The paper, titled “Decoding the societal and technical challenges of Artificial Intelligence: a comprehensive transdisciplinary approach,” was accepted on 20 May 2026 and published on 3 September 2026. Its central claim is deceptively simple but far-reaching: artificial intelligence is neither the utopian liberation promised by the techno-optimists nor the fatalistic doom feared by the techno-pessimists. Rather, the authors describe AI as a “diachronic dialectical continuum,” meaning that its character, trajectory, and distribution of benefits and harms reflect the social organization, property relations, and democratic arrangements of the societies that produce and govern it. Where those arrangements are unequal, the technology absorbs and amplifies that inequality.

To build this argument, the authors deploy what they call a transdisciplinary materialist framework, synthesizing insights from science and technology studies, political economy, philosophy, and what historians call the longue durée, the long-run history of technology stretching from stone tools through the industrial revolutions to the present. This is not merely an academic exercise in breadth. The framework allows the authors to treat seemingly separate phenomena, such as the mining of critical raw materials, the concentration of semiconductor fabrication, the exploitation of data-labeling labor, and the capture of AI governance by private interests, as dialectically interconnected moments of a single, historically determined techno-societal system. Each element feeds the others; none can be understood in isolation.

The material foundations of the AI conjuncture, as the authors term it, begin with physical infrastructure. The training and deployment of large-scale machine learning models depend on monopolized computational infrastructure, on the extraction of minerals such as those used in advanced chips, and on a semiconductor fabrication and GPU ecosystem concentrated among a handful of state-subsidized corporate actors. The paper points to the extraordinary market dominance of graphics processing units as evidence that the AI economy is not a democratized, distributed commons but a tightly held industrial complex. Projections cited in the article suggest the leading chipmaker could reach a market capitalization measured in the trillions of dollars, a scale of concentration that few industries in history have matched.

Equally central to the analysis is labor. Behind the polished interfaces of generative AI systems lies a global division of work that includes highly paid engineers at one pole and, at the other, precarious data annotation and content-moderation workers in the global South who perform the repetitive tasks that make machine learning possible. The authors frame this as part of a longer pattern of what scholars have called data colonialism, the appropriation of human life and knowledge as raw material for capital accumulation. AI, in this reading, is less an alien intelligence than a privatization of collective human knowledge, a genealogy the paper traces through the social history of computing.

The geopolitical dimension of the study is equally pointed. Drawing on world-systems analysis, which maps the relationship between core and peripheral regions of the global economy, the authors argue that the AI economy reproduces the asymmetric exchange patterns of earlier colonial eras. Computational resources, patents, and profits concentrate in the core, while peripheral geographies supply raw materials, labor, and data, and receive comparatively little of the value generated. China emerges as a notable exception to this pattern, pursuing a state-coordinated AI strategy that includes international cooperation initiatives and algorithmic recommendation regulations, a counterpoint to the market-dominated model of the United States and, more falteringly, Europe with its AI Act.

The paper is also a critique of how AI has been governed, or rather not governed. The authors document what they describe as the structural capture of AI governance by private interests, in which the corporations building the technology largely set the terms of its regulation. They highlight the phenomenon of “ethics washing,” the strategic use of ethical principles and advisory boards to forestall binding rules, and contrast the proliferation of soft-law frameworks, from OECD recommendations to UNESCO’s ethics declaration, with the weakness of enforceable international coordination. Against this backdrop, the paper notes proposals for institutions such as a G20 coordinating committee for AI governance, while stressing that meaningful regulation requires confronting the underlying property relations, not merely the outputs of biased algorithms.

Bias and accountability receive rigorous technical and social treatment. The study reviews the empirical literature demonstrating that machine learning systems absorb and amplify social prejudice: word embeddings encode gender stereotypes, commercial facial-recognition systems show sharply divergent error rates across skin tones and genders, and image generators produce racist and sexist outputs. The authors emphasize that these are not accidental glitches to be patched but predictable consequences of training systems on data drawn from unequal societies and deploying them through concentrated, opaque infrastructures. Algorithmic opacity, the “black box” problem, compounds the difficulty, since the internal reasoning of deep learning systems resists the transparency that accountability demands.

What distinguishes this study from much of the crowded AI ethics literature is its refusal of both dominant emotional registers. The authors explicitly position their argument against the techno-optimist utopianism associated with Silicon Valley manifestos promising abundance and singularity, and equally against the existential fatalism of those who warn that superhuman AI will inevitably destroy humanity. Both framings, they contend, depoliticize the technology by treating its future as predetermined by technical inevitability rather than as the outcome of contestable social choices. Historical perspective supports this view: the benefits of past general-purpose technologies, from electricity to computing, were distributed according to struggles over labor, institutions, and policy, not according to any intrinsic logic of the machines themselves.

The implications of the paper extend to labor markets and development. Citing economic research on automation and employment, the authors note that AI-driven automation both displaces existing tasks and creates new ones, with the balance determined by institutional context rather than technological necessity. Estimates of AI’s macroeconomic impact, including analyses from international financial institutions suggesting that a substantial share of global employment is exposed to generative AI, are read not as prophecy but as a measure of the policy choices ahead. For developing countries, the stakes are particularly high, as the paper’s framework of “dissymmetry” implies that without deliberate intervention the AI economy will widen existing gaps in ownership, access, and capability.

