A safety report about Anthropic has become a political demand. On November 6, 2026, New York Assemblymember Alex Bores quoted METR‘s new assessment and wrote that every AI company is trying “to fully automate AI research; to allow the AIs to build the next AI.” He called that recursive self-improvement, or RSI, and said “Congress should ban RSI, this year if possible.” The next evening, U.S. Rep. Greg Casar quoted the thread with two words: “Ban RSI.” Critics pushed back just as quickly. Rep. Rich McCormick argued that Democratic proposals to pause advanced AI development and ban artificial superintelligence would “give China the biggest AI advantage of our lifetime.” The exchange is a public call and a quote-tweet, not a newly introduced statute. The earlier Ban Artificial Superintelligence Act, proposed by Sen. Bernie Sanders and joined by Casar, is a separate bill already covered by i10x.
Quick Take
- The demand: Bores asked Congress to ban RSI this year, arguing that labs are racing toward AIs that build the next AI and that “we don’t get another election before AI companies reach this threshold.”
- The echo: Casar posted “Ban RSI.” at 7:25 p.m. Eastern on November 6, quoting Bores. The post drew a large reply thread and is not itself a bill text.
- The mechanism Bores sketched: Human sign-off — and personal criminal liability — on new training runs, major training-procedure changes, or internal deployment for AI research, with possible later off-ramps such as a third-party-accepted safety case.
- The pushback: McCormick said Democrats want to give China “the biggest AI advantage of our lifetime.” Bores replied to similar objections that a U.S. rule can reach foreign models by restricting use of systems trained without the required human sign-off.
What happened
METR published its Anthropic assessment on November 5. i10x covered the measurement in METR: AI Is Roughly Doubling Anthropic’s R&D Speed: on well-specified tasks, capabilities have been improving about twice as fast as between March 2024 and October 2025, with AECI scores rising at 1.9× the prior rate since October 2025 (95% confidence interval 1.5–2.8×) and a wider “would not be surprised” band of about 1.3–3×. METR’s rough guess was that a majority — perhaps about 80% — of the current acceleration comes from AI use, not from faster growth in training compute. This article takes those figures from that earlier report rather than re-deriving them.
Bores treated the paper as evidence that lab timelines are plausible. In a November 6 thread he wrote that OpenAI “says they are on track to do so by March 2028” and that “Anthropic employees say they are on track to do so by 2027.” “The METR report below suggests they are likely right.” He called fully automated RSI “a bad idea given current technology,” because “any mistake or misalignment in current AIs could be multiplied in each generation.”
He presented the demand as a broad coalition rather than a factional one, listing “various versions” of support from Bernie Sanders, Steve Bannon, Ralph Nader, Glenn Beck, Republican and Democratic attorneys general, whistleblower Daniel Kokotajlo, Turing Award winner Yoshua Bengio, Humans First, the Open Markets Institute, Ezra Klein and Princeton computer scientist Arvind Narayanan. “The exact mechanism can be debated. The ban doesn’t have to be permanent either,” he wrote. “This is not the only thing we have to do in regards to AI, far from it! But it is urgent. Congress must ban RSI.”
Casar’s quote-tweet — “Ban RSI.” — went up at 00:25 UTC on November 7, which is 7:25 p.m. Eastern on November 6. It quoted Bores rather than introducing legislative language. The Verge, writing about the METR report on November 7, noted Bores’s call as part of a wider political reaction.
Background: a call, not a new statute
Nothing in Bores’s thread, or in Casar’s quote-tweet, constitutes a newly filed bill. Bores said the mechanism “can be debated” and that criteria for later lifting a ban “can be figured out in the coming years,” while “the ban must happen in the coming months.” In replies he was explicit that “fully automated RSI does not exist, but the labs say it will soon,” and that he was talking about “turning over complete control to AI to build the next AIs,” not about banning AI-assisted coding in general.
That distinction matters because Congress has already seen a related proposal. The Ban Artificial Superintelligence Act, which i10x covered when Sanders introduced it, targets superintelligence rather than the narrower practice Bores is now naming. Casar’s “Ban RSI.” sits in the same political family as that earlier bill; it does not replace or rewrite it.
Bores offered a concrete example of how a ban could work. In the next post of the thread he wrote: “Require a human being to kick off each new model training, make substantial changes to training procedures, or to deploy a model internally for AI research. Make them take personal, criminal liability for what does wrong. We already apply a similar legal standard to grocery store owners or financial auditors.”
