- AI Week in Review
- Posts
- AI Week in Review - 8/23/25
AI Week in Review - 8/23/25
Public Sector
Chris Kraft
August 23, 2025
The Details
Federal
NIST has released a concept paper and action plan proposing SP 800‑53-based control overlays to secure AI systems, covering use cases from generative and predictive AI to agent systems and developer-specific controls, and has launched a stakeholder Slack channel for feedback.
The GSA launched USAi—a secure AI sandbox for federal employees to test AI tools from OpenAI, Anthropic, Google, and Meta, hosted on government cloud to protect agency data while reducing procurement barriers.
The U.S. Air Force’s DASH wargames—Decision Advantage Sprint for Human‑Machine Teaming—are testing AI microservices to enhance battle management decision-making, enabling rapid operator–vendor collaboration and reducing procurement timelines.
In August 2025, Nvidia and AMD secured U.S. export licenses to sell AI chips (H20 and MI308) to China by agreeing to pay the U.S. government a 15% cut of revenue—with critics warning the deal may breach constitutional and export-control laws.
Google’s new “Gemini for Government” offers federal agencies access to its full AI suite—including NotebookLM and Veo—via secure FedRAMP‑high infrastructure at $0.47 per agency through 2026, under a GSA deal.
Anthropic and the DOE’s NNSA co-developed and deployed a classifier on Claude that distinguishes benign from concerning nuclear‑related chats with approximately 96% accuracy, and they plan to share it industry-wide.
GSA’s new AI testing platform, USAi, is intended as a short-term, shared service for federal agencies to experiment with AI models until market-driven solutions become more viable.
State / Local
Amid the decline in entry-level tech hiring, California’s universities and community colleges are now offering free AI training—backed by Google, Microsoft, and others—to help students adapt to an evolving job landscape.
Pennsylvania’s year-long generative AI pilot engaged 175 employees across 14 agencies using ChatGPT Enterprise—saving about 95 minutes daily per user—while expanding to local governments and piloting AI in permitting and housing, though experts urge caution.
State and local jurisdictions need to establish dedicated AI governance bodies—backed by data governance, executive-level oversight, vendor accountability, and transparency—to responsibly manage risks like privacy, bias, and cybersecurity.
New Jersey’s in‑house NJ AI Assistant allows state employees to quickly generate and summarize documents, yielding dramatic time‑and‑cost savings at roughly $1 per user per month.
After a successful year‑long pilot using ChatGPT Enterprise, during which 175 employees across 14 agencies saved an average of 95 minutes per day, PA is now expanding access to AI tools in government workflows.
States are taking divergent approaches to AI governance—Colorado has passed comprehensive regulation, while others like California, Tennessee, and Utah focus on narrower safeguards for risk, rights, and transparency.
Fairfax County piloted an AI system to triage non‑emergency calls via its non‑emergency line—handling routine inquiries like permits and noise complaints—and plans to expand its use this fall to improve dispatcher efficiency.
Virginia is deploying agentic AI to supercharge a regulatory reform campaign—building on prior progress of cutting 26.8% of requirements—to slash red tape faster and more smartly.
International
Open‑source AI models offer the Global South a more adaptable, transparent, and locally relevant path to development—but U.S.–China competition, export controls, and high capital costs complicate their access and sovereignty.
Albania is piloting AI tools to fight corruption and bolster transparency—public procurement, tax, and customs are already benefiting—and even floated the possibility of an “AI-run ministry.”
South Korea has made AI investment its top policy priority, unveiling 30 major AI and innovation initiatives backed by a ₩100 trillion ($71.6 billion) fund to boost growth under President Lee Jae Myung’s administration.
DeepSeek released an upgraded DeepSeek‑V3.1 model featuring faster hybrid inference and enhanced agent capabilities, optimized for Chinese-made chips.
Everything Else
AI is doubling in performance every six months, upending economies, labor, and politics—escalating U.S.–China strategic rivalry as each grapples with innovation, infrastructure, and policy choices.
Apple is in early discussions with Google to use its Gemini AI to power a long‑delayed Siri overhaul—while the decision remains pending, both companies' stocks rose on the news.
Meta has paused hiring across its AI division, halting a high‑cost recruitment spree and restructuring its Meta Superintelligence Labs into four focused teams as part of broader organizational planning.
OpenAI CEO Sam Altman warned that the current AI investment frenzy may be forming a speculative bubble comparable to the dot-com crash—despite AI’s transformative potential and ongoing infrastructure spending plans.