You Can’t Have Responsible AI Without Responsible Data

Responsible AI starts at measurement—because that’s where bias begins.

Everyone talks about responsible AI. But too often, the focus starts at the model—long after the most important decisions have already been made. In reality, responsibility begins at measurement. It depends on how data is defined, collected, and documented—and whether that context travels with it. From neuroscience to clinical care, this post explores why data subjectivity is not a flaw to fix, but a fact to design for—and how FAIR² makes that possible.

Science aspires to objectivity. Our methods are designed to reveal truths that hold regardless of who is observing. Artificial intelligence has become an extension of this aspiration: a way to find patterns too complex or subtle for human reasoning alone. But neither science nor AI stands above data. They both rely on it. And data, contrary to what we might hope, is never truly objective.

All data reflects context. It carries with it the imprint of who collected it, what they chose to measure, how they defined their variables, and what they left out. This isn’t a flaw in data—it’s an unavoidable fact. The moment we begin to observe, select, and record, we introduce structure. And structure implies perspective. Data is not neutral. It is designed, interpreted, and situated.

This reality is easy to overlook when the data seems mechanical—such as a timestamp or a financial transaction. These feel stable, tightly defined, and machine-generated. And often, they are. Financial data, for example, tends to be highly structured and consistent across systems. A transaction record has an amount, a source, a destination, and a time—all agreed upon and encoded within digital protocols.

But most of the data we now use in science, healthcare, and artificial intelligence isn’t like this. Clinical data, behavioral data, social data, and even experimental scientific data are full of ambiguity, variation, and assumption. Diagnoses involve interpretation. Labels shift over time. Instruments differ between labs. Variables like race, housing status, cognitive function, or disease severity may be defined differently not just between countries, but between institutions—or even between individual clinicians.

This was confronted directly in my collaborative work with Dr. Laura Sikstrom and colleagues at CAMH, where we co-developed what became known as the Fairness Dashboard. Our aim was to make sociodemographic and clinical data more transparent and interpretable for data scientists building AI models in mental health care. We were working with variables that are often central to questions of equity and bias—things like race, gender identity, income level, or housing status. What we found is that these variables were not only inconsistently defined, they were inconsistently recorded. Entire categories were missing. Variables changed names or meaning depending on the system. And perhaps most concerning of all: the presence or absence of certain data often mapped onto deeper structural inequities.

This experience taught us that fairness in AI can’t be addressed solely through model-level auditing or bias mitigation algorithms. If the data itself is ambiguous, unrepresentative, or undocumented, the best-intentioned models will still make inappropriate inferences. The real work must begin at the level of data design. That means being clear not just about what’s in the data, but how it came to be.

Importantly, this isn’t a challenge unique to human data. The same lessons apply in basic science. Consider the example of neuroscience. In 2015, Tripathy et al. published an important study describing a project called NeuroElectro. Their goal was to aggregate published measurements of intrinsic electrical properties of neurons—things like membrane resistance, time constants, and firing thresholds—into a single searchable database. The results were startling. When the same cell type was measured by different labs, the reported values often differed substantially. But on closer inspection, these weren’t random errors. The variation could largely be explained by differences in experimental conditions: the temperature of the recording, the species used, the animal’s age, or the type of electrode.

By extracting and modeling this metadata—essentially, by accounting for context—the NeuroElectro team was able to normalize these datasets and make meaningful comparisons across studies. Their work showed that what looks like noise in scientific data is often structure that has been stripped of its context. And when that context is restored, the data becomes usable again—not just in theory, but in practice.

We saw this need for context even more deeply in our work on the Blue Brain Project (and now continued in the Open Brain Institute). Our challenge was to build biologically grounded simulations of brain microcircuits by integrating a wide range of data types: patch-clamp recordings, single-cell transcriptomics, neuronal morphologies, synaptic connectivity maps, cell-type densities, and more. Each of these data sources was valuable. But each also came with its own history—how it was collected, where it came from, which species it described, what tools and protocols were used.

To responsibly reuse this data, we had to make those histories explicit. So we built a provenance-aware knowledge graph—an infrastructure that preserved not just the facts, but the facts about the facts. Every element in the graph linked to structured metadata about how it was produced, under what assumptions, and using what instrumentation. This wasn't just helpful for downstream users. It was essential for ourselves: without this structure, we couldn’t meaningfully integrate the data, validate our simulations, or understand the sources of uncertainty in our results.

In both neuroscience and clinical care, the lesson was the same: data without context cannot be reused responsibly. That doesn’t mean we can’t reuse data. It means we must preserve the context that gives it meaning.

The FAIR principles—Findable, Accessible, Interoperable, and Reusable—were a transformative step in this direction. They emphasized the need to make data shareable and machine-readable. But they didn’t spell out how to embed meaning. They didn’t specify how to structure metadata for machine learning systems, or how to preserve subjective nuances across use cases.

That’s where FAIR² comes in. FAIR² builds on FAIR but adds what AI systems need in order to be responsible: structure for provenance, responsibility, and context. FAIR² data packages are AI-Ready: designed for modeling, with clear schemas and defined units. They are Responsible: aligned with ethical expectations for transparency, documentation, and traceability. And most importantly, they are Context-Rich: every feature in the dataset links to its method of collection, its definition, its instrumentation, and any known limitations or transformations.

This doesn’t just help humans understand the data. It allows AI systems to model the data responsibly—learning not just from values, but from their meaning.

The principle applies whether we’re working with human data or not. Responsible AI isn’t just about protecting people from harm. It’s about enabling trustworthy inference, wherever data is used. Whether you're modeling a neuron, predicting a diagnosis, or optimizing a treatment plan, the process only works if your data is interpreted in light of how it was created.

Objectivity isn’t the absence of subjectivity—it’s the ability to recognize it, document it, and design systems that can reason through it. That’s what we mean when we say responsible AI starts with responsible data. FAIR² helps make that possible—at scale.

Responsible AI is a broad and evolving discipline. It’s about more than just good data—it also includes transparent models, accountable systems, fair outcomes, and informed human oversight. This post has focused on the foundation: context. But FAIR² also supports these broader goals. By embedding provenance and traceability, it strengthens accountability. By requiring context-rich documentation, it enables interpretability and auditability. And by embedding structure at the level of data, FAIR² creates the conditions for AI systems that uphold scientific integrity and social responsibility.

FAIR² isn’t the whole answer to Responsible AI—but without it, we can’t even ask the right questions.

If you care about scientific rigor, fairness in AI, or reproducible insight, then you need to care about how data is measured—and whether that context is carried forward.