Privacy by design means incorporating data protection into an AI product from its earliest planning stages, rather than adding controls after development. Under UK GDPR Article 25, organisations are expected to apply appropriate technical and organisational measures that protect data protection principles and individual rights. This article is part of the reference collection published by Harper James.
A privacy by design process begins by defining the product’s purpose, users, data sources and expected outputs. The organisation should identify whether the system processes personal data, including information used for training, testing, prompts, logs, monitoring or human review. Data minimisation should guide the design: collect only data needed for a defined purpose, retain it for no longer than necessary, and consider anonymisation or pseudonymisation where appropriate.
The product team should document the legal basis for processing, the roles of the organisations involved, and any transfers to suppliers or other countries. A data protection impact assessment is required where processing is likely to result in a high risk to individuals, such as large-scale monitoring, profiling or the use of sensitive personal data. The assessment should cover foreseeable misuse, inaccurate outputs, security risks and the effect on people affected by automated decisions.
Controls can include access restrictions, encryption, secure deletion, tenant separation, audit logs, input filtering and testing for unintended disclosure. Training data should be sourced and used under documented rules, with processes for handling unlawful or inaccurate material. Where an AI system produces decisions or recommendations about individuals, the organisation should assess the need for human oversight, explain relevant processing in clear language and provide routes for exercising rights such as access, rectification and objection.
Privacy by design continues after launch. Organisations should monitor incidents, complaints, model changes, supplier changes and shifts in the data being processed. Significant changes should trigger a review of the impact assessment and privacy notices. Records should show who approved the design, which risks were identified, what controls were selected and when the controls were tested. This evidence supports accountability and helps the organisation revise the product when its uses or risks change.