GPT-5 and OpenAI’s strategy, Trump’s AI Action Plan, and the White House negotiates a 15% cut of Nvidia and AMD’s China chip sales

Subscribe to policy.ai
Worth Knowing

OpenAI Releases GPT-5: OpenAI debuted GPT-5 last week, the AI lab’s most anticipated release to date. With approximately 700 million weekly active users of ChatGPT, OpenAI’s models are the most widely scrutinized and the industry standard against which all other models are compared. Early results for GPT-5 show a mixed bag depending on how you look at it. What the company describes as “GPT-5” is not a single model but a “unified system” of models suited to different purposes. The most sophisticated of those models — GPT-5 with high levels of reasoning — took the top spot on the popular LMArena leaderboard across all categories and topped the ArtificialAnalysis composite leaderboard, which aggregates scores across popular benchmarks like GPQA Diamond, Humanity’s Last Exam, and LiveCodeBench.

But compared to what felt like a significant generational leap between GPT-3 and GPT-4 back in 2023, the difference between GPT-5 and the models that came before feels more like a small step. For one, while it beat out competitor models from the likes of xAI, Google, and Anthropic in the aggregate, it only does so by a few percentage points in most cases and even trails slightly on a handful of benchmarks. And even compared to OpenAI’s o3 and GPT-4.5 models, GPT-5 only shows modest improvement. That small step up may be disappointing if you were hoping for exponential takeoff or predicting a near-term AI takeover, but the other features of the GPT-5 release mean its impact will be felt more widely than a few more points on most benchmarks. On the day it was announced, GPT-5 rolled out to all ChatGPT users, including those using the tool for free. OpenAI’s model “router,” which selects the most appropriate GPT-5 model to answer a question, means that all of the company’s 700 million users will get much more exposure to reasoning models like OpenAI’s earlier o1 and o3 models. That will likely seem like a massive improvement for most users — according to OpenAI CEO Sam Altman, only a small percentage of ChatGPT paid users had used reasoning models on any given day.

Combined with other purported improvements focused on common use cases — lower hallucination rates, better health-related answers, less sycophancy — it’s clear OpenAI is more focused on quality-of-life improvements than pure benchmark victories. That’s a strategy that has worked before: ChatGPT’s instant popularity when it launched in 2022 wasn’t due to its underlying model, which had launched earlier that year with minimal fanfare (regrettably, we were more focused on Google’s comparatively massive PaLM language model at the time). It was because of its ease of use and consumer friendliness. With GPT-5, OpenAI seems to be running a similar playbook.
Concerns Grow about Mental Health Impact of AI Usage: As more stories crop up about AI users falling into “delusional spirals” — some resulting in hospitalization and even suicide — concerns are growing about the mental health effects of AI model use. With hundreds of millions of regular users, it’s still not clear whether the stories are a natural byproduct of mass adoption or the result of something inherent to the way the models work. But as CSET’s Helen Toner told the New York Times, the way models are trained means conversations can quickly spiral. Because models function like “improv machines,” their propensity to “yes, and…” — agree with the user and expand on the previous point — means that if the user is edging toward dangerous or harmful thinking, the models may not just have insufficient guardrails but can actively exacerbate the issue. OpenAI announced it would start to prompt users to take breaks during long chat conversations, and its GPT-5 release specifically highlighted the goal of “minimizing sycophancy,” but it may be difficult for companies to avoid the problem entirely. One of the biggest issues, as OpenAI realized after it deprecated its older GPT-4o model following the release of GPT-5, is that many users prefer a more flattering model. User outcry over the decision was significant enough that the company quickly resurrected GPT-4o, which can now be accessed again through ChatGPT. We’ll have to wait for hard data to come in before we know for sure whether “AI Psychosis” is a real phenomenon or more like the other panics that have accompanied new technologies.
Government Updates

