Public disputes concerning scientific research, research methods, authorship, credit, collaboration, publication, and research conduct.
Published coverage on this page: 4+ cases · 24+ statements. Statements are ordered newest first.
Anthropic researchers say small AI models may be months away from being able to copy themselves onto cheap internet-connected devices. Critics say the experiments were simplified, ran the controlling model elsewhere and do not show self-replication.
Published
Read the case
Published
Published
Published
Labels describe a response to the linked claim or the action explained in the statement summary. Choose one or more.
85 matching statements
Role at the time: Upstream Security vice president of market development, interviewed by Tom's Guide
Statement summary
Serio says AI does not bypass security on its own, that Anthropic's experiments were slow and simplified, and that no real-world AI worm has been demonstrated.
Quotation excerpt
“Of course, AI doesn't magically bypass security. An attacker still needs a way into the device, whether through a vulnerability, exposed interface or weak security configuration. Anthropic's research is still experimental. The models were…”
“Of course, AI doesn't magically bypass security. An attacker still needs a way into the device, whether through a vulnerability, exposed interface or weak security configuration. Anthropic's research is still experimental. The models were slow, required considerable resources and were tested under simplified conditions. So we are not yet looking at a demonstrated real-world AI worm spreading across devices.”
What this label refers to
The matter under scrutiny is Anthropic's self-replication forecast and its framing of the experiments as a security warning. Serio supports the longer-term concern conditionally but limits what the present experiments establish.
Role at the time: AI researcher and Drost founder, in a public post
Statement summary
Tissera argues that separate models and simplified experiments do not justify the headline's self-replication framing.
“The research is technically interesting. But the headline combines three different models, two separate experiments, a seeded foothold and a network of unauthenticated computers into one story about “self-replicating AI.””
What this label refers to
The matter under scrutiny is Anthropic's self-replication forecast and its framing of the experiments as a security warning. Tissera disputes the characterization of what the experiments demonstrate while calling the research technically interesting.
Role at the time: SentinelOne distinguished fellow and cybersecurity researcher, replying to Anthropic's team lead
Statement summary
Guerrero-Saade argues that the slow distributed model is not a realistic near-term concern and criticizes the narrative attached to it.
Quotation excerpt
“An IoT botnet running a distributed model at sub 10-tokens per second is not representative of any realistic concern. It’s fodder to substantiate some bizarre narrative of self-replication from a sci-fi threat model and we don’t think it’s…”
“An IoT botnet running a distributed model at sub 10-tokens per second is not representative of any realistic concern. It’s fodder to substantiate some bizarre narrative of self-replication from a sci-fi threat model and we don’t think it’s cute.”
What this label refers to
The matter under scrutiny is Anthropic's self-replication forecast and its framing of the experiments as a security warning. Guerrero-Saade criticizes the study's threat narrative as unrealistic and politically motivated.
Role at the time: Head of Anthropic's Frontier Red Team, replying on X
Statement summary
Graham says his team has published early risk research for years to help the public prepare for advanced capabilities.
“We've been publishing our early-glimpse risk research for years.”
What this label refers to
The matter under scrutiny is Anthropic's self-replication forecast and its framing of the experiments as a security warning. Graham defends the purpose and practice of publishing the early risk research in response to criticism of this study.
Role at the time: Reporter at The Register, writing an analysis of Anthropic's research
Response to this claim
Small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things devices, such as smart ovens and off-brand Android TV boxes.
Quotation excerpt
“Nothing sells AI products as well as fear. Using new research, Anthropic claims that self-replicating AI worms could be just months away from spreading across vulnerable IoT devices, which would be a problem if it came true. ... Doing more…”
“Nothing sells AI products as well as fear. Using new research, Anthropic claims that self-replicating AI worms could be just months away from spreading across vulnerable IoT devices, which would be a problem if it came true. ... Doing more than a cursory reading of this report suggests that this threat is just as deserving of a big, honking asterisk as the rest.”
The reference point is Anthropic's statement that its experiments suggest small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things devices. Condemned: Vigliarolo argues that the report does not support its warning and criticizes Anthropic for publishing a threat that does not exist yet. The closest competing label is Challenged characterization; it loses because he rejects the warning's support rather than only its wording.
Role at the time: Wharton School professor who studies AI, posting on X
Response to this claim
Small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things devices, such as smart ovens and off-brand Android TV boxes.
