This June, the second Vine to Mind Symposium convened in Lausanne, Switzerland, co-hosted by the Harvard Data Science Initiative (HDSI) and EHL Hospitality Business School. Launched in 2024, this series uses the wine industry as a lens to examine how data and AI operate in environments shaped by cultural, environmental, and regulatory complexity.
The 2025 symposium brought together economists, climate scientists, regulators, machine learning researchers, and winemakers to stress-test how data systems perform under real-world constraints. “Wine gives us a perfect excuse to ask hard questions about data science,” explains Xiao-Li Meng, Whipple V. N. Jones Professor of Statistics at Harvard and Editor-in-Chief of the Harvard Data Science Review (HDSR). “What do we measure? What do we ignore? And who decides?”
Over two days of keynotes, tastings, and roundtables, four key insights emerged:
1. Modeling assumptions are set by framing choices.
How data is structured—what’s measured, labeled, or left out—shapes algorithmic behavior. In domains rooted in human perception, framing choices determine not just model output but credibility.
Case study: Eric LeVine, Founder and CEO of CellarTracker, manages a massive database of user-submitted wine reviews. These notes, written in informal, metaphor-rich language, often describe the same wine in contradictory ways. One drinker might say ‘bright and citrusy,’ another says ‘harsh and sour.’ Traditional NLP systems struggle to process such inconsistency without collapsing it into a common denominator. LeVine’s challenge was to design a system that could structure the data without erasing its subjectivity. The system treats variation in language and perspective as meaningful information rather than noise. It organizes subjective input without reducing it to a single norm, preserving differences as part of the structure rather than smoothing them out.
Forward view: To support systems that engage with perception, preference, or ambiguity, data scientists must rethink evaluation itself. Instead of enforcing consistency, models should expose patterns in disagreement and offer structured representations of variability. Tools such as ensemble labeling, uncertainty-aware learning, and interface-level interpretability can help preserve disagreement as structured data—allowing models to represent variation explicitly to harvest useful information rather than force artificial agreement that can do harm.
2. Contextual relevance should be foundational.
Case study: Marc Brevot, Director of Robert-Jean De Vogüé Research Center for Moët Hennessy, described how traditional yield models—built on stable historical averages for temperature and rainfall—began to fail as climate patterns grew erratic. In response, his team rebuilt the model using variables once treated as peripheral—soil microbiomes, canopy structure, chemical composition—and restructured it to track how growing conditions shift from year to year. The result is a system capable of responding to highly localized conditions. The model incorporates recent data at fine-grained spatial and temporal resolution, capturing how soil chemistry, plant health, and microclimate interact in each season. This shift from generalized prediction to contextual responsiveness marks a fundamental change in how agricultural models are built and maintained.
Forward view: Context-aware modeling requires architectures that can be retrained incrementally, updated with new sensor or environmental data, and adapted to small-scale variation without global model resets. This includes localized calibration, feature reweighting based on seasonal inputs, and modular workflows that isolate change-sensitive components from core logic. Building in this flexibility allows models to remain valid as conditions evolve, without compromising traceability or interpretability.
3. A model’s value depends on its fit with human practice.
If human users can’t interpret, trust, or act on a model’s output, the system fails. Alignment with real-world decision-making—timing, constraints, expectations—is as critical as technical accuracy.
Case study: Kia Behnia, Co-Founder and CEO of Scout, builds AI tools that use tractor-mounted cameras to monitor vineyards. The system identifies disease risks and estimates yields using high-resolution imaging and predictive models. The system achieved high accuracy in detecting disease and predicting yield, but growers only adopted it when its outputs matched their observations and expectations. If a recommendation appeared inconsistent with what they saw in the field—such as marking a healthy-looking vine as diseased—it was disregarded. This revealed a gap between statistical performance and practical credibility. Behnia’s team adjusted the model to better align with grower priorities, such as timing of decisions, typical indicators of stress, and localized conditions—ensuring it supported rather than disrupted on-the-ground decision-making.
Forward view: Supporting expert decision-making requires systems that reflect how judgments are made in practice. This means surfacing the logic behind predictions, linking outputs to recognizable field conditions, and offering pathways to contest or refine results. Model design must anticipate how users think, what they notice, and when they act—embedding domain-specific cues, localized diagnostics, and feedback cycles that make the system intelligible and responsive in use.
4. Private metrics are reshaping public expectations.
Self-tracking tools—from wearables to wellness apps—generate behavioral data originally intended for personal insight. These metrics, rather than cultural norms or institutional consensus, are beginning to reshape collective standards for what is healthy and responsible.
Case study: Felicity Carter, Editorial Director of Areni Global, examined how personal data is reshaping public attitudes toward alcohol. Metrics like sleep quality, recovery time, and heart rate variability increasingly frame moderate drinking not as neutral, but as a deviation from optimized performance. These signals are amplified by wellness influencers and social platforms, reinforcing abstention as a moral good and indulgence as failure. This shift shows how platforms and wellness tools influence norm enforcement through default thresholds, performance scoring, and visibility algorithms, often without the user’s knowledge.
Forward view: When metrics designed for personal insight are repurposed to signal public norms, systems must make that transition visible. Designers should document how behaviors are categorized, disclose the thresholds that drive classification, and provide mechanisms for users to understand or challenge those judgments. Norm formation is a structural outcome of system design, and responsible design requires making those structures open to inspection.
These four insights trace a shift in what data science demands. First: how data is framed. Second: how it adapts to local conditions. Third: how it fits into expert judgment. Fourth: how it influences social expectations. Together, they outline a field that must be context-aware by design, situated, and accountable.
At the HDSI and the HDSR, this is the work we support. Vine to Mind uses viticulture as a complex, real-world system to expose how data science can be built to reflect environmental, institutional, and human constraints. What we observe in wine – and indeed, in agriculture and food systems – applies far beyond it. These same failures of interpretation, fit, and feedback occur in public health, climate science, medicine, finance and countless other fields.
“Vine to Mind confronts the same tensions we see in climate and health—where data must be adapted to local realities, and trusted by decision-makers,” shares Francesca Dominici, Professor of Biostatistics, Population, and Data Science at Harvard and HDSI Faculty Director. “These challenges are fundamental. If data science avoids them, it will repeatedly fail to create the impact we desperately need. But, if we take them seriously, we will build systems that work, for people and for the planet.”