Bad data is quietly breaking AI models — and the fix starts before training

Everyone talks about model architecture, parameter counts, and compute budgets. Fewer people want to discuss the unglamorous truth: the data feeding those models determines whether they actually work. The latest Ethics and Society Newsletter from the Montreal AI Ethics Institute puts data quality front and center, arguing that what goes into a model matters as much as how the model is built.

The newsletter frames data quality as fitness for purpose, not some abstract ideal. A heart disease prediction model needs detailed patient histories and medication dosages — but it probably shouldn’t have access to patients’ phone numbers. That’s relevance in action. The piece also highlights comprehensiveness, timeliness, and bias mitigation as core pillars. Outdated training data isn’t just suboptimal; it can render a system actively dangerous in fast-moving domains.

There’s a practical angle here that often gets lost in ethics discussions. High-quality data isn’t just about fairness or avoiding harm — it makes models more efficient. Curated datasets reduce noise, prevent overfitting, and lead to more compact models that require fewer computational resources. That’s a sustainability win and a cost win simultaneously. The newsletter also emphasizes governance and scientific reproducibility, pointing out that transparency about data provenance is what separates trustworthy systems from black boxes.

What’s notable is the timing. As companies race to train ever-larger models on ever-larger datasets scraped from the web, the conversation is finally shifting upstream. You can’t fix data quality problems after the fact. The architecture won’t save you. The compute won’t save you. The data either supports the use case or undermines it — and most organizations won’t know which until it’s too late.