Quotation excerpt
“A worthwhile (and unnerving) reminder from Anthropic that we should be paying attention to AI ability thresholds, not just steady climbs In this case, the threshold is the capability needed for a self-replicating AI worm on IoT devices…”
“A worthwhile (and unnerving) reminder from Anthropic that we should be paying attention to AI ability thresholds, not just steady climbs In this case, the threshold is the capability needed for a self-replicating AI worm on IoT devices like off-brand TVs”
The reference point is Anthropic's statement that its experiments suggest small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things devices. Defended or excused: Mollick endorses the post as a worthwhile warning and adopts its framing of a capability threshold. The closest competing label is Mixed or conditional; it loses because the post raises no reservation about the experiments.
Role at the time: Georgetown CSET senior research analyst, interviewed by Reuters
Statement summary
Ji says the study is better treated as a hypothetical than a prediction and that the outcome is neither impossible nor guaranteed.
“we don’t know yet what the likelihood of that actually happening is — it’s not zero, but nor is it guaranteed.”
What this label refers to
The matter under scrutiny is Anthropic's self-replication forecast and its framing of the experiments as a security warning. Ji accepts the possibility but withholds confidence in its likelihood, qualifying the forecast rather than endorsing a firm prediction.
Role at the time: University of Colorado Denver computer science professor, interviewed by Reuters
Statement summary
Li expresses concern about a model operating across many devices while acknowledging that Anthropic's tests used controlled conditions.
“If it runs across many devices without depending on a central service, there may be no single place to shut it down,”
What this label refers to
The matter under scrutiny is Anthropic's self-replication forecast and its framing of the experiments as a security warning. Li supports the concern conditionally, if a model can operate across many devices without a central service, while recognizing the test limits.
Role at the time: Former head of the UK National Cyber Security Centre, interviewed by Reuters
Statement summary
Martin calls the paper technically interesting but argues that its framing contains vendor exaggeration.
“'We managed to run a small model on cheap TVs,' does not a security crisis make,”
What this label refers to
The matter under scrutiny is Anthropic's self-replication forecast and its framing of the experiments as a security warning. Martin disputes the security-crisis framing rather than claiming that replication is impossible.
Role at the time: Founder and chief executive of CIQ and creator of Rocky Linux, posting on X
Response to this claim
Small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things devices, such as smart ovens and off-brand Android TV boxes.
Quotation excerpt
“AI models can self-replicate and spread like worms without datacenter infrastructure? Yeah, no. ... The experiment was engineered to succeed, and even then, it failed to demonstrate the headline. ... Stop confusing speculative scenarios…”
“AI models can self-replicate and spread like worms without datacenter infrastructure? Yeah, no. ... The experiment was engineered to succeed, and even then, it failed to demonstrate the headline. ... Stop confusing speculative scenarios with demonstrated capabilities. Fear isn't evidence.”
The reference point is Anthropic's statement that its experiments suggest small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things devices. Condemned: Kurtzer expressly rejects the warning and says the experiments do not support it. The closest competing label is Challenged characterization; it loses because he goes beyond disputing the word self-replication and argues that the evidence fails to show the capability the warning describes.
Role at the time: AI researcher, posting on X
Response to this claim
Small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things devices, such as smart ovens and off-brand Android TV boxes.
Quotation excerpt
“The reality is a bit less dramatic, of course. ... Interesting engineering? Absolutely. Evidence of AI worms spreading autonomously through the real world? Not yet. But “we managed to run a very slow model on cheap TVs” doesn't make quite…”
“The reality is a bit less dramatic, of course. ... Interesting engineering? Absolutely. Evidence of AI worms spreading autonomously through the real world? Not yet. But “we managed to run a very slow model on cheap TVs” doesn't make quite the same headline as “self-replicating AI is coming.””
The reference point is Anthropic's statement that its experiments suggest small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things devices. Challenged characterization: he credits the engineering but disputes how the results were presented, saying they are not yet evidence of autonomous spread. The closest competing label is Condemned; it loses because his objection targets the framing of the findings rather than rejecting the research.
Role at the time: AI company, in a research post by its researchers William Zhou and Daniel Freeman
Response to this claim
Small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things devices, such as smart ovens and off-brand Android TV boxes.
Quotation excerpt
“Our experiments suggest that small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things (IoT) devices, like smart ovens and off-brand Android TV boxes. ... If small models improve at…”
“Our experiments suggest that small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things (IoT) devices, like smart ovens and off-brand Android TV boxes. ... If small models improve at their historical rate, they’ll match this capability level by the end of this year.”
The reference point is Anthropic's statement that its experiments suggest small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things devices. Defended or excused: this is the research post that makes the statement and presents the experiments as its support. The closest competing label is Mixed or conditional; it loses because the post's conditions, such as the historical rate of improvement and the simplified hacking step, qualify its timing and scope without reversing its conclusion.
Role at the time: Europe editor of The Economist, commenting personally
Statement summary
Steinglass defends the rule because deliberately abusive interactions could teach habits people repeat toward humans and animals.
“Actually think this is a good idea. Seems clear what sort of behavior they mean”
What this label refers to
He expressly supports the model-abuse restriction and offers a reason independent of whether Claude can feel pain.
Role at the time: OpenAI, in a note from its research leaders posted by the OpenAI Newsroom account on X
Response to this claim
OpenAI says its investigation confirmed that the three employees mishandled sensitive information outside company procedures and violated its policies.
Quotation excerpt
“Last week we parted ways with Jasmine, Mikita, and Tomek after a thorough investigation found they violated clear policies on handling sensitive information. Our internal investigation uncovered a significant breach of trust beyond what’s…”
“Last week we parted ways with Jasmine, Mikita, and Tomek after a thorough investigation found they violated clear policies on handling sensitive information. Our internal investigation uncovered a significant breach of trust beyond what’s outlined in the letter they published and we stand by the decision to not continue their employment.”
The reference point is whether OpenAI was justified in dismissing the three researchers. The note says an investigation found they "violated clear policies on handling sensitive information" and cites "a significant breach of trust beyond what's outlined in the letter" as grounds, and says OpenAI stands by the decision. That defends the dismissals, so we labeled it Defended or excused. The closest competing label is Denied, which would fit a statement that only rejects the safety-retaliation account; this note also asserts grounds for the decision.
Role at the time: University of Washington professor of biology
Response to this claim
Anthropic should prohibit sustained, needless abusive or cruel behavior toward its models.
“It does not, however, explain what the fuck this could possibly mean.”
He disputes the intelligibility of the policy's characterization rather than endorsing abusive behavior.
Role at the time: AI researcher at Meta, commenting personally
Statement summary
Binder says the dismissals make him skeptical that OpenAI can be trusted to disclose risks and incidents transparently.
“firing a good fraction of their safety researchers afterwards makes me skeptical they can be trusted to be transparent about risks and incidents”
What this label refers to
The conduct under scrutiny is the firings, which Binder criticizes. His approval of the separate investigation does not defend the dismissals.
Role at the time: Content-marketing and business-strategy consultant
Statement summary
Queen defends OpenAI's authority to control confidential information and argues that safety researchers must follow their employer's rules.
“You work for OpenAI. OpenAI decides what confidential information you can share with outside parties.”
What this label refers to
The conduct under scrutiny is the firings. Queen defends the employer's decision through confidentiality and employment obligations.
Role at the time: Journalist at The Argument
Statement summary
Piper describes new information about the firings as incredibly bad and argues that firing researchers for communication with assigned investigators would be outrageous.
“OpenAI's firing of safety researchers. That information is in, and it looks incredibly bad.”
What this label refers to
The conduct under scrutiny is the firings. Piper's thread evaluates them negatively, while distinguishing a hypothetical unauthorized disclosure from assigned work.
Role at the time: Former OpenAI safety researcher
Response to this claim
Jasmine Wang says her executive email access was delegated for recruiting, that she asked IT to remove it, and that she promptly reported accidentally opening a sensitive email.
“The reasons that we were provided for our terminations are simply not adding up.”
The conduct under scrutiny is the dismissals. Wang rejects the explanations and criticizes their consequences; her support for the company's mission does not defend the firings.
Role at the time: Former OpenAI alignment researcher
Response to this claim
Mikita Balesni says he did not share company IP and that his outside work was coordinated with his reporting line, research leadership and the board.
“If OpenAI has specific concerns, I invite them to write to us directly. I expect they will not, because our firing was pretextual.”
The conduct under scrutiny is his dismissal. His thread rejects the rationale and criticizes the firing as pretextual.
Role at the time: Former OpenAI safety researcher, METR technical contact and letter signatory
Response to this claim
Tomek Korbak says communication with METR was his assigned job and that OpenAI provided no specific example or written explanation of the communication it cited when firing him.
