When an AI system makes a mistake, citizens don't see an isolated error. They assume the same failure could happen again tomorrow - and the day after - until someone finds and fixes the underlying problem. That perception gap, argues Snowflake Field CTO Fawad Qureshi, means accountability will become one of government's most important AI capabilities, not raw intelligence.
"If a machine makes a mistake, we are a lot less tolerant because we know that mistake has occurred because of some systematic flaw, some data gap or some bias," Qureshi said. "Unless somebody fixes it, it could be repeated millions of times."
As governments embed AI into public services - helping clinicians review medical images, supporting caseworkers, powering customer service - that distinction matters. AI can improve services and speed decisions. But if citizens cannot understand why a recommendation was made, trust evaporates as quickly as efficiency improves.
Accountability predates AI
Qureshi argues the principle of accountability isn't new. It has always sat at the heart of public administration. If a visa application is refused or a planning application rejected, people expect an explanation. They may not agree with the outcome, but they expect government to justify how it got there.
AI should not lower that standard. It should raise it.
"If I apply for a visa, I have a right to ask the Home Office why my visa was rejected," Qureshi said. "Whenever a government decision is being made... it needs to be backed by evidence, and that traceability needs to be built into the model."
Every AI-assisted decision should be supported by what Qureshi calls a "data supply chain" - a complete record of where data originated, how it was processed, which models contributed, and who approved the final decision. Lineage and audit trails are not technical concerns. They are critical public infrastructure.
More AI agents, more complexity
The challenge grows as AI becomes more autonomous. Tomorrow's public services may rely on multiple AI agents exchanging information across departments before producing a recommendation. Understanding who is responsible will become significantly more complex.
"The more moving parts you have in the chain, the more imperative it becomes to know who owns the model, who owns the data, who approved the prompt, who validated the recommendation and who signs off," Qureshi said.
He recommends governments move beyond shared responsibility toward a "shared destiny" model, where failures anywhere within a connected AI ecosystem can undermine public confidence in every organisation involved.
Lessons from past governance failures
Recent history shows why explainability matters. Qureshi cites the Post Office Horizon scandal as a case study in what happens when technology becomes difficult to question. "It wasn't an AI failure," he said. "It was an accountability failure. It was a governance problem." The lesson is not that governments should avoid technology, but that they should always be able to explain, audit and challenge the decisions technology helps produce.
He draws similar lessons from algorithm-generated A-level grades during the COVID-19 pandemic. Public confidence evaporated because people could not understand how decisions affecting students had been reached. When systems operate as opaque black boxes, controversy is almost inevitable.
Why this matters for government professionals
For civil servants deploying AI today, the takeaway is concrete: do not prioritise speed and automation over traceability. Every process you build should include a verifiable record of data origins, model contributions, and human signoffs. Without that, the tools you develop will face public resistance regardless of how well they perform. As Qureshi put it, trust will become government's competitive advantage. "The trusted governments will be the ones who build the most accountable AI." If you work in AI for Government, the technical work of auditability is the political work of earning trust.
Your membership also unlocks:
