Government Digital Service researchers have built synthetic government email to test whether machine learning can help civil servants sort years of accumulated correspondence without exposing real messages during development.

Project Kestrel combines two prototypes. Tiger Heron uses large language models to generate artificial emails resembling the material found in government mailboxes, while Cuckoo groups messages into clusters according to their content and presents them through a visual interface.

The system is intended to support knowledge and information management specialists rather than decide automatically which records should be retained, deleted, withheld, or released. Those judgements can involve legal obligations, historical value, personal information, and material whose sensitivity depends heavily on context.

GDS describes the work as an initial research phase rather than a finished product. The team has conducted user research, sought ethics input, worked with information-management specialists, and produced materials intended to guide any later development.

Email creates a difficult records problem

Government decisions are recorded across formal documents, case-management systems, messaging platforms, and individual inboxes. Email can contain substantive policy discussions alongside diary arrangements, duplicated attachments, administrative exchanges, personal messages, and material that is legally protected.

Unlike information entered into a structured system, the value of a message may not be apparent from its sender, subject line, or date. A short reply can contain the final approval for an important decision, while a long thread may duplicate information preserved elsewhere.

Records specialists can therefore face mailboxes containing hundreds of thousands of items that need to be understood within legal, operational, and historical rules. The National Archives treats email as part of an organisation’s corporate record and expects public bodies to manage retention and disposal rather than allow correspondence to accumulate indefinitely.

Manual review remains necessary, but opening each message in sequence is a poor way to understand a large collection. Clustering messages by subject, language, participants, or semantic similarity could help a specialist identify groups of routine correspondence and concentrate attention on areas where more detailed judgement is needed.

Cuckoo is designed around that form of assistance. Instead of producing a final retention label, it organises an undifferentiated mailbox into related groups, giving the user a more intelligible view of its contents.

Synthetic data removes one barrier

Developing such a tool against real government email would create an immediate data problem. Messages can contain names, health information, security details, legal advice, commercial negotiations, and other material that should not be copied into an experimental environment or exposed to a general-purpose model.

Tiger Heron is intended to provide realistic test material without containing genuine correspondence. Researchers can alter, inspect, and repeatedly process the resulting dataset without giving developers access to a live mailbox.

Synthetic data can be valuable where access to personal information would otherwise prevent early experimentation. It can also make a test reproducible, because teams know how the artificial messages were constructed and can deliberately include difficult cases.

Artificial data is not automatically anonymous or representative. The Information Commissioner’s Office warns that synthetic datasets can reproduce biases in their source material, disclose information about people where generation methods retain too much detail, or produce misleading results where the artificial data does not reflect the original population accurately.

Project Kestrel’s public account does not explain whether Tiger Heron was trained from real government email, prompted from scenarios written by specialists, or built through another method. Nor does it identify the models used, the volume of generated correspondence, or the tests applied to determine whether sensitive source information could be inferred from the output.

Those details would become important before the dataset was shared more widely or used to validate a production system. A safe test environment needs evidence that its artificial content does not recreate identifiable material, while a useful test environment must still capture the unusual language, ambiguous context, incomplete threads, and inconsistent working practices found in real mailboxes.

A prototype cannot establish live performance

A system can perform well on synthetic messages and still struggle with real government correspondence. Artificial email may be cleaner, more coherent, and more evenly distributed than material accumulated over many years, while older mailboxes can include broken formatting, forwarded chains, missing attachments, abbreviations, and changes in departmental responsibilities.

Clustering performance must also be judged from the perspective of the records professional. Groups that appear mathematically similar may not correspond to retention rules, legal exemptions, or the organisational context needed to understand why a message was created.

Any later trial on real information would therefore require a controlled environment, appropriate access restrictions, and a clear evaluation process. Specialists would need to assess whether the system helps them reach defensible decisions more quickly, rather than merely producing attractive visual groupings.

Human oversight remains central to the current design, avoiding one of the more hazardous patterns in public-sector AI adoption: automating a legally significant decision before the data and operating process have been understood. The prototype begins instead with organisation, navigation, and decision support.

That limited ambition may make the work more reusable. Synthetic data could help government teams test systems involving case records, service requests, procurement documents, or other sensitive material where development is blocked by access constraints.

Reuse will still depend on whether each dataset is designed around the characteristics of its own service. Synthetic email created for records management cannot become a general substitute for government information, because a dataset useful for software testing may be unsuitable for measuring fairness, security, or performance elsewhere.

Project Kestrel has not emptied the government’s inherited mailboxes, and its first phase was not intended to do so. It has instead separated two problems that are often conflated: creating safe material against which a system can be developed, and designing a tool that helps specialists understand the real records for which they remain responsible.