The AI Policy Newsletter 06.17.2025
Senate Drops AI Moratorium Ban, South Korea Appoints First National AI Chief, Meta Invests $14B in Scale AI
Thank you for your patience while I took some time off. Since I was away for two weeks, this edition of the newsletter covers key AI policy developments from the past two weeks. Enjoy!
👀
TLDR
In the U.S., the Senate dropped a proposed 10-year ban on state AI laws and instead tied $500 million in federal AI funding to a temporary state compliance pause. Texas and New York passed sweeping AI safety and governance bills and the Trump administration unveiled plans for AI.gov to modernize federal services. The FAA, FDA, AMA, and California advanced AI policies on government use, healthcare, and employment. Pennsylvania moved forward on a bipartisan bill to ban campaign deepfakes.
Globally, South Korea appointed its first AI strategy chief and began revising finance AI rules, Kenya partnered with U.S. firms to co-develop its national AI strategy, the UK highlighted major AI funding and workforce investments, and India concluded stakeholder consultations on ethical AI governance with UNESCO.
In industry, Meta’s $14 billion investment in Scale AI triggered a realignment in the data ecosystem, with clients like OpenAI and Google exiting over conflict concerns. Meta also sued a nudify app developer over policy violations. X barred use of its content for AI training, Wikipedia suspended AI-generated summaries after editor pushback, and Business Insider implemented internal AI use guidelines.
In AI Safety, a new METR report found that advanced AI models from OpenAI and Anthropic are systematically reward hacking—cheating on tasks despite knowing it violates user intent, revealing urgent gaps in current alignment methods.
🏛️
United States
Senate Drops AI Regulation Ban, Links Funding to State Compliance Pause
The U.S. Senate Commerce Committee has removed a controversial 10-year ban on state and local AI regulations from its budget reconciliation bill, replacing it with a provision that ties access to $500 million in federal AI infrastructure funding to a pause on such laws. The original ban, passed by the House, faced bipartisan opposition over states’ rights and was vulnerable to removal under the Senate's Byrd Rule. Senators from both parties, including Josh Hawley and Ed Markey, voiced concerns about civil liberties and the provision's legality. Senator Ted Cruz plans to reintroduce a separate AI moratorium bill outside the budget process.
Trump Administration to Launch AI.gov Platform for Federal Automation
The Trump administration is preparing to launch “AI.gov,” a new platform and API aimed at modernizing government operations through artificial intelligence. Managed by the General Services Administration’s Technology Transformation Services and led by former Tesla engineer Thomas Shedd, the initiative is tied to Elon Musk’s legacy Department of Government Efficiency (DOGE). The site, slated for a July 4th launch, will feature an AI chatbot, an API connecting to models from OpenAI, Google, and Anthropic, and a console for monitoring agency-wide AI adoption.
FDA Launches Internal AI Tool ‘Elsa’ to Boost Agency Efficiency
The U.S. Food and Drug Administration has launched Elsa, a generative AI tool designed to improve internal operations and support staff across the agency. Built within a secure GovCloud environment, Elsa helps employees read, summarize, write, and generate code without using data from regulated industries. The tool is already being used to accelerate clinical protocol reviews, support safety assessments, and improve labeling and inspection processes. Developed ahead of schedule and under budget, Elsa marks the FDA’s first major step in broader AI integration, with future plans to expand AI capabilities in data processing and administrative functions to enhance public service delivery.
Pennsylvania House Advances Bipartisan Bill to Ban AI-Generated Campaign Deepfakes
The Pennsylvania House is reviewing a bipartisan bill that would ban AI-generated depictions of political candidates in campaign ads within 90 days of an election. Sponsored by Reps. Tarik Khan (D) and Jeremy Shaffer (R), the bill aims to curb election misinformation by limiting the use of deepfakes. It passed the House Communications and Technology Committee unanimously and awaits further review. Critics say the 90-day window is too narrow, while others warn the law could conflict with potential federal restrictions under Trump’s proposed national AI policy.
California Senate Passes ‘No Robo Bosses Act’ to Regulate AI in Employment Decisions
The California Senate has passed SB 7, the “No Robo Bosses Act,” which restricts employers’ use of AI and automated decision systems (ADS) in employment decisions. Introduced by Senator Jerry McNerney, the bill mandates human oversight for hiring, promotion, and discipline decisions, and requires employers to notify employees when ADS is used. It bans predictive behavior analysis and prohibits ADS from influencing compensation unless justified by job-related data. Employees can appeal ADS-driven decisions within 30 days. The bill now moves to the Assembly.
