Month: January 2016

[Summary: an extended discussion of issue arising from today’s discussion of UK data sharing open policymaking discussions]

I spend a lot of time thinking and writing about open data. But, as has often been said, not all of the data that government holds should be published as open data.

Certain registers and datasets managed by the state may contain, or be used to reveal, personally identifying and private information – justifying strong restrictions on how they are accessed and used. Many of the datasets governments collect, from tax records to detailed survey data collected for policy making and monitoring fall into this category. However, the principle that data collected for one purpose might have a legitimate use in another context still applies to this data: one government department may be able to pursue it’s public task with data from another, and there are cases where public benefit is to be found from sharing data with academic and private sector researchers and innovators.

However, in the UK, the picture of which departments, agencies and levels of government can share which data with others (or outside of the state) is complex to say the least. When it comes to sharing personally identifying datasets, agencies need to rely on specific ‘legal gateways’, with certain major data holders such as HM Revenue and Customs bound by restrictive rules that may require explicit legislation to pass through parliament before specific data shares are permitted.

That’s ostensibly why the UK Government has been working for a number of years now on bringing forward new data sharing proposals – creating ‘permissive powers’ for cross-departmental and cross-agency data sharing, increasing the ease of data flows between national and local government, whilst increasing the clarity of safeguards against data mis-use. Up until just before the last election, an Open Policy Making process, modelled broadly on the UK Open Government Partnership process was taking place – resulting in a refined set of potential proposals relating to identifiable data sharing, data sharing for fraud reduction, and use of data for targeted public services. Today that process was re-started, with a view to a public consultation on updated proposals in the coming months.

However, although much progress has been made in refining proposals based on private sector and civil society feedback, from the range of specific and somewhat disjointed proposals presented for new arrangements in today’s workshop, it appears the process is a way off from providing the kinds of clarification of the current regime that might be desirable. Missing from today’s discussions were clear cross-cutting mechanisms to build trust in government data sharing, and establish the kind of secure data infrastructures that are needed for handling personal data sharing.

I want to suggest three areas that need to be more clearly addressed – all of which were raised in the 2014/15 Open Policymaking process, but which have been somewhat lost in the latest iterations of discussion.

1. Maximising impact, minimising the data shared

One of the most compelling cases for data sharing presented in today’s workshop was work to address fuel poverty by automatically giving low-income pensioners rebates on their fuel bills. Discussions suggested that since the automatic rebate was introduced, 50% more eligible recipients are getting the rebates – with the most vulnerable who were far less likely to apply to recieve the rebates they were entitied to the biggest beneficiaries. With every degree drop in the temperature of a pensioners home correlating to increased hospital admissions – then the argument for allowing the data share, and indeed establishing the framework for current arrangements to be extended to others in fuel poverty (the current powers are specific to pensioners data in some way), is clear.

However, this case is also one where the impact is accompanied by a process that results in minimal data actually being shared from government to the private companies who apply the rebates to individuals energy bills. All that is shared in response to energy companies queries for each candidate on their customer list is a flag for whether the individual is eligible for the rebate or not.

This kind of approach does not require the sharing of a bulk dataset of personally identifying information – it requires a transactional service that can provide the minimum certification required to indicate, with some reasonable level of confidence, that an individual has some relevant credentials. The idea of privacy protecting identity services which operate in this way is not new – yet the framing of the current data sharing discussion has tended to focus on ‘sharing datasets’ instead of constructing processes and technical systems which can be well governed, and still meet the vast majority of use-cases where data shares may be required.

For example, when the General Records Office representative today posed the question of “In what circumstances would it be approciate to share civil registration data (e.g. Birth, Adoption, Marriage and Death) information?”, the use-cases that surfaced were all to do with verification of identity: something that could be achieved much more safely by providing a digital service than by handing over datasets in bulk.

Indeed, approached as a question of systems design, rather than data sharing, the fight against fraud may in practice be better served by allowing citizens to digitally access their own civil registration information and to submit that as evidence in their transactions with government, helping narrow the number of cases where fraud may be occurring – and focussing investigative efforts more tightly, instead of chasing after problematic big data analysis approaches.

(Aside #1: As one participant in today’s workshop insightfully noted, there are thousands of valid marriages in the UK which are not civil marriages and so may not be present in Civil Registers. A big data approach that seeks to match records of who is married to records of households who have declared they are married, to identify fraudulent claims, is likely to flag these households wrongly, creating new forms of discrimination. By contrast, an approach that helps individuals submit their evidence to government allows such ‘edge cases’ to be factored in – recognising that many ‘facts’ about citizens are not easily reduced to simple database fields, and that giving account of ones self to the state is a performative act which should not be too readily sidelined.)

