This arXiv preprint (abs/2604.22071), not yet peer-reviewed, offers a contrarian analysis arguing that large language models are symptoms rather than root causes of AI's most pressing failures. The authors use a mixed methodology: a literature synthesis of over 40 prior studies on AI incidents, combined with theoretical modeling of data-to-deployment pipelines and qualitative case reviews of 25 real-world failures spanning content generation, decision automation, and scientific applications. They explicitly note limitations including reliance on publicly documented cases (potentially skewed toward high-visibility Western examples) and the absence of new large-scale empirical experiments.

The paper's central claim is clear: hallucinations, bias amplification, and misuse commonly blamed on LLMs actually originate upstream in data curation practices, economic incentives that prioritize scale over reliability, and governance gaps that allow unchecked deployment. This challenges mainstream hype from both AI critics and boosters. While outlets like The New York Times have framed ChatGPT-era errors as inherent model flaws, they missed the recurring pattern of 'blame-the-artifact' reporting that ignores how these models merely compress and reflect existing societal data ecosystems.

Synthesizing this preprint with Gebru et al.'s influential 'Datasheets for Datasets' (arXiv:1803.09010, peer-reviewed version in Communications of the ACM) and Cathy O'Neil's 2016 book 'Weapons of Math Destruction' reveals consistent throughlines others overlook. Gebru's work demonstrated that undocumented training data inevitably bakes in historical inequities; the new preprint extends this by showing LLMs act as high-fidelity mirrors rather than originators. O'Neil's analysis of opaque algorithms in hiring, policing, and finance maps directly onto current LLM deployments in education and healthcare, where profit-driven companies deploy them at scale without addressing the feedback loops that reinforce inequality.

The genuine insight lies in the overlooked connection: focusing on model size or 'emergent abilities' distracts from systemic rot. Concentrated control of training data by a handful of tech firms creates self-reinforcing cycles—models trained on internet scrapes absorb cultural biases, then flood the internet with generated content that pollutes future training runs (a dynamic also explored in model collapse research). Regulatory attention on compute thresholds or safety testing, while useful, misses the deeper economic reality: corporate incentives reward rapid deployment over deliberate, representative data practices.

This contrarian lens suggests AI limitations discourse has been captured by spectacle. The real crisis isn't whether an LLM can reliably cite sources—it's that we've built an information economy where these tools are deployed into high-stakes domains without corresponding accountability structures, diversity in development teams, or investment in non-statistical complementary approaches. Large language models are not the problem. They are the clearest evidence of the problems we've refused to solve.