When a physician uses OpenAI's Whisper transcription tool and the software converts a routine observation about a patient wearing a necklace into a fabricated account of a terror attack, the incident is easy to dismiss as a software glitch. For Timnit Gebru, one of the world's most prominent AI ethicists, it is something far more serious: a symptom of a fundamentally broken approach to building artificial intelligence that is failing communities from Lagos to Lahore.
"We've now been pushed to a paradigm that is ridiculous, that is never going to be safe because you don't have a well-defined task," Gebru said in remarks published by the AI Now Institute. The Whisper error she cited — transcribing "I think he was wearing a necklace" as "He was holding a terror knife and he killed a bunch of people" — is not, she argues, an anomaly. It is the predictable output of systems built to do everything at once, for everyone, everywhere.
A Monoculture With Global Consequences
The critique Gebru and her colleagues at the Distributed AI Research Institute (DAIR) are advancing goes well beyond concerns about hallucinations or bias. At its core, it is an argument about market structure and global equity. The headlong rush to build ever-larger, all-purpose models — pursued by a handful of American and Chinese technology giants — has, they contend, systematically crowded out the specialised, resource-efficient AI tools that previously served diverse communities around the world.
"This idea that we can just use one way of doing things for everything in the world, one giant model for everything has introduced problems we didn't even have before, and also results in subpar tools for many people around the world," Gebru said.
The linguistic dimension of this problem is acute. Of the roughly 7,000 languages spoken globally, the vast majority are severely underrepresented in the training data used by leading AI systems. Languages spoken by hundreds of millions of people across Africa, South and Southeast Asia, and Latin America — from Swahili and Hausa to Bengali and Quechua — receive a fraction of the attention lavished on English. The result is that AI-powered tools for healthcare, education, legal services, and commerce perform markedly worse for speakers of these languages, compounding existing inequalities.
When Big Tech Announcements Kill Small Companies
The market dynamics are particularly punishing for the startups and research organisations attempting to fill this gap. Gebru describes a pattern in which announcements from OpenAI or Meta of new multilingual models directly trigger investor withdrawals from smaller, community-focused AI ventures. "A number of potential investors in these smaller organisations literally told them to close up shop," she explained.
This consolidation effect is not limited to the United States. Across Africa, researchers and entrepreneurs building natural language processing tools for local languages have reported similar pressures. In India, where the government has invested in indigenous AI initiatives such as the Bhashini platform, the challenge of competing with resource-rich American models remains formidable. In Brazil, efforts to develop Portuguese-language AI tailored to local idiom and context face the perpetual gravitational pull of English-centric systems.
The competitive advantage enjoyed by the large players is self-reinforcing. As Gebru frames it, the ability to accumulate vast datasets — acquired through practices that critics describe as large-scale appropriation of online content — and to outspend rivals on GPU infrastructure and data centres, are treated by the industry not as ethical liabilities but as core strategic assets. "Industry has absolutely no incentive to look at less resource-intensive things because they view their stealing of data as a competitive advantage," she said.
The Regulatory Battleground: From Sacramento to Brussels
The political response to these dynamics is fragmenting along geographic lines, producing a patchwork of regulatory approaches that reflects sharply divergent visions of how AI development should be governed.
In California, Governor Gavin Newsom's veto of SB 1047 — legislation that would have imposed meaningful safety obligations on large AI developers — was widely interpreted as a victory for incumbent technology companies and a signal that comprehensive oversight remains politically contested in the jurisdiction that is home to most of the world's leading AI firms.
The contrast with Europe is stark. The European Union's AI Act, which entered into force in 2024 and is being phased in progressively, represents the most ambitious attempt yet to impose binding obligations on AI developers operating at scale. It establishes risk-based categories, mandates transparency requirements, and places explicit restrictions on certain high-risk applications. European regulators have made clear that the Act applies to any AI system deployed in Europe, regardless of where it is developed — a jurisdictional reach that puts American and Chinese developers directly in scope.
Elsewhere, approaches vary considerably. China has implemented its own AI governance framework, focused in part on algorithmic recommendation systems and the management of generative AI content. Several Gulf states are actively courting AI investment while developing parallel regulatory structures. India is proceeding cautiously, wary of stifling a sector it sees as central to its economic ambitions. In much of the Global South, formal AI regulation remains nascent, raising concerns that populations with the most to gain from well-designed AI — and the most to lose from poorly designed systems — have the least input into how those systems are built and governed.
A Decentralised Alternative Takes Shape
Against this backdrop, a loosely affiliated international movement is coalescing around a different model. Its participants include academic researchers, civil society organisations, open-source developers, and policymakers who argue that the problems Gebru identifies are not technical inevitabilities but the products of specific choices — choices that can be made differently.
The movement's priorities include: investment in specialised, task-specific AI systems that are more auditable and less prone to catastrophic failure; development of AI tools built with and for specific linguistic and cultural communities rather than imposed upon them; open-source alternatives to proprietary models that allow scrutiny and local adaptation; and governance frameworks that give affected communities meaningful standing in decisions about AI deployment.
Initiatives such as Masakhane, a pan-African NLP research community, and AI4Bharat, focused on Indian languages, exemplify this approach. The BigScience workshop, which produced the open multilingual BLOOM language model through a globally distributed collaboration, demonstrated that large-scale AI development need not be the exclusive preserve of well-capitalised corporations.
Whether this alternative ecosystem can survive the gravitational pull of Big Tech's resources and reach remains an open question. What is clear is that the choices made in the next few years — in boardrooms, legislatures, and research labs — will determine not just the technical trajectory of AI, but who gets to participate in the world it is reshaping, and on whose terms.
Sources:
1 News Report, "Frugal AI"
2 News Report, "Linguistic Diversity"
3 News Report, "The Impact of AI in Education: Navigating the Imminent Future"