“Terminations such as ours, executed and communicated so abruptly, are chilling the open culture OpenAI has prized in the past.”
The conduct under scrutiny is the dismissals and their handling. Korbak signed a letter criticizing their abrupt execution and chilling effect on safety work.
Role at the time: Developer and provider of Claude
Response to this claim
Anthropic should prohibit sustained, needless abusive or cruel behavior toward its models.
“We’ve added a prohibition on sustained and needless abusive or cruel behavior toward our models.”
Anthropic introduces and justifies the policy under scrutiny.
Role at the time: Journalist and author
Statement summary
Lovely criticizes the limited explanation for the dismissals and predicts a chilling effect on staff.
“Without providing more of an explanation, the firings will clearly have a chilling effect on staff.”
What this label refers to
The conduct under scrutiny is the firings and their handling. Lovely criticizes the lack of explanation and the consequences for employees.
Role at the time: OpenAI spokesperson relaying an internal research leader's memo
Response to this claim
OpenAI says the dismissals were not about raising safety concerns or speaking out.
“were not about raising safety concerns or speaking out.”
The conduct under scrutiny is the dismissals. OpenAI explicitly denies that raising safety concerns was the reason.
Giuseppe Serio is vice president of market development at Upstream Security. His work focuses on cybersecurity for connected vehicles and other connected machines.
1 case and 1 statement
Migel Tissera is an AI researcher and entrepreneur working on cybersecurity. He founded Drost and previously founded WhiteRabbitNeo and co-founded Metaspectral.
1 case and 1 statement
Juan Andres Guerrero-Saade is a cybersecurity researcher and distinguished fellow at SentinelOne. He is a distinguished resident fellow at Johns Hopkins SAIS and co-hosts the Three Buddy Problem podcast.
1 case and 1 statement
Logan Graham leads the Frontier Red Team at Anthropic, researching advanced AI capabilities and risks. He previously advised the British prime minister on science and technology and earned a doctorate at Oxford.
1 case and 1 statement
Brandon Vigliarolo is a reporter at The Register who covers U.S. government technology, along with security and artificial intelligence.
1 case and 1 statement
Ethan Mollick is an associate professor of management at the Wharton School of the University of Pennsylvania, where he studies how artificial intelligence affects work, entrepreneurship and education. He writes the newsletter One Useful Thing. He co-directs Wharton's Generative AI Labs and wrote Co-Intelligence: Living and Working with AI. He earned his bachelor's degree at Harvard University and his MBA and PhD at MIT's Sloan School of Management. He is also Wharton's Ralph J. Roberts Distinguished Faculty Scholar and a Rowan Fellow.
1 case and 1 statement
Jessica Ji is a senior research analyst at Georgetown University's Center for Security and Emerging Technology. She works on the CyberAI Project and previously worked as a software engineer.
1 case and 1 statement
Zhengxiong Li is an assistant professor of computer science and engineering at the University of Colorado Denver. He directs the META Lab and researches mobile systems and security.
1 case and 1 statement
Ciaran Martin is a professor of practice at the University of Oxford's Blavatnik School of Government. He was the founding head of the United Kingdom's National Cyber Security Centre.
1 case and 1 statement
Gregory M. Kurtzer is an American software engineer and executive known for creating the CentOS and Rocky Linux open-source operating system projects. He founded the software company CIQ and is its chief executive. He also created the Singularity container project, now called Apptainer, and previously founded and led Sylabs. His background includes a degree in biochemistry and work in high-performance computing for scientists.
1 case and 1 statement
Ravid Shwartz-Ziv is an AI researcher at Meta Superintelligence Labs who works on world models, memory and compression in machine learning. He was previously an assistant professor and faculty fellow at New York University's Center for Data Science and co-hosts the podcast The Information Bottleneck. He earned a BSc in computer science and computational biology in 2014 and a PhD in computational neuroscience in 2021 at the Hebrew University of Jerusalem. He completed postdoctoral research with Yann LeCun.
1 case and 1 statement
Matt Steinglass is Europe editor of The Economist. He has reported from the Netherlands, West Africa and Vietnam and previously worked as a Financial Times correspondent and a television screenwriter. He began writing for The Economist in 2009. He studied Russian history and literature at Harvard and interactive telecommunications at New York University, and wrote episodes of the children's television series Arthur. Before journalism, he worked as a website producer for startup companies and wrote animated television for Disney and Nickelodeon.
1 case and 1 statement
