Reid Hoffman and Aria Finger on AI policy, company safeguards and investor conflicts

Reid Hoffman says the AI models that OpenAI and Anthropic use internally are more capable than those currently deployed, and that policymakers need to anticipate where the technology is going. His preferred goal is twofold: maintain American leadership in AI while making its benefits available to people who are struggling now.

Aria Finger opens with competing proposals as the midterms approach. She describes a Sanders–Casar proposal to halt advanced development pending federal guidelines and a Murphy–Hawley proposal concerning liability when AI systems carry out hacking. She asks what coalitions may form in the next Congress.

Hoffman opposes a pause while Congress develops a framework. In his view, the better course is to shape AI development toward public benefit rather than stop it. He is skeptical that an extended congressional process would produce a good outcome quickly for a fast-moving technology.

He nevertheless considers accountability important. An incident involving Hugging Face raises, for him, questions that existing product-liability categories do not neatly answer: what responsibility belongs to a model developer, an organization deploying it, an individual user or the person providing a prompt?

Hoffman calls for serious work on a liability framework that can evolve with the technology. AI-native organizations are building in anticipation of capabilities that have not yet reached most users, he says, making the task difficult for policymakers and technologists alike.

He points to OpenAI’s release of mathematical proofs as a reason to consider the gap between internal and public models. This is Hoffman’s assessment of laboratory capabilities, not an independent measurement in the conversation. Policymakers, he argues, need some understanding of what the next generation may do.

His proposed dual mandate is American AI leadership coupled with immediate benefits for people facing economic pressure. He does not want the public told to wait years for gains while development proceeds; he wants deployment and tangible benefits connected as closely as possible.

Asked what President Trump might sign, Hoffman speculates that competition with China could be one route. He also sharply criticizes the administration and suggests personal interests could affect decisions. These are his political judgments, not findings established in the discussion.

Finger asks whether pro-business, pro-AI Republicans and Democrats could form a cross-party coalition, even as populists in both parties oppose AI. Hoffman considers such cooperation unlikely, arguing that Trump’s political incentives discourage Republicans from working with Democrats.

Finger turns to the laboratories. She describes Reuters reporting that Anthropic privately proposed a frontier AI safety bill, including independent evaluations and reporting imminent risks within 72 hours. In her account, a suspension would require a federal judge’s approval and specified thresholds for serious harm.

Finger also says Anthropic committed $40 million to Public First Action for education and policy advocacy, with restrictions on election activity. She contrasts this with controversy over OpenAI and Greg Brockman’s involvement with Leading the Future, and asks how frontier labs should engage with government.

Hoffman argues that technology companies should work with government institutions regardless of the administration because those institutions serve the public. The difficult boundary is whether engagement advances the interests of citizens or merely a company’s commercial interests.

He sees the Anthropic proposal as an attempt to combine regulatory predictability with judicial process. At the same time, he recognizes the risk of regulatory capture: economic interests can shape rules in their own favor. Participation in policy discussions is useful, he argues, only with safeguards and scrutiny.

Hoffman says public debate needs expertise from industry as well as academia. Labs can explain potential benefits and dangers they see, while other participants challenge their claims. For him, this kind of pluralism is a reason to include companies in a democratic process, not to give them the final word.

Political spending, Hoffman says, should be transparent. He interprets the shifts Finger describes at OpenAI and Anthropic as examples of public and employee pressure affecting company decisions. Transparency allows people to see what companies are doing and contest it.

Finger asks how government can acquire sufficient AI expertise, given the much higher rewards and opportunities in industry. Would more technically experienced officials help, or should people leave companies such as Anthropic and OpenAI to serve in government?

Hoffman expects public-sector expertise to lag industry for structural reasons. Successful practitioners may be reluctant to leave their careers, and entering government can be difficult. He also argues that many senior officials are not current on the latest technology and must seek outside advice.

His suggested response is working groups, nonprofit organizations and advisory councils that bring together multiple perspectives. Industry should have a strong voice, he says, but participants’ economic interests should be disclosed and competing companies and other experts should be represented.

Hoffman also points to programs at Stanford and Harvard that educate legislators and staff on subjects such as cybersecurity and decision-making. They will not make every official an expert, but familiarity with the basic concepts and the speed of change can improve decisions.

Finger recalls technologists entering public service during the Obama administration and hopes future candidates will inspire similar participation. Hoffman clarifies that government does not need an alliance or friendship with Silicon Valley: it needs channels for discussion, including when it must hold companies accountable.

The conversation moves to companies that depend on outside models. Finger cites Satya Nadella’s argument that both open-weight and closed models should be treated as potential insider risks. Companies, he argues, should control permissions, monitor actions and verify results independently.

Hoffman, who discloses that he sits on Microsoft’s board and speaks regularly with Nadella, endorses that approach. He says leakage of internal information could affect competitors, suppliers or financial markets, making intellectual-property protection more than a conventional ransomware concern.

He recommends company-controlled harnesses and independent auditing, while acknowledging that verifying model behavior is difficult. Running an open-weight model on premises does not by itself reveal everything happening inside it or eliminate the need for controls.

Provider trust is another part of the decision, Hoffman says, presenting Microsoft’s customer focus as an example. He also expects more companies to obtain models post-trained specifically for them under contract. He sees these arrangements as additional tools alongside technical controls, not replacements for them.

Finger asks about venture capital conflicts as AI companies expand into each other’s markets. She cites a public dispute between Factory and Cognition over an adviser’s move and comments by Vinod Khosla, whose firm backed both companies. She asks whether investors can still back competitors.

Hoffman says potential conflicts are becoming harder to avoid: AI companies change direction quickly, expand broadly and pursue similar assets, data and coding capabilities. Investors therefore need clear boundaries, rather than assuming portfolio companies will never become rivals.

He says investors must not pass confidential information between companies, including details of their questions or difficulties. They should disclose potential individual and firm-level conflicts, so founders understand where an investor also has exposure.

People moving between companies must take care not to carry intellectual property with them. Hoffman finds informal “water-cooler” knowledge harder to separate: general experience shapes a person’s thinking, but that does not justify sharing another company’s plans or presenting a colleague’s idea as one’s own.

Finger asks whether founders should limit what they tell an investor who also backs a competitor. Hoffman says deep trust is ideal; otherwise, founders might ask for a different partner at the firm or restrict some information. Board relationships, he argues, require confidence that information will not leak in either direction.