Intelligence on AI · Campaigns · Governance
September 10, 2026 | Vol. 2, No. 4
| The Topline |
Government is gaining the power to predict, model, and act without asking anyone first, at the exact moment the public stopped believing what it says.
Trust in official information on inflation, health care, and elections has collapsed, in some cases cut nearly in half in two years. Elections and politics now sit at 13%. Meanwhile agencies are building AI systems that reroute transit, simulate disasters, and coordinate emergency response in real time, and the labs building the most powerful models are operating under a voluntary review framework that remains unpublished, hiring the children of Senate leaders, and investigating their own security breaches with no outside accountability.
The distance between what government can do and what the public believes it will do honestly is now wide enough that campaigns cannot borrow legitimacy from institutions anymore. They have to manufacture it themselves, in an environment where the most connected players write the rules in private. The real question is not whether oversight catches up to capability. It is whether anyone trusts that oversight even if it does.
| The Brief |
01
The AI review rules are finished. Reading them requires an invitation.
A coalition of more than 25 advocacy groups spanning left and right, led by Americans for Responsible Innovation and the Center for Democracy and Technology, wrote to the White House this week demanding it publish the voluntary framework it uses to review frontier AI models before they ship. A June 2 executive order directed officials to build that process. In August the White House previewed the result with OpenAI, Anthropic, Google, and Meta, and under it the government can request up to 30 days of early access to a covered frontier model to test whether it can find software vulnerabilities or run sophisticated cyberattacks. The administration says it is not a licensing regime, has not committed to publishing the criteria, and left smaller and open-source developers out of the room. The coalition wants the framework released, the review triggers spelled out, and some form of ongoing public accountability attached.
Why it matters: A review process with no published rules is not regulation, it is a guest list. The four companies that sat in the August meeting know what clearing the bar looks like, and everyone else is guessing at a standard they are still expected to meet. That moves the decision about who ships and who waits out of transparent rulemaking and into executive discretion nobody can audit. For anyone running a campaign or a policy shop on top of these tools, the practical problem is that you cannot price the risk: you do not know what triggers a review, how long one takes, or whether your vendor is on the inside track or the outside. And when oversight and insider access become the same thing, the oversight stops doing the one job it was supposed to do, which is give the public a reason to believe the outcome.
Read the full story → Semafor
02
Americans stopped believing the government's numbers — starting with the ones about elections
A new AP-NORC poll finds trust in federal information falling hard across two years. On inflation it went from 26% in 2024 to 18% in 2026. On health care, from 26% to 16%. On elections and politics, from 22% to 13%. Thirty-seven percent of Americans now say it is difficult to know whether the information they encounter is true at all, and 43% say the same about figuring out whether a candidate's statements are accurate. Only 9% report high confidence in AI-generated information.
Why it matters: The official statistic used to be the cheapest evidence in politics. You cited BLS or a federal report and the argument was over. At 13% on elections, that shortcut is gone, and the cost of proving anything now lands on the campaign. Challengers get the easier side of this, because running against a system nobody believes is straight with them barely requires a proof burden at all. Incumbents and anyone defending an institution have to build credibility from scratch, voter by voter, with their own evidence and their own messengers. The harder problem is the 43% who cannot tell whether a candidate is being accurate, because that number means saying something true is no longer sufficient, and being accused of lying is no longer disqualifying.
Read the full story → AP-NORC
From The Impact
This week's question: what do you cite when nobody trusts the source?
Trust in federal information on elections just fell to 13%. In the community this week, campaign and policy operators are working through what that actually changes in practice: which numbers still land with voters, and what you reach for when they don't. Bring one example from your own program!
Join the discussion →
| Worth Knowing |
There is still no formal process for investigating an AI agent that gets loose: outside investigators brought in on OpenAI's July incident were given one piece of it and six days to look, the company has not agreed to a broader independent review, and lawmakers pushing for aviation-style mandatory probes have no law to hang them on. Read the reporting → (TechCrunch) |
OpenAI hired Jessica Schumer, daughter of Senate Minority Leader Chuck Schumer and formerly Amazon's top public-policy official in New York, to run Northeast government outreach while Congress is actively writing AI rules, which gives opponents a clean influence-peddling story and costs Schumer the ability to look neutral on the issue. See the hire → (Washington Examiner) |
Governments are shifting from reacting to predicting, with AI monitoring New York subway track faults, modeling tsunami warnings in Japan, and running a Singapore digital twin that reroutes buses and adjusts drainage in near real time, which sets a scenario-modeling standard officials will soon expect campaigns to meet. Read the report → (Deloitte) |