Sustainable data products and practices
Sustainable AI is not only about watts per FLOP. Data products—features, tables, labels and pipelines—shape storage, recomputation and risk. Lean, well-governed data reduces cost, emissions and harm.
Data minimisation
Collect and retain only what you need. Over-collection increases storage energy, widens breach impact and complicates consent. Minimisation aligns with GDPR-style principles and often improves model focus when paired with strong definitions.
Practices include selective retention, aggregation for analytics where raw detail is unnecessary, and anonymisation or pseudonymisation when identifiers are not required.
Data quality and integrity
Sustainable models need trustworthy inputs: validation rules, automated checks, versioned datasets and clear ownership. Poor quality forces retraining, hotfixes and rework—all of which burn energy and trust.
Responsible data governance
Governance connects policy to implementation:
- Access control limits sensitive data to authorised roles.
- Audit trails show who changed what and why—critical for regulated domains.
- Bias and fairness assessments catch discriminatory patterns before they scale; see Algorithmic Justice League and IBM discussion of real-world bias.
Energy-aware algorithms and hardware
Pair lean data with efficient computation: smaller models where adequate, quantisation, specialised chips and vendor tools that deduplicate storage (for example NetApp-style dedup/compression patterns in enterprise storage).
Key point
Sustainability wins stack: less data movement + higher data quality + more efficient compute beats optimising only one layer.
Action item: Quiz
- A. Storing every click forever “just in case”
- B. Defining the minimum fields required for the decision, with documented retention and purpose limitation
- C. Duplicating raw exports to every analyst laptop
- D. Disabling access controls to speed onboarding
Feedback: Minimisation is intentional scope and retention, not maximal hoarding.
- A. They replace the need for testing
- B. They guarantee models are unbiased
- C. They reduce repeated forensic work and enable targeted fixes when issues arise
- D. They remove legal obligations
Feedback: Traceability shortens incident response and prevents endless re-discovery of the same gaps.