Data preparation remains one of the most time-consuming steps in analytics and machine learning. Cleaning messy datasets, inferring schemas, and mapping fields across sources are essential but tedious. As artificial intelligence moves from hype to pragmatism, a useful framework is to view AI as 'normal technology'—a tool that humans can and should control. As the Knight First Amendment Institute at Columbia University explains:
This perspective describes AI as a tool that humans can and should control, not as a superintelligent entity.
Just as future hardware-design tools will let engineers prompt AI to produce review checklists, similar approaches could apply to data wrangling tasks like schema inference and cleaning. By analogy, an AI system might flag a column containing 'N/A' and 'null' as potentially representing missing data and suggest an imputation method. Such automation could accelerate the initial cleaning phase, but the model's suggestions would still need domain review.
Type inference and schema mapping offer another promising area. When combining data from multiple sources, an AI might propose that a 'DOB' field matches a 'dateofbirth' column and recommend casting both as dates. However, mismatches can slip through, so automated checks and human oversight remain critical to ensure accurate mappings.
A key risk is that AI may introduce or reinforce biases during cleaning. For example, if income data is missing disproportionately for part-time workers, mean imputation could understate their true earnings. Treating AI as a tool under human control—rather than as an autonomous agent—allows practitioners to audit transformations and ensure equitable treatment across subgroups.
Validation workflows are essential for reliable automation. Statistical checks like distribution comparisons and review checklists aligned with project requirements provide guardrails. In hardware design, engineers will be able to prompt AI to produce review checklists that leverage approved components and libraries. By analogy, a similar approach in data wrangling would have data scientists verify each automated decision against business rules. This human-in-the-loop model keeps AI as a practical co-pilot, not an unchecked operator.
