The foundation of any machine learning system is its training data. Yet despite the centrality of data to model performance, the ethical dimensions of dataset construction often receive insufficient attention during the development process. As AI systems become more powerful and widely deployed, the decisions we make about training data—what we include, where it comes from, and how we document it—have profound implications for fairness, safety, and societal impact.

At Drane Labs, we believe that rigorous data sourcing isn't just a compliance checkbox. It's a fundamental component of building AI systems that deserve trust. This post outlines our framework for thinking about training data ethics, focusing on three pillars: provenance, consent, and documentation.

The Provenance Problem

Dataset provenance refers to the complete lineage of training data—where it originated, who created it, how it was collected, and through what intermediaries it has passed. In an era where massive web scrapes and pre-aggregated datasets dominate the landscape, provenance has become increasingly opaque.

Consider a typical computer vision dataset. Images might be scraped from public websites, aggregated through third-party services, filtered by automated systems, and labeled by crowd workers—all before a single model is trained. At each step, important context is lost: the original photographer's intent, the circumstances of capture, the potential biases in what was deemed worth photographing in the first place.

We've seen this problem manifest in real-world harms. Facial recognition datasets containing mugshots without consent. Medical imaging datasets that overrepresent certain demographics. Datasets scraped from online forums that encode toxic social norms. In each case, insufficient attention to provenance allowed problematic data to enter the training pipeline.

The solution isn't to avoid all complex data sources—that would be impractical for most modern applications. Instead, we need robust provenance tracking systems. At Drane Labs, every dataset we use includes a provenance document that answers:

  • Origin: Where did this data first appear? Who created it and why?
  • Collection method: How was it gathered? What were the selection criteria?
  • Chain of custody: What organizations and systems has it passed through?
  • Known limitations: What populations, scenarios, or contexts are underrepresented or absent?

This documentation travels with the data throughout the ML pipeline. When we evaluate a model's failure modes, we can trace them back to dataset characteristics. When we consider deployment contexts, we can assess whether our training data's provenance aligns with the target domain.

Informed Consent and Data Rights

The question of consent in AI training data is more complex than it initially appears. At first glance, the principle seems straightforward: data about people should only be used with their permission. But the reality involves numerous edge cases and evolving norms.

For data explicitly created for AI training—think labeled datasets constructed by research institutions—consent frameworks are relatively clear. Participants can be informed about how their data will be used, what models will be trained, and what deployment contexts are anticipated. They can make an informed decision about participation.

The challenge arises with data originally created for other purposes. A photograph posted to a social media platform in 2010. A scientific paper published before large language models existed. A surveillance camera feed from a public space. In these cases, the data creators often had no way to anticipate AI training as a use case.

Some argue that public availability equals consent—if something is posted online, it's fair game. We reject this framing. Public visibility doesn't imply consent to any imaginable future use. The reasonable expectations of data creators matter.

Our approach at Drane Labs involves several principles:

Respect original context: If data was created for a specific purpose, we consider whether AI training aligns with that purpose. Scientific publications intended to advance research are different from personal photos intended for friends and family.

Enable opt-out mechanisms: Where possible, we implement systems that allow data creators to remove their contributions from training datasets. This is technically challenging for already-trained models, but it's feasible for ongoing data collection and future model versions.

Compensate when appropriate: For certain types of data, particularly creative works and specialized expertise, we believe compensation is ethically necessary. This is still a developing area, but we're exploring models that fairly credit data contributors.

Prioritize purpose-built datasets: When feasible, we invest in datasets explicitly created for AI training, where consent is unambiguous. While more expensive than web scraping, this approach provides a more stable ethical foundation.

Documentation as Accountability

The third pillar of our framework is rigorous documentation. Training data documentation serves multiple functions: it enables reproducibility, facilitates bias auditing, supports model debugging, and creates accountability for dataset decisions.

We've adopted and extended the "Datasheets for Datasets" framework proposed by Gebru et al., which calls for standardized documentation covering motivation, composition, collection process, preprocessing, uses, distribution, and maintenance. Our datasheets answer questions like:

  • Why was this dataset created? What gap does it fill? What applications was it designed to enable?
  • What's in it? What do the instances represent? How many instances are there? What's the distribution across relevant categories?
  • Who's missing? What populations, scenarios, or examples are underrepresented or absent entirely?
  • How was it collected? What mechanisms were used? Who performed the collection? What ethical review was conducted?
  • What preprocessing was applied? What transformations were made? Was anything filtered out?
  • What are the known limitations? Where should this dataset not be used? What failure modes are anticipated?

Importantly, these datasheets are living documents. As we discover issues during model development and deployment, we update the documentation. When we learn about unintended uses or harms, we add warnings. This creates an institutional memory that prevents recurring mistakes.

Documentation also enables external accountability. We make selected dataset documentation publicly available, allowing researchers and civil society to audit our decisions. While competitive concerns prevent full disclosure of all datasets, transparency where possible builds trust and enables community feedback.

The Path Forward

Training data ethics is not a solved problem. As AI capabilities advance, new challenges emerge. Synthetic data raises questions about authenticity and representation. Multimodal models create novel consent challenges when combining text, images, and audio. Federated learning enables training on distributed data while preserving privacy, but introduces new provenance challenges.

What remains constant is the need for principled frameworks and institutional commitment. Training data ethics can't be an afterthought or a compliance exercise. It must be integrated into the core of how we build AI systems.

At Drane Labs, we're still learning. We make mistakes. We encounter edge cases our frameworks don't cover. But we're committed to approaching these challenges with humility, transparency, and a genuine belief that how we source and handle training data matters—not just for regulatory compliance or risk management, but because it's the right foundation for AI systems that serve humanity.

The decisions we make today about training data will shape AI's trajectory for decades. We owe it to the people whose data powers these systems, and to society more broadly, to make those decisions with care.

Priya Sandoval is Principal Research Scientist at Drane Labs, where she leads the AI Ethics and Safety team. She holds a PhD in Computational Ethics from MIT and has published extensively on algorithmic fairness and responsible AI development.