Ultimately, the study is a call to see AI as it is: a material system embedded in capitalism, colonial history, and democratic deficit, but also a system that can be reorganized. The authors argue that because AI’s direction reflects the social body that produces it, changing that direction requires changing the underlying relations of property, governance, and participation. Proposals for digital commons, public computational infrastructure, and genuinely transnational governance are treated not as idealism but as structural necessities. As the AI revolution accelerates through smart cities, epidemiology, gene editing, policing, and even warfare, the paper insists that the decisive question is not what machines will do to us, but what kind of society we will build through them.

The study in question scrutinizes multiple AI language models across nine different languages, breaking new ground by extending beyond the commonly examined English-language parameters. Traditionally, much of the research has been limited to English, leaving a gap in understanding how biases manifest in a multilingual context. To bridge this divide, the researchers developed a new framework known as the Multilingual Assessment of Gender Bias in Image Generation, abbreviated as MAGBIG. This framework employs carefully curated occupational terms to assess biases and stereotypes in AI’s image generation processes.

The study categorized prompts into four distinct types: direct prompts utilizing the ‘generic masculine’, indirect descriptions that refer to a professional role in a gender-neutral manner, explicitly feminine prompts, and ‘gender star’ prompts designed for gender neutrality. This approach enables a nuanced examination of how different linguistic expressions influence the AI’s output. Notably, the research took into account languages that possess gendered occupational titles, such as German, Spanish, and French, as well as languages like English and Japanese, which utilize a single grammatical gender yet have gendered pronouns. Moreover, the study considered languages devoid of grammatical gender, exemplified by Korean and Chinese.

Upon examining the output generated from various prompts, the researchers found a consistent pattern: direct prompts employing the generic masculine resulted in the most pronounced gender biases. Particularly in professions typically associated with numbers and authority, like “accountant,” the AI predominantly presented images depicting white males. Conversely, roles associated with caregiving, such as nursing, tended to yield images of women, reinforcing long-standing gender stereotypes. Even gender-neutral options or ‘gender-star’ prompts offered only minimal relief from these biases, while explicitly feminine prompts yielded overwhelmingly female representations.

Interestingly, while utilizing neutral prompt structures appeared to mitigate gender stereotypes, it also resulted in diminished quality regarding the fidelity of the generated images. In essence, while striving for neutrality in prompts, the AI’s overall effectiveness in image generation was compromised. This trade-off raises critical questions for users and developers alike, urging them to consider how the specific wording of their queries can yield profoundly different visual outcomes.

The impact of language on AI image generation raises alarm bells, as articulated by Alexander Fraser, a professor specializing in data analytics and statistics at the Technical University of Munich. He emphasized the crucial role language plays in guiding AI systems, warning that varying phrasing can significantly alter the nature of the images produced, potentially enhancing or mitigating societal stereotypes. This caution is particularly relevant in Europe, where multiple languages coexist and intersect, exemplifying the need for fair AI that accounts for linguistic nuances.

The research also indicates that biases do not uniformly correlate with grammatical structures across languages. For instance, the shift from French to Spanish prompts resulted in a notable uptick in gender bias, despite both languages sharing similar methods for delineating male and female occupational terms. This unexpected divergence signals that underlying cultural perceptions and societal norms may exert a critical influence, irrespective of linguistic grammar.

The implications of these findings extend beyond academic inquiry; they resonate with practical applications in technology, marketing, and entertainment. As AI image generation becomes increasingly integrated into sectors ranging from corporate branding to social media, the potential for bias to shape public perception and reinforce stereotypes necessitates urgent attention. Therefore, stakeholders must advocate for AI systems that not only recognize but also consciously confront and rectify gender biases.

The revelations derived from this study signal a pivotal moment in the intersection of language, culture, and technology. As artificial intelligence continues to permeate various facets of life, understanding the implications of gender bias in AI-generated imagery becomes paramount. Developers and users are urged to embrace a more thoughtful approach to AI interactions, implementing language sensitivity that acknowledges and addresses these biases. As AI evolves, the pursuit of more inclusive and fair algorithms must remain at the forefront of discourse on ethical AI development.

In conclusion, the exploration of AI’s role in perpetuating gender biases lays bare the complexities that arise from the coupling of technology and linguistic frameworks. With AI image generation increasingly shaping societal narratives, stakeholders must commit to creating systems that not only reflect diverse realities but also actively dismantle historically entrenched stereotypes. This imperative endeavors to ensure that the future of AI is grounded in fairness, equity, and representation.

As scholars, developers, and policymakers converge to address these pressing challenges, the essence of the conversation will continually revolve around how our languages shape the technologies we rely on, and in turn, how these technologies reflect and refract the prevailing attitudes and norms that govern our societies.

Subject of Research: Multilingual bias in AI-generated image outputs
Article Title: Multilingual Text-to-Image Generation Magnifies Gender Stereotypes
News Publication Date: 27-Aug-2025
Web References: DOI
References: Not provided.
Image Credits: Not provided.

Keywords

AI, gender bias, language models, image generation, stereotypes, multilingual research