A third post listed possible tests for lifting the ban: “1. A safety case accepted by third parties 2. A speed limit on capabilities increases 3. Require that models used for training new AIs be publicly released for a minimum period of time.” Those, he said, can be designed later. The prohibition itself should not wait.
Asked whether RSI could be banned outside the United States, Bores said yes, in effect: if the mechanism is human sign-off on training runs, the United States could then “ban the usage of any model that has achieved a certain level of capability and doesn’t have a human sign-off,” enforced by “strict liability for any harms,” sanctions or cutting off access to US dollars, or criminal penalties — though he added, “I don’t think we would need to go anywhere near that far.” Competition, he argued, is the current driver; take US companies out of a race toward “runaway RSI before we know it’s safe,” and the risk/reward changes for everyone else, “not to mention a lot of foreign progress is based on distilling our models.”
How the argument breaks down
The political fight is not really about METR’s 1.9× point estimate. It is about what that acceleration implies. If AI is already speeding up AI research inside a leading lab, Bores’s camp says the next step — systems that design and train their successors with little human control — should be treated as a distinct, ban-able practice. If other countries will do it anyway, McCormick’s camp says a US ban is a gift to Beijing.
McCormick’s November 7 post put the China argument in his own words: “America’s AI advantage is on the line. Democrats want to give China the biggest AI advantage of our lifetime. They are proposing a complete pause on advanced AI development and a ban on artificial superintelligence.” He asked how to protect Americans “without compromising our national security and technological leadership,” and answered: “By giving American innovators the freedom to keep building, competing, and leading.” He was speaking about an advanced-AI pause and an ASI ban, which overlaps with but is not identical to Bores’s RSI-specific ask.
The Verge reported that the METR paper also revived criticism of Anthropic’s own posture: the company has warned about rapid AI development while appearing to accelerate, and it redacted much of METR’s most informative evidence from the public report. Those transparency disputes belong to the earlier i10x article. They are the backdrop for why Bores felt he could say the report “suggests they are likely right” about near-term RSI timelines.
| Item | What was said | Status |
|---|---|---|
| Bores’s call | “Congress should ban RSI, this year if possible.” Posted November 6, quoting METR’s report | A public thread, not a newly filed bill |
| How he defined RSI | “Fully automate AI research; to allow the AIs to build the next AI.” He later wrote that fully automated RSI “does not exist, but the labs say it will soon.” | He distinguished this from ordinary AI-assisted research |
| Proposed mechanism | Require a human to kick off each new training run, make substantial changes to training procedures, or deploy a model internally for AI research, with personal criminal liability | Offered as an example; “the exact mechanism can be debated” |
| Possible off-ramps | A safety case accepted by third parties; a speed limit on capability increases; a minimum public-release period for models used to train new AIs | “These provisions can be figured out in the coming years. The ban must happen in the coming months.” |
| Casar | Quoted Bores with “Ban RSI.” at 7:25 p.m. Eastern on November 6 (00:25 UTC on November 7) | Casar previously joined Sen. Bernie Sanders on the Ban Artificial Superintelligence Act |
| McCormick | “Democrats want to give China the biggest AI advantage of our lifetime. They are proposing a complete pause on advanced AI development and a ban on artificial superintelligence.” | Posted November 7; he groups an ASI ban with a pause on advanced development |
| METR context (from i10x’s earlier coverage) | AECI growth about 1.9× the pre-October 2025 rate (95% CI 1.5–2.8×), a roughly 2× pace with a 1.3–3× uncertainty band; METR’s rough guess that a majority (perhaps about 80%) of the current acceleration is from AI use | See the separate METR article; this piece does not re-litigate the measurement |
Pros and cons
For users
- Pro: A bright-line rule against handing successor-training to AI systems is easier for the public to debate than a stack of model-by-model safety cases.
- Con: Users of ordinary AI tools may hear “ban RSI” as a ban on useful AI assistance, and Bores’s thread does not yet define the boundary in statutory language.
For developers
- Pro: Bores’s example — a named human signing off on each training run or internal research deployment — would keep a person on the hook for the step that actually creates a more capable model.
- Con: Personal criminal liability for training decisions could chill legitimate research, and the off-ramps he lists (safety cases, speed limits, public release) are still slogans rather than tests.
For companies
- Pro: A uniform US rule would reduce the fear that a single lab is punished for slowing down while rivals race, which is the coordination problem Bores is trying to solve.
- Con: Companies argue, as McCormick does in broader terms, that a US-only constraint hands an advantage to Chinese labs, and Bores’s cross-border enforcement ideas (sanctions, usage bans, liability) would themselves be a major trade and diplomatic fight.