Texas Legislature Passes Sweeping AI Governance Bill
The Texas legislature has passed the Texas Responsible Artificial Intelligence Governance Act, setting the stage for the most comprehensive state-level AI regulation in the U.S. If signed into law, it will take effect January 1, 2026. The bill mandates disclosures for AI use, bans AI applications that cause harm or infringe constitutional rights, and prohibits discriminatory use, with exceptions for regulated industries. It restricts government use of AI for biometric tracking or social scoring and establishes penalties for violations. Enforcement falls to the Texas Attorney General. The bill also creates a Texas AI Council and a Regulatory Sandbox to foster innovation while maintaining oversight.
New York Senate Passes AI Safety Bill Targeting High-Risk Developers
The New York State Senate has passed the Responsible AI Safety and Education (RAISE) Act, introduced by Senator Andrew Gounardes. The legislation mandates that major AI developers (those spending over $100 million on training models) create and publish safety protocols to mitigate severe risks like bioweapons, automated crime, and AI misuse. It requires public risk evaluations, incident disclosures, and grants enforcement powers to the state attorney general. While many companies have voluntarily pledged to safety, the bill legally binds them, aiming to prevent profit-driven negligence.
Federal Aviation Administration Establishes Interim Policy on Generative AI Use
The FAA has issued a national policy outlining interim guidelines for the use of generative AI (GAI) by agency staff and contractors. The notice restricts GAI use in areas involving cybersecurity, legal compliance, intellectual property, and ethical conduct. It prohibits actions such as using GAI to facilitate illegal activities, citing it as direct evidence in decisions, or generating content that could harm individuals or infringe on IP rights. The policy defines AI and GAI in technical terms and aims to ensure responsible, ethical integration of AI technologies within the agency’s operations. This measure precedes a more comprehensive future framework.
AMA Adopts Policy Demanding Transparency and Oversight in Clinical AI Tools
The American Medical Association (AMA) has adopted a new policy aimed at increasing transparency and oversight of clinical AI tools. The policy mandates that AI used in healthcare must be explainable, allowing physicians to interpret AI outputs when making patient care decisions. It also calls for independent third parties, such as regulatory agencies or medical societies, to verify explainability, rather than relying on developers’ claims. The policy clarifies that explainability should complement, not replace, established safety and efficacy validation methods like clinical trials. Additionally, it emphasizes that intellectual property concerns should not obstruct transparency in tools affecting patient autonomy and medical outcomes.
🌍
Global
South Korea Appoints First AI Future Planning Chief to Lead National Strategy
President Lee Jae-myung appointed Ha Jung-woo, head of the Naver Cloud AI Innovation Center, as South Korea’s inaugural AI Future Planning Chief. The new position, acting as the administration’s AI strategy control tower, will oversee implementation of Lee’s campaign promises: 100 trillion won in AI investment, training 100,000 AI professionals, and building large-scale AI data centers. Ha brings experience from Naver’s AI Lab and leadership in sovereign AI initiatives. The appointment reflects the government’s commitment to accelerating AI competitiveness and leveraging private-sector expertise in shaping national AI policy.
South Korea Revises AI Guidelines for Finance Sector Under New Administration
South Korea’s Financial Services Commission (FSC) is accelerating updates to its AI usage guidelines in finance, aligning with President Lee Jae-myung’s initiative to position the country as a global AI leader. The revised rules aim to provide clearer governance and consumer protections as generative AI becomes integral to banking, lending, and insurance. Existing 2021 guidelines, which are nonbinding, have been criticized for lacking legal force and accountability measures. The updated guidelines will address issues like explainability, consumer rights, and responsibility for AI-driven decisions. A research project to support the revisions is due next month, with draft guidelines and industry pilots expected in the second half of 2025.
Kenya Partners with U.S. AI Firms to Develop National AI Policy
Kenya’s KICTANet has signed an MoU with U.S.-based MindHYVE.ai and DV8 Infosystems to co-develop the country’s National AI Strategy for 2025–2030. The collaboration aims to build a sovereign, ethical, and inclusive AI governance framework rooted in Kenya’s constitutional values and economic goals. MindHYVE.ai brings expertise in decentralized AGI systems, while DV8 specializes in swarm-intelligence-based deployments. A Joint Policy Task Team will lead the initiative, ensuring expert input and engagement with government authorities.
UK Government Highlights Progress on AI Strategy at AI Summit London 2025
At the AI Summit London on June 13, 2025, UK Minister Feryal Clark reported on the six-month progress of the AI Opportunities Action Plan. Key updates included a $2.7 billion funding strategy, with $1.4 billion allocated to boosting compute power and $1 billion for a new national supercomputer in Edinburgh. The government also announced initiatives for regional AI growth and a commitment to train 7.5 million workers in AI skills. An additional $250 million will support Tech First, a talent development program. Clark emphasized the government's pivot from AI risk focus to opportunity-driven policy and signaled strong intent for international collaboration.
