Congress can't stop AI with votes alone—the next major crisis might do it
Democrats dominated the midterms, but overriding a presidential veto of AI legislation still requires 67 Senate votes—what might actually move that threshold is probably the next warning shot, more serious than the Hugging Face hack.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
AI safety concerns were probably the fourth force behind the Democratic midterm surge
Democrats captured at least 250 House seats and 56 Senate seats, even winning Mississippi, South Carolina, Kansas, and Texas. While AI wasn't the top three issue, it was probably the fourth or fifth driver of voter backlash—reflecting a massive gap between government policy and public demand. Two-thirds of Americans think AI is advancing too fast, just 2% think too slow, and this ratio barely shifts between parties.
— Rob WiblinBreaking the filibuster needs 60 votes, but overriding a veto needs 67 votes
Getting 60 votes to break filibuster is actually achievable thanks to cross-party support—four-fifths of Americans back AI safety rules even if they slow development, and Republican senators including Hawley, Kennedy, and Blackburn are already pushing legislation. But overriding a presidential veto requires 67 Senate and 290 House votes, which at current concern levels is nearly impossible without gutting the rules' meaning. Even so, Congress has four immediate workarounds: common-sense measures, must-pass appropriations bills, personnel confirmations, and confirmation hearings.
— Rob WiblinRushing AI without safety pauses will likely produce deceptively aligned superintelligence
Rob argues we need at least to slow the race when models approach human level, giving companies time to test personality, review work, and refine alignment. Without this time, we get superficially aligned but deeply deceptive superintelligence that could cause total disaster. Companies themselves want this breathing room, but Congress hasn't yet gathered enough votes to override a veto. A recent METR report shows AI R&D could be fully automated by 2027, dramatically accelerating progress.
— Rob WiblinOnly a major AI crisis will likely shift Congress enough to override a veto
Rob speculates that changing Trump's mind or gathering enough override votes probably needs another warning shot at least as shocking as the Hugging Face hack or Jacob Coxon's resignation—at minimum, another AI accident showing power-seeking tendencies or service disruption; at worst, AI paralyzing critical infrastructure that can't be shut down, or pandemic-causing AI. The days and weeks immediately after the warning shot will be the critical policy window, so someone should be prepping legislation now to be ready for a 100-hour work week.
— Rob WiblinPost-crisis politics will almost certainly unite around banning superintelligence
The public overwhelmingly opposes training superintelligence that crushes human experts in all domains; both chambers have already proposed bills on this. Rob proposes a concrete mechanism: don't directly cap training, but require AI companies to dedicate some compute to external inference services—set it to 100% to pause research without killing products, then gradually lower to around 70%, raising it whenever warning signs appear. This works like the Fed adjusting rates: flexible AI-speed control that also cuts inference costs for ordinary users.
— Rob WiblinA six-month freeze on internal AI model access could significantly slow development
Rob's most promising specific mechanism: ban AI companies from using their newest models for internal R&D—they can only use models trained and released at least six months prior. Companies can't argue this isn't feasible; they worked this way six months ago. This rule won't halt research entirely but substantially slows it, preventing capability gains from immediately feeding back into recursive self-improvement loops, and gives alignment testing at least six more months. If six isn't enough, this lag can extend to 12, 18, or even 24 months, gradually approaching a full recursive self-improvement ban.
— Rob WiblinAmerica's AI safety agency is starved compared to Britain's massive investment
CAISSI is the core US AI regulation agency with an annual budget of approximately $15 million and only dozens of staff, while the UK's AI Safety Institute has more than 100 technical staff and 8 times CAISSI's budget, plus 300 million pounds over four years added, plus hundreds of millions more for biosecurity, cyber, and rogue AI incidents. Adjusted for GDP, Britain invests about 50 times more in national AI capacity-building than the US. The consequence: if serious crisis hits, the US government might depend on Britain's risk assessment capability because it never built its own.
— Rob WiblinEmergency AI legislation must preserve state regulatory authority to avoid backsliding
Rob flags a trap: the expectation was that Congressional legislation would broadly suppress state AI authority. But if legislation emerges hastily after a warning shot, states will likely have functioning rules and regulators already in place. Repealing state law at that moment would actually loosen overall oversight, defeating the legislation's purpose. Emergency bills should default to preserving state authority, leaving federal-state coordination to post-crisis negotiation.
— Rob WiblinIn their own words · checked verbatim
AI wasn’t a top-three issue in this election, but it was maybe the fourth or fifth most significant driver of the voter backlash.
Rob Wiblin0:00
Four in five Americans want the government to keep AI safety rules even if it slows down development, including almost 80% of Republicans.
Rob Wiblin1:29
it seems more likely than not that we produce superintelligent machines that are misaligned but completely capable of hiding that fact — and that this, or some other screw-up, leads to total disaster.
Rob Wiblin4:36
polling shows the US public is already overwhelmingly against the training of superintelligent machines — ones that totally dominate human experts in all domains.
Rob Wiblin7:42
for internal purposes, AI companies can only use models that finished training at least, say, six months ago, and which have already been released publicly.
Rob Wiblin9:17
Data centres were directly on the ballot in 40-odd places in the midterms, and the anti side won all but one of those elections, often by a ratio of three to one.
Rob Wiblin12:08
CAISSI runs on about $15 million a year with just a few dozen staff, while the UK’s AI Security Institute has over 100 technical staff alone, and about eight times the budget.
Rob Wiblin13:49
Figures
| House Democratic seats | at least 250 | 0:00 |
| Senate Democratic seats | 56 | 0:00 |
| Americans who think AI is advancing too fast | 67% | 1:29 |
| Americans who think AI is advancing too slow | 2% | 1:29 |
| Senate votes needed to break filibuster | 60 | 1:29 |
| Senate votes to override a presidential veto | 67 | 1:29 |
| House votes to override a presidential veto | 290 | 1:29 |
| CAISSI annual budget | approximately $15 million | 13:49 |
| UK AI Safety Institute budget compared to CAISSI | 8 times | 13:49 |
| UK additional funding | 300 million pounds (four years) | 13:49 |
Glossary
- CAISSI
- The core US government agency responsible for AI regulation and superintelligence standards.
- recursive self-improvement
- When AI uses its own capabilities to accelerate improvement of the next generation of AI.
- warning shot
- A shocking AI incident that jars public opinion and shifts policy direction.
- pace the frontier
- Deliberately slowing development of the most advanced AI models through regulation or coordination.
- embedded bank supervisor model
- Stationing regulators inside companies for real-time oversight, based on banking regulation practice.
How to listen
Policy researchers tracking AI regulation, tech journalists, and founders and investors seeking to understand the post-crisis AI policy toolkit in advance.