White House Unveils Sweeping AI Action Plan and Accompanying EOs: On July 23, the White House released its 23-page AI Action Plan, outlining the administration’s strategy for maintaining U.S. leadership in artificial intelligence while operationalizing key proposals through three accompanying executive orders. The plan — authored by OSTP Director Michael Kratsios, Special Advisor David Sacks, and Secretary of State Marco Rubio — was put together at the direction of President Trump’s January executive order on AI and incorporates feedback from over 10,000 respondents to a federal request for information issued earlier this year. While the plan points to long-term goals — “winning the AI race” — its focus is decidedly on near-term, actionable policies. We recapped the plan in greater depth on CSET’s website, but at a high level, the plan is organized around three pillars:
  • “Accelerate AI Innovation” through deregulation and strategic investment, including directing OMB to identify and repeal restrictive regulations, a proposal to withhold federal funds from states with burdensome AI laws, a plan to establish regulatory sandboxes and “AI Centers of Excellence” for testing, and a DARPA-led program for breakthroughs in AI interpretability and control. The plan backs open-source models “founded on American values” and calls for giving all federal employees access to frontier language models (see the following entry on the federal government’s deal with OpenAI for more). The section also outlines the administration’s plan to use federal contracting to combat perceived ideological bias in AI models. An executive order signed by President Trump the same day — Preventing Woke AI in the Federal Government — operationalized the plan.
  • “Build American AI Infrastructure” through streamlined permitting and energy expansion, including categorical exclusions for data center construction, making federal lands available for AI facilities, prioritizing new energy generation from geothermal and nuclear sources (notably excluding other renewables), and “revamping” the CHIPS Program Office to remove “extraneous policy requirements” like childcare provisions. The accompanying executive order on data center permitting operationalizes many of these recommendations.
  • “Lead in International AI Diplomacy and Security” through export promotion and enhanced controls, including creating an “American AI Exports Program” to facilitate full-stack technology packages for allies (as elaborated by the third and final same-day executive order), countering Chinese influence in international standard-setting bodies, exploring location verification features on advanced chips to prevent diversion, plugging loopholes in semiconductor manufacturing export controls, and using tools like the Foreign Direct Product Rule and secondary tariffs to ensure allied alignment.
The plan’s accompanying executive orders made some elements immediately concrete — and DOE announced four federal sites for AI data centers within days — but other provisions will require sustained political support. Initial reception was largely positive, but civil rights organizations and congressional Democrats sharply criticized the “Preventing Woke AI” order, and the proposal to withhold funds from states with AI regulations risked reigniting a fight that nearly scuttled the president’s Big Beautiful Bill earlier this summer. As implementation unfolds, those contentious provisions may get bogged down in politics and litigation, but the infrastructure proposals, export plans, and internal executive branch initiatives are likely to proceed, with prioritization and bandwidth being the biggest impediments.

OpenAI and Anthropic Strike $1 Deals to Provide AI Tools to the USG: Last week, OpenAI and the General Services Administration announced a first-of-its-kind agreement to provide ChatGPT Enterprise to the entire executive branch for just $1 per agency for one year. Not to be outdone, Anthropic quickly matched the offer with its Claude models, extending access to all three branches of government (not just the executive branch) at the same price. Under the OpenAI deal, millions of federal workers across all executive branch agencies will be able to use ChatGPT for their work, with unlimited access to premium capabilities like Deep Research during the first 60 days. The deals leverage GSA’s Multiple Award Schedule, allowing agencies to bypass lengthy procurement processes and onboard immediately. Both partnerships align with the Trump administration’s AI Action Plan goal to “accelerate AI adoption in government.” That process was already underway, with the Pentagon recently inking $200 million contracts with multiple AI providers and GSA approving OpenAI, Anthropic, and Google as federal vendors, but the practically free trials will likely speed up adoption significantly. For the AI companies, the calculus is a classic enterprise software move: prove your value during the trial period and lock in a paying customer for life. Industry observers dubbed it a “land and expand” strategy — OpenAI and Anthropic are betting that after a year of integration into workflows, agencies will find the tools indispensable enough to pay standard rates. But several questions remain about the partnerships’ long-term viability, including how much agencies will ultimately pay when promotional pricing expires and whether the administration’s executive order banning “woke AI” from government use will create friction as the tools are deployed at scale. 

