Public‑Sector AI Cautions

Washington is moving into artificial intelligence the way it moved into the cloud: fast, with big promises, and before the plumbing is ready. That is the point of ProPublica’s new warning, published on April 6, 2026. The story does not argue that agencies should avoid AI. It shows something more basic and more troubling. Federal leaders keep treating a hard implementation problem like a branding exercise, even after years of evidence that the government still struggles to secure and govern the digital systems it already has (propublica.org). ProPublica ties that warning to a recent case that has nothing to do with chatbots and everything to do with institutional memory. In March, it reported that FedRAMP, the federal program meant to certify cloud security, approved Microsoft’s GCC High environment for sensitive government data even after internal reviewers said the company’s documentation was so poor they lacked confidence in its overall security posture. One reviewer was blunter. The product, that person said, was “a pile of shit.” FedRAMP still authorized it, with a warning attached, because Microsoft was already too embedded to easily stop (propublica.org). That history matters because AI is arriving on top of the same stack. Agencies do not deploy large models into a vacuum. They plug them into cloud systems, identity systems, procurement systems, and data stores that were already uneven before the AI boom. The Biden White House tried to impose some structure in March 2024, when OMB’s M-24-10 memo required agencies to designate Chief AI Officers, inventory their AI uses, and apply minimum risk-management practices to systems that could affect rights or safety (whitehouse.gov). NIST followed with a generative AI profile in July 2024, and on April 7, 2026, released a new concept note for a critical-infrastructure profile, which is another sign that the rulebook is still being written while deployment continues (nist.gov). The scale of that deployment is no longer hypothetical. A 2025 GAO report found that reported AI use cases across 11 selected federal agencies rose from 571 in 2023 to 1,110 in 2024. Reported generative AI uses jumped from 32 to 282 over the same period. Growth like that would be hard for a well-staffed private company to manage. In government, it is colliding with procurement bottlenecks, skills gaps, and old security debt that never really went away (gao.gov). And yet the demand is real. That is the other half of the story. On April 6, Plug and Play announced the demo day for the New Jersey AI Innovation Challenge, a statewide program backed by the New Jersey Economic Development Authority. The challenge drew 285 applications and selected 10 teams building tools for government services, clinical intelligence, diabetes care, climate resilience, student support, cybersecurity, and energy forecasting. NJEDA has set aside up to $3.34 million in sub-grant funding to help the winners move toward pilots and commercialization in New Jersey (prnewswire.com). That New Jersey program is small compared with the federal machine, but it reveals the missing piece. Public-sector AI is not failing because nobody has ideas. It is failing when institutions skip the dull parts that make ideas safe to use: data quality, security review, procurement discipline, clear ownership, and enough technical staff to understand what they bought. New Jersey’s challenge at least starts with bounded pilots and concrete use cases. Washington keeps talking as if scale comes first. On April 6 in New Brunswick, the 10 winning teams were presenting minimum viable products. In Washington, agencies are still trying to prove they can securely run the maximum imaginable one (prnewswire.com)