapore, March-April 2023.

Additional information

Notes on contributors

Marina Kaneti

Dr. Marina Kaneti is Assistant Professor in International Affairs at the Lee Kuan Yew School of Public Policy, National University of Singapore. She draws on visual methods in order to explore questions of global development, including migration, rights, and environmental governance. Dr. Kaneti's work appears in a number of peer reviewed journals, including Citizenship Studies, Human Rights Review, and Sustainable Development. She is currently working on two book manuscripts. The first explores the politics of migrant inclusion. The second interrogates the nexus of global mobility and ancient conceptualizations of hospitality.