Posts
Get my latest updates and learn more about privacy protection and emerging tools and technologies
Developers on the App Store Small Business Program* can now get access to a server-based Large Language Model (LLM) via Private Cloud Compute (PCC) at Apple with no cloud API cost (subject to usage limits).
Based on the capabilities required for their specific use cases, developers on Apple platforms can choose whether to use Apple’s local (on-device) model, or Apple’s server model (via PCC).
Apple highlight that PCC: i) does not store data, ii) only uses data to process requests, iii) has been independently verified by researchers.
Both the local (on-device) model and the server model (via PCC) provide developers with some assurance over the flows of (potentially personal) data that arise from interaction with an LLM.
Worth noting that data flows arising from interaction with an LLM (eg. prompts and responses) are likely to represent only a subset of the flows into, and out of an application.
Developers need to understand how these data flows integrate with other flows in an application to ensure that personal data is processed in line with expectations.
*Additionally, app must have less than 2 million total first-time app store downloads
View ‘What’s New in Apple Intelligence’ at: https://developer.apple.com/apple-intelligence/whats-new/
View the ‘Build with the new Apple Foundation Model on Private Cloud Compute’ video at: https://developer.apple.com/videos/play/wwdc2026/319/
View the ‘WWDC26 Platforms State of the Union’ at: https://developer.apple.com/videos/play/wwdc2026/102/
Is the development, deployment and usage of AI systems in accordance with data protection requirements fundamentally any different to other data processing systems?
Whether or not a system is considered to be an ‘AI system’, organisations need to understand the (personal) data flows and data processing tasks that are associated with each stage of the system lifecycle.
Complexity of data flows and data processing tasks that comprise modern AI systems (and data processing systems in general) may increase the expertise, time and resources required to gain this understanding. However, such understanding remains essential to the fulfilment of data protection requirements at each stage in the system lifecycle.
Any systems (AI or otherwise) for which (personal) data flows and data processing tasks are not fully understood are unlikely to fulfil data protection requirements, and are likely to place individuals and organisations at risk of harm.
View the ICO’s response to their consultation series on Generative AI at: https://ico.org.uk/about-the-ico/what-we-do/our-work-on-artificial-intelligence/response-to-the-consultation-series-on-generative-ai/executive-summary/
View the ICO’s Tech Futures report on Agentic AI at: https://ico.org.uk/about-the-ico/research-reports-impact-and-evaluation/research-and-reports/technology-and-innovation/tech-horizons-and-ico-tech-futures/ico-tech-futures-agentic-ai/
Royal Observatory makes some excellent points about the potential effects of AI dependence on individuals and society. Dependence of individuals on AI tools, such as chatbots, raises some important questions regarding trustworthiness and trust. For example:
- To what extent can information provided by an AI tool be considered trustworthy?
- How do users determine the trustworthiness of the information provided by AI tools?
- Are users making ‘good’ decisions about placing (or refusing) trust in the information provided by AI tools?
A key challenge for users in determining trustworthiness and making ‘good’ decisions is that AI tools may communicate trustworthy and untrustworthy information in the same (or similar) way(s).
Users need to be able to distinguish trustworthy information from untrustworthy information, but may have limited information, capacity and capability to do so.
View the BBC News article at: https://www.bbc.co.uk/news/articles/c2023l60370o
Some reasons why employers and workplaces may shape the early governance of wearable technologies in society:
- Employers have duties and obligations in areas that are directly impacted by use of wearable technologies in the workplace. For example, safety, equality and data protection.
- Employers can identify and articulate specific harms that could arise from use of wearable technologies in the workplace. For example, harms arising from intimidation of staff, ‘chilling effects’ on organisational discourse, and breaches of confidential data.
- Organisational policies can adapt quicker than legislation to the challenges posed by wearable technologies; defining expectations relating to their use (or not) in the workplace, establishing norms that are adopted more widely in society, and providing experiences that inform development of legislation and guidance.
