CIOs warn AI governance risks

State government tech leaders are no longer talking about AI like a future experiment. They’re talking about it like an operational risk sitting inside procurement, security, data policy, and public trust right now. That was the real message coming out of the NASCIO Midyear Conference in Philadelphia, held April 26–29, where CIOs and CISOs kept circling back to the same point: AI is useful, but the hard part is governing it before it spreads everywhere. (nascio.org) ### What changed here? The shift is that AI is no longer just one more innovation topic on the conference agenda. In the conference discussion and in NASCIO’s new cybersecurity study, AI showed up as both a threat vector and a defensive tool — which means leaders have to manage upside and downside at the same time. That is a very different posture from simple “let’s pilot a chatbot” optimism. (govtech.com) ### Why are CIOs treating this as governance first? Because the first failure mode is not magic superintelligence. It’s ordinary institutional mess. Who can buy an AI tool? What data can it touch? Which model is approved? Who owns the output? How do you log decisions, review bias, and shut something off when it goes sid(govtech.com)enterprise controls so innovation does not outrun accountability. (nascio.org) ### Why does AI make cybersecurity harder? Because it compresses time and lowers the skill floor for attackers. The 2026 NASCIO-Deloitte study says foreign adversaries and cybercriminals are increasingly using AI to exploit vulnerabilities, while states are also trying to use AI to strengthen defense. So defenders are being pushed into an arms race they didn’t exactly ask for. (nascio.org)t the same time. (nascio.org) ### What did leaders sound worried about? Confidence. Dan Lohrmann’s conference write-up says CIOs and CISOs talked openly about losing confidence in their ability to stop and recover from cyberattacks, even while governor support remained high in some places. He also captured the practical stress points — budget cuts, unfilled jobs, pressure to prove hard savings, and fear of paying twice while old tools and new AI systems overlap. (govtech.com) ### What’s the “double-bubble” problem? Basically, states want AI-era tooling, but they are still paying for legacy stacks. Nobody wants to run duplicate systems forever. That creates a nasty transition problem: leaders need better visibility, automation, and analytics, but ripping out old systems too fast can create ne(govtech.com)silience. (govtech.com) ### Why does public-sector AI feel different from private-sector AI? Because the state has to answer for fairness, records retention, procurement rules, privacy, and critical services all at once. A bad enterprise AI rollout at a company can be expensive. A bad rollout in government can affect benefits, schools, local i(govtech.com)not just productivity. (nascio.org) ### So what are they actually saying to do? Use AI, but put rails on it early. Start with outcome-focused projects. Tie them to existing systems and measurable value. Clarify who approves tools, who monitors them, and what data boundaries are non-negotiable. In other words — don’t scale first and invent governance later. That sounds slower, but turns out it is the only way state leaders think this stays manageable. (govtech.com) ### Bottom line? The warning from Philadelphia was not anti-AI. It was more practical than that. State CIOs and CISOs seem to be saying the same thing: if AI enters government as a procurement trend, it becomes chaos; if it enters as a governed enterprise capability, it might actually help. (govtech.com)