“We shape our tools and thereafter our tools shape us.” John Culkin
Culkin’s quote welcoming Bloomberg terminal users at sign in one day this week could not have been happenstance. Meanwhile, most statements lately about AI policy have not been memorable, certainly not in pre-election Washington. As a legislative and market risk clearing event, the midterm results will force a recalibration of DC power balance. I see Congress reasserting authority on a few issues regardless of election outcomes, excepting the unlikely outcome of Republicans gaining seats in the House. In this power rebalance, AI safety/governance policy is on a compact list of ‘must do’ agenda items (fiscal, infrastructure, and defense are also on my list).
An inexhaustive roundup of AI-relevant developments continue 2H26 indicators of federal policy action next year.
- At a Senate Homeland Security subcommittee hearing on Wednesday witnesses stressed four common points
- Inadequacy of current voluntary self-regulation
- Urgent need for mandatory independent model testing oversight and incident reporting
- Efficacy of existing safety testing regimes degrades quickly as model capability grows.
- Dangerous gap between “AI capability” and progress in alignment science (reliable performance according to human safety intent during training and deployment). (HERE)
- A bipartisan Senate duo involved in that process plan to introduce legislation mandating states collect and report on data centers’ environmental impacts.
- BoE’s Bailey told a London School of Economics audience that AI buildout leverage will produce losers as well as winners, noting the market currently prices all as winners. He does not favor rigorous industry regulation at present but has instructed the Bank to develop stress tests for lenders and other providers of buildout leverage.
Circumstances matter, but stakeholders eventually will fill the AI policy vacuum. Pace of that policy process and frontier model development matters less than consequences of those activities. Among private and public policy decision makers, voters have access to the latter, which is their only method of influencing the former. Midterm election results and admitted sandbox transgressions – even those delayed despite internal warnings – are the more telling developments to watch for policy directional signals.
A constructive path to collaborative safety oversight presents positive market and commercial opportunities, and garners broad support in polling. This issue will prove easier and more urgent to resolve next year than the data centers challenge if only because of fragmented regulatory oversight in all states. In this scenario, wasted breath and time on ‘doomers versus tech billionaires’ gives way to addressing obvious risks discussed at the Wednesday Senate hearing and every day in most media.
- Congress could benefit from unbiased analysis and advice. From 1972 to 1995, the Legislature relied on the Office of Technology Assessment to bridge insufficient knowledge and pressing national issues. Bringing OTA back next year would address an obvious need related to AI and other complex issues. In 2019, the General Accountability Office, one of four surviving congressional expert analysis shops, opened a division to help fill the void. Congress needs resolute, permanent tech advisor.
- Dodd-Frank created financial market utility designation status for organization providing core market functions, including central clearing activities (Title VIII) designed to mitigate systemic risk. These entities help reduce counterparty risk (enterprise end-user, in the case of AI), price and risk transparency, marketplace standardization (ranking of model capabilities, for example).
Policymakers’ AI oversight challenge suffers from a lack of motivation but not demand. Whether frontier models racing to superintelligence can provide broadly shared good or represent a path to certain destruction elected officials will not avoid responsibility. Regarding federal lawmaking, the Eighth Section of the Constitution obligates Congress to “provide for the common defence (sic) and general welfare of the United States.” AI policy commentary and actions by various governments globally broadly address national security and technological human impacts. Regulating “Commerce with foreign Nations, and among the several States” is another responsibility. AI science and application fit within this governance mold. Distractive noise aside, plausible societal risks include national security, especially biosecurity, cybersecurity, and competition policy. Morally binding commitments from people motivated to win fall short of legally binding rules.