View the BBC News article at: https://www.bbc.co.uk/news/articles/cj37z8357e5o
Automation of software development using AI agents raises important governance questions, including:
- How do AI agents operationalise human values such as privacy and fairness in their processes and outputs? Human values are typically expressed using high-level principles that must be ’translated’ into practice in a given context. Given the significant challenge for humans in undertaking such ’translation’ in the design and evaluation of systems, it is unclear how (or even whether) AI agents could undertake and justify such ’translation’ with sufficient rigour for systems to be trustworthy with respect to these values
- Who is responsible for the process and outputs of these AI agents? AI agents can not only generate code, but also generate the code that tests the code. Humans move further away from not only the code itself, but the assurance of the code with respect to functional and non-functional properties. Additionally, as the scope of AI agent work increases, the volume and complexity of code (and tests) generated may exceed that for which one or more humans have the capability or capacity to comprehend and reasonably oversee. Individuals within organisations may be asked to assume legal and moral responsiblity for processes and outputs that they cannot or do not comprehend, and for which their participation in assurance processes may be limited
- How is the risk of supply chain attacks mitigated? AI agents may generate code that introduces dependencies on components from external package repositories. AI agents therefore require criteria on which specific repositories and packages within these repositories can be robustly evaluated to determine whether or not they should be trusted. In the absence of robust and rigorous evaluation processes, AI agents may introduce dependencies on components that are not trustworthy and place organisations at risk of supply chain attacks. Reputational and financial impacts of such attacks would negate some of the economic benefits on which use of AI agents for software development may be predicated
Read the Guardian article at: https://www.theguardian.com/technology/2026/may/07/your-craft-is-obsolete-wisetech-staff-in-limbo-as-ai-touted-as-better-than-humans
As pointed out in the article below, transformation of governance will be a key factor in determining the success of the proposed ‘Health Data Research Service’.
Some brief thoughts on ways in which the governance of health data for research purposes might be improved:
- Move away from use of abstract terms such as ‘gold standard’ to describe security and privacy measures. Such terms have no meaningful semantics. Additionally, they are likely to elicit different expectations for different parties and therefore cause misunderstandings. Focus is better placed on clearly articulating the security and privacy measures that are in place, and why they are necessary and sufficient in a particular context
- Harmonise the interpretation of legal frameworks and ethical principles applicable to use of health data for research purposes. Legal frameworks and ethical principles define abstract concepts and principles that must be interpreted and then operationalised in a given context. Variation in interpretation and operationalisation between the different organisations in the health data ecosystem is likely to be a source of misunderstanding and lead to ineffective and inefficient decision-making processes
- Develop people and processes to make and clearly articulate informed decisions based on balancing of different (and potentially conflicting) individual, organisational and societal interests. Effective and efficient decision-making processes in the governance of health data are constrained if people and organisations do not have the expertise to confidently make and articulate these inherently complex decisions
UK Government announcement available at: https://www.gov.uk/government/news/prime-minister-turbocharges-medical-research
Read the Lancet article at: https://www.thelancet.com/journals/landig/article/PIIS2589-7500(26)00017-8/fulltext
Worth noting that:
- ‘Anonymised’ data is not ‘anonymous’. Additional technical and organisational measures are typically used in addition to ‘anonymisation’ of the data to ensure measures are adequate to protect privacy. For example, contracts and information security assurances
- No ‘personally identifiable information’ does not mean no risk. Given the nature of the data items that are likely to be included in the data, it may be feasible for individuals or organisations with the willingness and ability to re-identify individuals. For example, using additional data to which they may have access
- UK Biobank data involved in the incident is now potentially ‘in the wild’; simply removing from the listing(s) from a website does not mitigate the continuing risks posed by the potential existence of one or more copies of the ‘anonymised’ data outside of additional technical and organisational measures
UK Biobank update available at: https://www.ukbiobank.ac.uk/news/a-message-to-our-participants-uk-biobank-data-security-update/
Read the BBC article at: https://www.bbc.co.uk/news/articles/cpvxgl3n138o