Dark data in agrifood — collected but not surfaced
Global
Content
Dark data in agrifood is data that is collected but not surfaced for broader use. This is the inequality problem in agrifood AI: data exists, is captured, is sometimes aggregated, but the benefits flow to vendors and aggregators, not to the farmers whose operations generated it.
Categories of dark data in agrifood
- IoT sensor data — soil moisture probes, barn environment sensors, climate sensors in greenhouses. Collected by individual farms or vendors but aggregated only at the vendor level (and not surfaced publicly).
- Machine telemetry — combine yield sensors, planter population sensors, sprayer volume sensors. Captured by equipment makers (Deere, AGCO, CNH) but aggregated and operationalised within vendor platforms.
- Government-classified satellite data — high-resolution satellite imagery held by governments for national-security purposes, not released for agricultural use. Some countries (India, China, Russia) have meaningful classified satellite capacity that could theoretically benefit agricultural monitoring but is not released.
- Aggregator-pooled data — vendors aggregate farm data into pooled datasets (anonymised, supposedly) but do not surface the pooled data for farmer benefit. The aggregate value flows to the vendor and to the model’s training, not back to the contributing farmers.
The structural problem with dark data
The farmers whose data trains the model receive the model’s recommendations back, but the aggregated insight (regional patterns, market signals, supply-chain trends) is held by the vendor. This is the value capture concern that the FCC ecosystem-not-technology framework names as “data governance” as one of the four systemic constraints.
The dark-data problem is structurally different from the proprietary-data problem:
- Proprietary data is known to exist but governed by restrictive terms.
- Dark data is collected without the farmer’s full awareness of how it is used in aggregated form.
The structural asymmetry
The vendor / aggregator captures:
- The aggregated value (regional pattern insights, ML training data improvements)
- The supply-chain optimisation opportunities (predictive logistics)
- The market positioning (knowing farmer-level input demand across a region)
The contributing farmer receives:
- The model’s recommendation back
- Possibly a better-than-baseline prediction
- Not the aggregated insight or the supply-chain value
This asymmetry is the structural inequality that the FCC’s “data governance” framework constraint names, and that Civil Eats captures in the practitioner reality quote.
Related concepts worth distinguishing
- Data minimization (a privacy principle) — collect only what is needed for the stated purpose. Distinct from the dark-data problem (where data is collected regardless of minimization).
- Data sovereignty — control over data by communities / states / Indigenous peoples. The CARE Principles articulate this. Dark data is without sovereignty.
- Data altruism / data cooperatives — farmers pooling data for mutual benefit. Worth naming as a response to the dark-data problem. Worth tracking for the field guide.
What the dark-data problem means for talks
- Most agrifood AI training data is captured without explicit farmer consent to the aggregated use.
- Vendor-aggregated value (regional patterns, supply-chain optimisation) does not flow back to contributing farmers.
- Government-classified satellite data (where it exists) is a structural missed opportunity.
- The data cooperative / data altruism response is a substantive alternative worth naming.
What this unit is doing in the taxonomy
Anchors the data-substrate × dark cell as a framework claim-type. First unit in the field guide that explicitly captures the dark-data problem.
Distinct from:
- Open data ecosystem unit (
open-data-ecosystem.md) — open / state / civil society cell. - Proprietary farm data unit (
proprietary-farm-data.md) — proprietary / industry cell. - Farm-data ownership critical voice unit (
farm-data-ownership-critical.md) — critical lens.
Why it matters for talks
- Dark data is the inequality problem in agrifood AI. Worth naming in any talk that engages seriously with the data substrate.
- The structural asymmetry between vendor-aggregated value and farmer-received recommendation is concrete and worth naming.
- The data cooperative / data altruism response is a substantive alternative worth surfacing.
- The government-classified satellite angle is a structural missed opportunity worth knowing.
Critical context
- The dark-data problem is structurally distinct from proprietary data (different problem) but adjacent (often same vendors).
- Data altruism / data cooperatives are an active alternative but under-deployed in our scan.
- The structural asymmetry between vendor-aggregated value and farmer-received recommendation is the substantive concern.
- The FCC’s “data governance” framework constraint names this problem in the Canadian context.