Butterfly Data has called on public sector organisations to prioritize data provenance in artificial intelligence (AI) development, emphasizing that this issue extends beyond traditional data quality concerns. Maja Strawinska, a data scientist at Butterfly Data, noted that many teams mistakenly believe that cleaner data alone can address issues of fairness, accuracy, and governance. She highlighted that even well-structured datasets might be unsuitable for AI if organisations cannot clarify their origins, purposes, and legal reuse conditions.
Strawinska distinguished between “clean data” and “trustworthy data,” particularly within the public sector, where automated systems can significantly impact service access and care delivery. In such contexts, the dataset’s history is equally important to its format or completeness. “The important question we need to ask is simple: where did this data actually come from?” Strawinska said. This inquiry involves understanding who collected the data, under what conditions, for what purpose, and whether those circumstances pose risks for current applications.
To underscore her point, Strawinska compared data provenance to the farm-to-table approach in the food industry, where trust is not solely based on the final product, but also on a transparent supply chain. This is particularly vital in the public sector, where many datasets have evolved through legacy systems over time. Although technical improvements like data migration and standardization can enhance quality, they do not resolve questions about the original data collection methods or the terms of its current usage.
The issue of data provenance also encompasses compliance and oversight. Strawinska argued that it should not merely be viewed as a technical concern, but rather as an integral aspect of responsible AI, directly linked to data protection obligations amid increasing regulatory scrutiny. Her remarks reflect a broader trend in AI governance, particularly within government and public services, where there is growing pressure to explain not only what an AI model does, but also the foundations on which it is constructed. In this regard, maintaining a data audit trail is becoming increasingly essential for justifying the deployment of AI systems.
While acknowledging the value of standard data quality efforts—such as removing duplicates and standardizing formats—Strawinska cautioned that such initiatives cannot address every challenge. For instance, data collected without valid consent or for a different purpose cannot be deemed appropriate for a new application simply because it has undergone cleaning and validation. She illustrated this with the analogy of food grown in contaminated soil, explaining that even if vegetables are washed and prepared, they can still be unsafe due to their origins. The same reasoning applies to datasets whose origins may introduce legal, ethical, or representational issues.
This challenge is especially pronounced for public bodies managing information gathered over decades. Much of this data was collected prior to the establishment of current data protection standards, complicating efforts to apply modern AI techniques to older records. Strawinska also emphasized the significance of understanding when bias enters an AI system. Discussions around AI bias typically focus on model outputs and fairness testing; however, biases may originate much earlier during the data collection and assembly phases.
If a dataset over-represents certain demographics, regions, or timeframes, the resulting AI model may reflect these discrepancies. For instance, systems trained predominantly on urban data may perform poorly in rural settings, while models built on data collected during periods of unusual demand may falter when conditions normalize. For public services, Strawinska insisted that these limitations should be identified prior to deployment rather than after, with data provenance helping organisations to assess a dataset’s true representativeness and its potential gaps.
As AI systems grow larger and draw from diverse data sources, the task of maintaining a clear account of data handling becomes increasingly complex. Strawinska argued that organisations incorporating provenance tracking from the outset will be better equipped to navigate audits, oversight committees, and public scrutiny. In the public sector, the ability to elucidate these decisions is closely linked to public trust in AI applications. “Data provenance—the ability to trace where data came from, who handled it, and how it has changed—is often seen as a niche technical topic. It isn’t. It is at the heart of what responsible AI requires,” she stated.
See also
OpenAI’s Rogue AI Safeguards: Decoding the 2025 Safety Revolution
US AI Developments in 2025 Set Stage for 2026 Compliance Challenges and Strategies
Trump Drafts Executive Order to Block State AI Regulations, Centralizing Authority Under Federal Control
California Court Rules AI Misuse Heightens Lawyer’s Responsibilities in Noland Case
Policymakers Urged to Establish Comprehensive Regulations for AI in Mental Health
