Can Northern Ireland’s AI Strategy deliver?

Earlier this month, the Executive published its long-awaited draft Artificial Intelligence Strategy. While it recognises that AI can support economic growth and the public sector, the strategy also gives consideration to the vast risks the technology generates. As such, the strategy’s eight principles for responsible AI include human oversight; accountability and redress; data governance; technical safety and security; fairness and transparency; sustainability; societal benefit; and training and literacy.

At times, though, there exists a tension between the document’s appropriate caution about AI’s risks and its far greater confidence in AI’s potential.

The “tortoise and hare” metaphor has been regularly referenced to describe a mismatch between the slow-moving pace of legislation and policymaking and the lightspeed pace of AI development. In my view, the Executive’s strategy represents another instance of retrospective policymaking, attempting to catch up with technology adoption that’s already well underway, with the document admitting that both the Department for Agriculture, Environment and Rural Affairs and the Department for Infrastructure had already adopted AI technology prior to the Executive developing its strategy.

Aside from concerns surrounding pace, difficult questions remain – particularly with respect to the use of AI in the public sector. The strategy claims that AI “does not replace humans”, but instead “supports human decision-making”. That framing is a bit too tidy: AI may not make the final decision, but its use within decision-making processes can still fundamentally shape the outcome. Automation bias, where people are prone to trusting the suggestions and choices of machines over their own judgement, is a well-documented phenomenon within behavioural psychology. In addition, research has suggested that human oversight requirements (like the strategy’s proposal for human oversight teams of “critical thinkers” with “acuity”) not only fail, but can provide false reassurance and legitimacy for algorithmic systems. This approach leaves several vital questions unanswered: Are we able to tell how AI reached a decision? How easy is it to reject that decision? Do rejections have to be justified? Does accepting a decision follow the same process, or is it easier?

This danger has real consequences for our lives. Public bodies in the UK have already experienced serious failures with automated decision-making: the 2020 A-level algorithm had to be abandoned after it produced “inconsistent and unfair” results; the Home Office has faced legal challenges over “discriminatory” algorithms (in 2020 and 2023) and has faced wider scrutiny of its AI usage (2024 and this year); and a Department for Work and Pensions algorithm wrongly flagged 200,000 people for housing benefit fraud and error. In Northern Ireland, faulty and incomplete data – also a risk the strategy acknowledges – caused HMRC to wrongly target nearly 90% of the NI families flagged during a child benefit crackdown.

Humans were still involved in these processes, but their presence ultimately didn’t prevent harmful outcomes. With the draft strategy identifying possible applications in healthcare, justice and education, errors and bias could have immense consequences.

A pressing concern is whether the departments and public bodies expected to procure and manage AI genuinely have the capacity to manage complex technological projects of this kind. The Northern Ireland Audit Office (NIAO) previously found that 14 of 24 (58%) live major IT projects across the NI Civil Service were already experiencing serious delivery problems. Twelve were rated amber, meaning they had significant but potentially resolvable problems, and two were red, meaning that they were unlikely to meet their time, cost and quality objectives and may need to be reconsidered altogether. In almost every case, delays meant that contracts for existing systems had to be extended multiple times simply to keep services running – for almost eight years on average. This raises legitimate questions about whether departments currently have the ability to harness AI safely.

The strategy, described as a “call to collective action and a guiding framework” rather than a “technical manual or an economic blueprint”, is lacklustre on delivery. It does propose a number of “recommended actions”, such as developing public sector standards for AI risk assessment and establishing an AI Citizens’ Panel. But, in familiar Executive fashion, many of the actions are vague – things to be developed or explored, largely without named owners, budgets, or deadlines.

By comparison, Scotland’s strategy identifies 10 clear actions for completion by March 2027, while Ireland’s strategy contains 90 individual deliverables, with a timeline and lead department for each. Under the UK Government’s Algorithmic Transparency Recording Standard (ATRS), publication of information about the algorithmic tools in use is already mandatory for government departments and some public-facing bodies – with this standard launched over four years ago. Northern Ireland proposes a similar “publicly available AI Register of Systems” but doesn’t specify how it will work, or even when it might exist.

None of this is an argument for standing still: the strategy identifies several potential use cases for AI, especially within small businesses and the private sector more generally. Likewise, we should give credit where credit is due – the document displays a clear intent to take ethics and responsible use seriously. However, it’s imperative that the consultation process (open until 7 October) is followed by sustained engagement with not only proponents of AI but those who critically examine its consequences.

Credibility will ultimately depend on knowing where AI is useful and, more importantly, where it isn’t.

Dr Karli Gibson is a research executive at Chambré. She completed her doctorate through the Leverhulme Interdisciplinary Network on Algorithmic Solutions.