Nvidia and AMD Agree to Give the U.S. a 15% Cut of Chip Sales to China: This week, President Trump announced a deal with U.S. chip design firms Nvidia and AMD that will net the U.S. government a cut of the companies’ sales to China of specific high-powered chips. Both companies will pay the U.S. government 15% of revenues from the sales of their AI accelerators developed for the Chinese market — Nvidia’s H20 and AMD’s MI308 — in order to obtain export control licenses. The move is the latest development in what has been a roller-coaster of an approach to Chinese chip sales. The administration had blocked exports of Nvidia’s H20 and equally powerful chips in April, only to reverse course in July. That reversal was met with criticism, including from prominent Republicans; House Select Committee on China Chair John Moolenaar (R-MI) was particularly outspoken, publicizing his opposition in a letter to Commerce Secretary Howard Lutnick. Neither Moolenaar nor his Democratic counterpart on the China Committee, Ranking Member Raja Krishnamoorthi (D-Ill.), seems pleased by the latest deal with Nvidia and AMD. “Export controls are a frontline defense in protecting our national security, and we should not set a precedent that incentivizes the Government to grant licenses to sell China technology that will enhance its AI capabilities,” Moolenaar told Politico. Both he and Krishnamoorthi also raised questions about the legality of the deal. While the path is seemingly clear for both Nvidia and AMD to begin exporting, there are some signs that demand may not be as robust as expected: Chinese authorities have reportedly advised against using Nvidia’s H20 in recent weeks, particularly for government or national security-related work, possibly due to concerns about potential “back-door” risks of using those chips. But the paucity of comparably powerful China-made alternatives likely means that Beijing’s advice will do little to sap demand.

Trump Hints at 100% Chip Tariffs, But Expect Major Exceptions: Last week, President Trump said he plans to place a 100% tariff on semiconductors produced abroad, though he said that companies with U.S. manufacturing commitments would be exempt. And while exemptions for companies that manufacture at least some of their semiconductors in the United States would spare foreign industry leaders like TSMC, Samsung, and SK Hynix, it could have a significant impact on the market for legacy chips, which are largely produced abroad and do not command the margins that would typically incentivize significant domestic investment. For now, it’s not clear how or when the tariffs would be imposed — a White House official told Politico that more details would be forthcoming as part of an official announcement and that “the tariffs will be nuanced and phased in to reshore manufacturing while minimizing supply chain disruptions.” President Trump has said that announcement could come as soon as this week, but as of writing, nothing formal has materialized.

Job Openings

We’re hiring! Please apply or share the role below with candidates in your network:
  • Data Research Analyst: This role bridges CSET’s data and analysis teams, combining technical skills with subject-area expertise. Ideal candidates have experience in data visualization, programming, and statistical analysis. Apply by September 2, 2020.
Please bookmark our careers page to stay up to date on all active job postings. You can also subscribe to receive job announcements by updating your subscription preferences in the footer of this email.

In Translation
CSET’s translations of significant foreign language documents on AI
If you have a foreign-language document related to security and emerging technologies that you’d like translated into English, CSET may be able to help! Click here for details.

What’s New at CSET

REPORTS

PUBLICATIONS AND PODCASTS

EMERGING TECHNOLOGY OBSERVATORY

IN THE NEWS

What We’re Reading

Paper: Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens, Chengshuai Zhao, Zhen Tan, Pingchuan Ma, Dawei Li, Bohan Jiang, Yancheng Wang, Yingzhen Yang, and Huan Liu (August 2025)

Paper: Subliminal Learning: Language models transmit behavioral traits via hidden signals in data, Alex Cloud, Minh Le, James Chua, Jan Betley, Anna Sztyber-Betley, Jacob Hilton, Samuel Marks, and Owain Evans (July 2025)

Paper: Scaling Laws Are Unreliable for Downstream Tasks: A Reality Check, Nicholas Lourie, Michael Y. Hu, and Kyunghyun Cho (July 2025)