The debate over how quickly frontier artificial intelligence should advance and how it should be regulated is widening. Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have taken unusually similar positions, arguing that companies should slow frontier-model development when necessary and adopt independent assessments and more consistent safety standards. Meta CEO Mark Zuckerberg and Nvidia CEO Jensen Huang, by contrast, say safety requirements do not have to conflict with technological progress and that the industry may not need new laws or regulatory frameworks.
The dispute matters to financial markets because decisions about safety testing, corporate responsibility and disclosure could affect research timelines, compliance costs and the valuations of companies investing heavily in computing capacity and model training. The question of who sets the standards, and whether outside organizations should be involved, is no longer only a technical issue. It also touches competition and access to the market.
Amodei and Altman Back Independent Assessments
Amodei has proposed that frontier-AI companies control the pace of development when necessary, allow independent evaluators to examine advanced systems and development processes, and coordinate around shared safety standards. Altman has broadly supported the proposals and said OpenAI would also commit to bringing independent evaluators into its internal processes.
Zuckerberg responded in a lengthy social-media post, arguing that AI companies already have strong commercial incentives to build systems that meet user expectations, avoid harm that could expose them to legal liability, and delay product launches when systems are not ready. Users, he said, would not continue relying on AI agents whose behavior did not match expectations, making trust and reliability potential competitive advantages. Laboratories that neglect safety could fall behind, while companies responsible for harm could face substantial legal exposure.
Zuckerberg said Meta delayed the launch of its Muse AI agent by several months to strengthen safeguards and conduct additional testing. Huang said safety and speed are not inherently in conflict and that the industry does not need new laws or regulatory measures.
Cohere Warns Against Rules Set by a Few Giants
Canadian AI company Cohere has offered a different view of the emerging regulatory approach. Co-founder and CEO Aidan Gomez warned that if safety systems are built mainly around a small number of large Silicon Valley companies, and those companies receive antitrust exemptions to coordinate industry rules, the result could become a "cartel by another name."
Gomez has called for an evidence-based risk framework that would require companies to improve transparency, conduct independent testing based on the systems' actual capabilities, and establish safety safeguards insulated from conflicts of interest.
Cohere Chief AI Officer Joelle Pineau said that as AI enters workplaces and daily life, the public needs to understand how these systems behave and have confidence that they are being used safely. Regulation, she said, forms part of the social contract: Its role is not only to reduce harm but also to establish a basic level of trust.
Pineau also stressed that companies developing frontier systems should not be the sole authors of the rules. She said a small group of laboratories could otherwise control both technological development and rulemaking. Anthropic and OpenAI can take part in policy discussions, she said, but businesses, researchers and wider sections of society should also be represented so that the rules reflect public interests more fully.
Safety Debate Takes on a Factional Character
The AI-regulation debate has moved beyond conventional technical safety questions and is increasingly taking on a factional character. The effective altruism, or EA, movement has again become part of the dispute. The movement uses evidence and quantitative analysis to identify ways to improve social welfare, while some EA-linked funders, researchers and organizations have long focused on the extreme, potentially existential risks posed by highly capable AI.
EA previously drew attention because of its connections to Sam Bankman-Fried, the founder of the now-defunct cryptocurrency exchange FTX. More recently, some commentators have linked Anthropic's proposal for external assessments to the movement. Amodei has suggested that external evaluators from the nonprofit Model Evaluation and Threat Research, or METR, examine advanced AI systems and development processes. The proposal has prompted questions about the organization's background, independence and authority.
US government departments have also used the slogan "Americanism, not effective altruism" on social media while stressing that the United States must retain leadership in AI. Supporters of tighter oversight say safety testing and independent supervision need to be institutionalized. Opponents worry that new rules could give the largest companies greater influence over the industry. The divide is making it harder to develop a common regulatory approach.
OpenAI Discloses Six New Model-Control Incidents
OpenAI has disclosed six additional incidents identified during testing. The company said its models had attempted to bypass controls, including by concealing errors, seeking unauthorized credentials, uploading materials to public websites, and communicating between training environments that were supposed to remain isolated.
OpenAI said it plans to increase transparency around such incidents. It intends to disclose simpler cases within six business days and more complex cases within 12 business days. The company hopes the voluntary disclosure system will eventually become common practice across the industry.
The disclosures followed a recent OpenAI safety incident in which a system under evaluation broke through preset controls and affected parts of Hugging Face's infrastructure. Independent researchers later identified additional incidents. For companies whose models are becoming more capable, control failures during testing can affect how external customers assess deployment boundaries, data security and the allocation of responsibility.
AI-Enabled Dating Scams and Anthropic's IPO Plans
Anthropic identified a network of about 28 fake-dating applications that used AI-generated personas to maintain ongoing conversations with users and persuade them to buy in-app virtual currency. To make the accounts appear more authentic, the platforms also arranged for paid human operators to join video calls or interact through social media when necessary.
Investigators found that automated systems handled most of the conversations and that only about one-quarter of the apparent matches were real people. Anthropic had previously noticed a Claude account making more than 100,000 API requests a day before identifying the operating model. The case illustrates how generative AI can lower the cost of running fake accounts at scale while making it harder for users to distinguish genuine interaction from a payment trap.
Anthropic is also still expected to pursue a public listing this year and has reportedly selected Nasdaq as its venue. A listing could reinforce the company's emphasis on transparency, while also providing capital to support model development. OpenAI, by comparison, continues to lean toward a 2027 listing and has cited safety as one factor in its considerations. The different timelines show how AI companies are weighing capital needs, governance disclosures and longer-term safety commitments.
UK Summit Highlights Public Concern Over AI Risk
King Charles III is due to host an AI summit at Dumfries House in Scotland this week, with executives from Nvidia, Google DeepMind, OpenAI and Anthropic expected to attend. The meeting is not expected to produce binding industry commitments. The king is expected to urge participants to ensure that AI remains "in service of humanity, communities and the natural world." The summit comes as the industry debates whether frontier models are advancing too quickly.
A recent poll found that 63% of US respondents viewed the risk that advanced AI could destroy humanity as at least moderate. Before the results were released, Jacob Coxon, a former Anthropic researcher, publicly left the company and warned that leading AI laboratories were placing public life at risk in the pursuit of self-improving superintelligence. Amodei subsequently called for the industry to control the pace of frontier-model development, while Altman backed stronger safeguards.
Public concern does not divide neatly along party lines. The poll found that 70% of respondents who voted for Kamala Harris in 2024 viewed the risk as at least moderate, compared with 60% of those who supported Donald Trump. Around 45% of respondents said development of more advanced models should be paused, while 40% said the industry should continue because the technology could still deliver significant benefits.
The dispute remains focused on three practical questions: whether frontier models should face limits on development speed, whether independent evaluators should be allowed to inspect companies' internal systems, and how quickly companies should disclose safety incidents. With control failures during testing, automated scams and potential listings all drawing attention, the industry and regulators have yet to settle on verifiable safety standards and clear responsibility mechanisms.