A look at researchers' use of deep neural networks to examine historical documents and the related risks, like ML slipping bias or falsifications into records
and the ways in which artificial intelligence could upend our relationship to history itself https://twitter.com/... Yannis Assael / @iassael : Thank you @techreview for the insightful article! The fusion of AI and Humanities in interdisciplinary research has an immense potential to illuminate our historical understanding. @TSommerschield @BrendanShilling @NandoDF https://twitter.com/...
Context & Ripple Effects
Deep learning's move from ImageNet-era benchmarks into the archive has been building for years: the field's own risk literature — including the draft paper co-authored by Timnit Gebru on large language model risks — flagged training-data bias as a structural problem, and separate reporting found AI tools lack the training data to grasp the cultural meaning of material from underrepresented cultures. This MIT Technology Review piece applies that critique to a new domain: historians and researchers like Yannis Assael using deep neural networks to read historical documents.
First-order effects
- Historians and archivists adopting neural-network tools for reading documents now face a direct quality problem: model errors and biases can enter the historical record as if they were scholarly findings.
Second-order effects
- The gap between model capability and cultural grounding — documented in coverage of AI's weak training data for underrepresented cultures — means whose history gets read and reconstructed depends on which archives the training data overrepresents.
Third-order effects
- If AI-assisted readings become routine in scholarship, peer review inherits the burden of catching machine-introduced falsifications, extending the oversight gap researchers have long warned about in AI ethics into the humanities itself — a concern later case studies showing frontier models handling PhD-level historical analysis only sharpen.
The trend: AI is moving from analyzing the present to reconstructing the past, forcing historical scholarship to build verification practices for machine readings of the record.