National Security Law Today · National Security Law Today

Innovation Without Guardrails: The Rise of OpenAI and the Upcoming National Security Law Conference

November 5, 2025·53 min·4 clips

OpenAI expanded pornographic material in DALL-E 3 because removing it would have failed data quantity requirements and degraded the model's ability to generate women's faces.

1. National Security Law Today re-airs a conversation between host Elisa Poteet and Karen Howe, author of Empire of AI, ahead of the 35th Annual Review of the Field of National Security Law CLE Conference. 2. Karen Howe is a bestselling author and award-winning AI reporter who trained at MIT with an engineering degree; she has trained over 2,000 journalists worldwide on AI accountability reporting through the Pulitzer Center's AI Spotlight Series. 3. The episode explores how OpenAI was founded, how its internal culture diverged from its public identity, and what the global consequences of unchecked AI development may be. 4. Howe began covering OpenAI at MIT Technology Review in 2019, initially fascinated by the company's plan to nest a for-profit within the nonprofit as a model for aligning capital incentives with public-interest mission. 5. After interviewing roughly four dozen people for a profile, she discovered a fundamental disconnect between OpenAI's public identity — antithetical to Silicon Valley values — and its actual internal culture, which she describes as 'part and parcel of the reckless, competitive, secretive culture ethos of Silicon Valley innovation.' 6. Howe notes that OpenAI's non-disparagement agreements included a clause allowing the company to claw back employee equity if the agreement was broken or even revealed — which she describes as a 'dramatic, aggressive tactic' that distinguished OpenAI from standard Silicon Valley NDAs. 7. She describes Sam Altman as 'the Michael Jordan of listening' in face-to-face meetings, yet documents a consistent pattern in which he tells different people different things at different times, and longtime collaborators cannot articulate what he truly believes or values. 8. Howe argues the nonprofit-to-for-profit transition was not a corruption of an idealistic original mission but a logical consequence of its founding premise: Elon Musk and Altman created the nonprofit specifically to out-compete Google and build AI 'in their image,' which made capital accumulation structurally necessary. 9. The OpenAI origin story included the framing of an 'AI Manhattan project' — Sam Altman's pitch to Musk — and this language was used in new employee onboarding for years, creating a mythology that justified why a small group needed to control a civilization-scale technology. 10. Howe explains that the 'good AGI before the bad guys build the bad one' narrative mirrored the nuclear-weapon analogy and served both a sincere motivating belief for some employees and a rhetorical mechanism for suppressing outside democratic participation. 11. On training data, Howe describes how OpenAI moved from 8 million curated articles in early GPT models to petabytes of unfiltered Common Crawl data, scraping every accessible web page — including transcribing YouTube videos and pirating books — to meet data quantity demands. 12. She reveals that DALL-E 3 was trained on dramatically expanded pornographic material compared to DALL-E 2, because at the required data scale, removing it would have failed quantity requirements and also degraded the model's ability to generate faces of women. 13. Howe argues these decisions illustrate the core governance failure: a small group philosophizing based on their own values, without public input or domain expertise, made choices whose consequences — including enabling synthetic child pornography — a child safety expert could have predicted immediately. 14. On trust and safety, Howe describes an early AI safety team focused entirely on theoretical long-term consciousness risks — not real-world harms — while ChatGPT was scaling to millions of users, leaving real-world safety retroactively addressed by a team that was 'completely insufficient.' 15. She draws a parallel to the subprime mortgage crisis: superficial safety mechanisms stapled onto a system built on rotten foundations, where the metastasis into the global system made remediation nearly cataclysmic. 16. On energy, Howe cites a McKinsey report projecting AI data centers would consume the equivalent of 2 states of California in the conservative scenario and 6 states of California in the accelerated scenario within five years. 17. She notes that the US and EU had flatlined or declining energy demand for a decade before AI data center expansion; AI is solely responsible for the current historic rise in global energy consumption. 18. The conversation is structured as an extended interview, with Poteet providing policy and historical framing and Howe providing insider investigative reporting — a reflective rather than adversarial format. 19. Lawyers, journalists, policy professionals, and technology ethicists tracking AI governance, OpenAI's corporate structure, or the national security implications of unregulated AI development will find this substantive. 20. Listeners seeking current product updates, technical AI tutorials, or neutral coverage of OpenAI will find the episode sharply critical and investigative in framing.

As heard by us

A sober account of OpenAI's rise as a national security law story.

National Security Law Today uses this re-airing to frame Karen Howe's Empire of AI as a legal and institutional story, not merely a company story.

Read the full review in PlayNext →

Why you'd press play

If OpenAI's rise feels bigger than tech, this gives it legal shape.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from National Security Law Today