GDS publishes guidance on AI coding assistants

The government has published guidance for software engineers working in government departments on how they should use artificial intelligence (AI)-based coding assistants.

The Government Digital Service (GDS) AI coding assistants for developers in HMG document warns that if a production service is developed, maintained and deployed from a single environment, using AI coding assistants may introduce unacceptable risks.

“The closer a development platform and deployment infrastructure is to good practice, the less concern you should have about the specific use of AI coding assistants,” GDS said. It recommended that software engineering teams within government departments can “greatly reduce the risks of employing AI coding assistants in their development environment by working in the open and employing main branch protections”.

GDS’ guidance recommends software engineering teams in government departments also maintain the strict separation and audit of production secrets access and use multi-stage deployment, which needs to include sufficient test coverage and vulnerability scanning for continuous deployment in software development pipelines.

Due to the non-deterministic nature of the models underpinning AI coding assistants, the GDS guidance recommends that source code and build pipeline should never rely on a specific response to a prompt unless the software engineering team is willing to test these responses extensively and accept the risk of frequent breakage.

Publication of the guidance follows on from a four-month trial with more than 1,000 software engineers using AI to improve programme productivity.

The Department for Science, Innovation and Technology (DSIT) reported that the pilot shows that AI assistants has the potential to save government software developers the equivalent of 28 working days a year – almost an hour every day.

The boost in efficiency from this AI has meant that more than 1,000 developers who took part in the trial were able to build more software to support government-led digital initiatives. DSIT predicted that AI assistants could help the government build the technology it needs more quickly, targeting £45bn in savings to the taxpayer by making the public sector more efficient.

Developers and engineers across 50 government departments trialled AI coding assistants from Microsoft, GitHub Copilot, and Google, Gemini Code Assist.

The trial found widespread satisfaction with the tools among coders, with 72% of users agreeing they offered good value for their organisation. Over half of participants (58%) said they would prefer not to return to working without AI assistance, while 65% reported completing tasks faster and 56% said they could solve problems more efficiently.

The AI-based coding assistants were used to produce first drafts of source code, which could then be amended by government software engineers, or using them to review existing code. DSIT said only 15% of code generated by the AI coding assistants was used without any edits, showing that engineers were taking care to check and correct AI-generated code where needed.

Technology minister Kanishka Narayan said: “These results show that our engineers are hungry to use AI to get that work done more quickly and know how to use it safely. This is exactly how I want us to use AI and other technology to make sure we are delivering the standard of public services people expect, both in terms of accuracy and efficiency.”