(Aside #2: The case of civil registers also illustrates an interesting and significant qualitative difference between public records, and a bulk public dataset. Births, marriages and deaths are all ‘public events’: there is no right to keep them private, and they have long been recorded in registers which are open to inspection. However, when the model of access to these registers switches from the focussed inspection, looking for a particular individual, to bulk access, they become possible to use in new ways – for example, creating a ‘primary key’ of individuals to which other data can be attached, eroding privacy in ways which was not possible when each record needed to be explored individually. The balance of benefits and harms from this qualitative change will vary from dataset to dataset. For example, I would strongly advocate the open sharing of company registers, including details of beneficial owners, both because of the public benefit of this data, and because registering a company is a public act involving a certain social contract. By contrast, I would be more cautious about the full disclosure of all civil registers, due to the different nature of the social contract involved, and the greater risk of vulnerable individuals being targetted through intentional or unintentional misuse of the data.)

All of which is a long way to say:

• Where the cross-agency or cross-departmental use-cases for access to a particular can be reduced to sharing assertions about individuals, rather than bulk datasets, this route should be explored first.

This does not remove the need for governance of both access and data use. However, it does ease the governance of access, and audit logs of access to a service are easier to manage than audit logs of what users in possession of a dataset have done.

Even the sharing of a ‘flag’ that can be applied to an individuals data record needs careful thought: and those in receipt of such flags need to ensure they govern the use of that data carefully. For example, as one participant today noted, pensioners have raised fears that energy companies may use a ‘fuel poverty’ flag in their records to target them with advertising. Ensuring that later analysts in the company do not stumble upon the rebate figures in invoices, and feed this into profiling of customers, for example, will require very careful data governance – and it is not clear that companies practices are robust enough to protect against this right now.

2. Algorithmic transparency

Last year the Detroit Digital Justice Coalition produced a great little zine called ‘Opening Data’ which takes a practical look at some of the opportunities and challenges of open data use. They look at how data is used to profile communities, and how the classifications and clustering approaches applied to data can create categories that may be skewed and biased against particular groups, or that reinforce rather than challenge social divides (see pg 30 onwards). The same issues apply to data sharing.

Whilst current data protection legislation gives citizens a right to access and correct information about themselves, the algorithms used to process that data, and derive analysis from it are rarely shared or open to adequate scrutiny.

In the process of establishing new frameworks for data sharing, the algorithms used to process that data should be being brough in view as much as the datasets themselves.

If, for example, someone is offered a targetted public service, or targetted in a fraud investigation, there is question to be explored of whether they should be told which datasets, and which algorithms, led to them being selected. This, and associated transparency, could help to surface otherwise unseen biases which might otherwise lead to particular groups being unfairly targetted (or missed) by analysis. Transparency is no panacea, but it plays an important role as a safeguard.

3. Systematic transparency of sharing arrangements

On the theme of transparency, many of the proposals discussed today mentioned oversight groups, Privacy Impact Assessments, and publication of information on either those in receipt of shared data, or those refused access to datasets – yet across the piece no systematic framework for this was put forward.

This is an issue Reuben Binns and I wrote about in 2014, putting forward a proposal for a common standard for disclosure of data sharing arrangements that, in it’s strongest form would require:

• Structured data on origin, destination, purpose, legal framework and timescales for sharing;
• Publication of Privacy Impact Assessments and other associated documents;
• Notices published through a common venue (such as the Gazette) in a timely fashion;
• Consultation windows where relevant before a power comes into force;
• Sharing to only be legally valid when the notice has been published.

Without such a framework, we are likely to end up with the current confused system in which no-one knows which shares are in place, how they are being used, and which legal gateways are functioning well or not. With a scattered set of spreadsheets and web pages listing approved sharing, citizens have no hope of understanding how their data is being used.

If only one of the above issues could be addressed in the upcoming consultation on data sharing, then I certainly hope progress could be made on addressing this missing piece of a robust common framework for the transparency principles of data sharing to be put into practice.

Towards a well governed infrastructure?

Ultimately, the discussion of data sharing is a discussion about one aspect of our national data infrastructure. There has been a lot of smart work going on, both inside and outside government, on issues such as identity assurance, differential privacy, and identifying core derived datasets which should be available as open data to bypass need for sharing gateways. A truly effective data sharing agenda needs to link with these to ensure it is neither creating over-broad powers which are open to abuse, nor establishing a new web of complex and hard to operate gateways.

Further reading

My thinking on these issues has been shaped in part by inputs from the following:

Data & Discrimination – Collected Essays

White House Report on Big Data, and associated papers/notes from The Social, Cultural & Ethical Dimensions of “Big Data.” conference