What X is saying
The opening posts are short, and the replies are not.
Every AI company is trying to fully automate AI research; to allow the AIs to build the next AI. This is called Recursive Self Improvement (RSI). OpenAI says they are on track to do so by March 2028. Anthropic employees say they are on track to do so by 2027. The METR report below suggests they are likely right. This is a bad idea given current technology. Any mistake or misalignment in current AIs could be multiplied in each generation. We don’t get another election before AI companies reach this threshold. Congress should ban RSI, this year if possible.
— Alex Bores (@AlexBores) November 6, 2026 · View on X
Ban RSI.
— Congressman Greg Casar (@RepCasar) November 6, 2026 · View on X
America’s AI advantage is on the line. Democrats want to give China the biggest AI advantage of our lifetime. They are proposing a complete pause on advanced AI development and a ban on artificial superintelligence. That raises a serious question. How do we protect Americans without compromising our national security and technological leadership? By giving American innovators the freedom to keep building, competing, and leading. Protecting tomorrow starts today.
— Congressman Rich McCormick (@RepMcCormick) November 7, 2026 · View on X
In the same thread, Bores told critics he was not asking civilization to stagnate: “We don’t have full RSI yet and civilization is moving forward. I said we should ban a specific reckless practice.” To another user he wrote, “Let’s focus on that and not full RSI before we know how to do it safely.”
Critical analysis
Bores is doing something more precise than “stop AI,” and something less precise than introducing a bill. The precision is the target: not chatbots, not coding assistants, but the step at which an AI system is given control over creating a more capable successor. The imprecision is everything else — who counts as the responsible human, what “substantial changes to training procedures” means, how a third-party safety case would be accepted, and how a usage ban on unsigned models would work at the border.
Casar’s two-word endorsement amplified the demand without adding that detail. That is effective politics and thin policymaking. Readers should not treat “Ban RSI.” as the text of a law.
McCormick’s China argument is the standard counter, and it should be quoted as he wrote it, not paraphrased into a cartoon. He is attacking a “complete pause on advanced AI development and a ban on artificial superintelligence,” which is a wider target than Bores’s human-sign-off example. Both sides are, in other words, arguing partly past each other: one about a specific future training practice, the other about American technological leadership in general.
The METR report is doing a lot of work in this debate that the public version cannot fully support. As i10x noted in the earlier article, Anthropic redacted much of the evidence METR found most informative. Bores can fairly say the public findings point to faster AI-assisted research. He cannot fairly say the redacted report, on its own, proves that fully automated RSI is months away.
What’s still open
- Will Bores, Casar or others file legislation that defines RSI and the human-sign-off rule, or will the demand stay on X?
- How would a ban treat today’s AI-assisted research, which METR says already accounts for much of Anthropic’s speedup, without becoming a de facto pause?
- Can the off-ramps Bores listed — third-party safety cases, speed limits, public release of trainer models — be specified well enough to lift a ban later?
- How would the United States try to apply a usage rule to foreign models, and what would China and other labs actually do in response?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis draws on Alex Bores’s November 6 thread and follow-up replies, Greg Casar’s “Ban RSI.” quote-tweet, Rich McCormick’s November 7 post, The Verge’s November 7 report on the METR reaction, and i10x’s earlier article on METR’s Anthropic assessment. METR figures are recapped from that article, not from secondary blogs. This is coverage of a public call, not of a newly filed bill. Related reading: METR: AI Is Roughly Doubling Anthropic’s R&D Speed and the Ban Artificial Superintelligence Act.
- Alex Bores: Congress should ban RSI
- Bores: example mechanism (human sign-off)
- Bores: possible criteria for lifting a ban
- Rep. Greg Casar: “Ban RSI.”
- Rep. Rich McCormick on a China advantage
- The Verge: Why a new report on Anthropic automating AI development is alarming researchers
- i10x: METR: AI Is Roughly Doubling Anthropic’s R&D Speed
i10x Perspective
“Ban RSI” is catching on because it sounds like a firebreak: pick one dangerous future practice and draw a line in front of it. The hard part is that the line sits just beyond a capability METR says is already rising fast with human researchers still in the loop. If lawmakers write a definition that captures only fully automated successor-training, they will have a debateable rule. If they write one that captures today’s AI-assisted R&D, they will have written McCormick’s pause by another name. Until someone files text, the responsible reading of this moment is simpler than the rhetoric: two lawmakers asked Congress to stop a practice that, by Bores’s own account, does not fully exist yet — and a third warned that trying could help China. The argument is now public. The statute is not.