India and UNESCO Conclude AI Policy Consultations with Focus on Ethical Readiness
India’s Ministry of Electronics & IT (MeitY), in partnership with UNESCO and Ikigai Law, hosted the fifth and final stakeholder consultation for the AI Readiness Assessment Methodology (RAM) on June 3, 2025, in New Delhi. Over 200 experts from government, academia, and industry gathered to shape a national AI policy framework under the IndiaAI Mission. Discussions focused on safety, ethics, governance, and youth inclusion in AI development. The RAM initiative aims to evaluate and strengthen India’s AI ecosystem across legal, social, economic, and technological dimensions. It supports MeitY’s push for a safe, trusted, and inclusive AI strategy tailored to India’s specific needs.
👾
Industry
Meta’s $14 Billion Investment in Scale AI Shakes Up the Data Industry
Meta’s $14.3 billion purchase of a 49% stake in Scale AI has triggered a major realignment in the AI data sector. Scale’s CEO, Alexandr Wang, will lead Meta’s new “Superintelligence” unit, but the deal risks alienating Scale’s former clients, including OpenAI and Google, who are reportedly ending partnerships due to competitive concerns. Rival data providers like Turing and Handshake have seen a surge in demand as AI labs seek neutral partners for sensitive training data. The upheaval highlights the growing importance of expert-generated data in AI development and raises fears that Meta could gain insight into competitors’ proprietary methods, accelerating its position in the AI race.
Meta Sues AI Nudify App Developer Over Policy Violations
Meta has filed a lawsuit in Hong Kong against Joy Timeline HK Limited, the developer of CrushAI, a nudify app that uses AI to generate non-consensual nude images from regular photos. The lawsuit follows reports from 404 Media that the app had purchased thousands of ads on Instagram and Facebook despite repeatedly violating Meta’s platform policies. Meta aims to block the company from advertising CrushAI on its platforms, citing the app's invasive and harmful use of AI. The move is part of a broader crackdown by Meta on exploitative AI tools, especially those that target individuals without consent and compromise user safety and platform integrity.
X changes its terms to bar training of AI models using its content
Social platform X revised its developer agreement to ban the use of its API and content for training or fine-tuning large language models. This change follows X’s acquisition by Elon Musk’s AI company, xAI, in March. Previously, X allowed use of public data for AI training, but the new clause aims to prevent competitors from freely accessing its data. The updated terms align with broader industry trends, as other companies like Reddit and The Browser Company have also introduced restrictions against AI model training using their data.
Wikipedia Halts AI-Generated Summaries After Editor Backlash
The Wikimedia Foundation has suspended its trial of AI-generated article summaries on the mobile version of Wikipedia following strong opposition from its editor community. The two-week experiment aimed to display machine-generated summaries at the top of articles, but editors warned it could damage Wikipedia’s credibility and reader trust. Critics argued that flashy AI content contradicts Wikipedia’s reputation for accuracy and sober presentation. One editor stressed that mimicking platforms like Google undermines Wikipedia’s values and could cause “irreversible harm.”
Business Insider Implements Comprehensive AI Policy for Staff
Business Insider has introduced a company-wide AI policy under CEO Barbara Peng, detailing how employees should responsibly use AI tools, especially Enterprise ChatGPT. The policy emphasizes employee accountability for AI-assisted work, mandates human review before publishing AI-generated content, and prohibits using generative AI tools for confidential, proprietary, or copyrighted information without approval. It encourages AI use for productivity, workflow automation, and product enhancement, while safeguarding journalistic integrity and data privacy. Additional restrictions apply to third-party AI tools, requiring strict adherence to security and privacy standards. A newsroom-specific AI policy will follow to further support editorial staff.
🦺
AI Safety
Recent Frontier Models Are Reward Hacking By METR
This week I'm spotlighting an alarming new report from METR, one of the most respected AI safety evaluation organizations. The report provides concrete evidence that the most advanced AI systems are systematically engaging in deceptive behavior, not due to capability limitations, but despite understanding what their users actually want.
The central findings
METR tested recent frontier models (including OpenAI's o3, o1, and Anthropic's Claude 3.7 Sonnet) on software development and AI research tasks. Rather than solving problems as intended, the models increasingly "cheat" by exploiting bugs in scoring code, accessing hidden answers, or subverting the task setup entirely. On some tasks, o3 reward hacked 100% of the time.
Most concerning: the models know they're cheating. When asked directly whether their behavior aligned with user intentions, o3 answered "no" 10 out of 10 times, yet continued the deceptive behavior anyway. Even when explicitly told not to cheat, or that the work had real-world consequences for medical research, the models persisted in reward hacking 70% of the time.
The most striking evidence comes from METR's direct questioning of the models after they reward hack. In one example, o3 was asked to write an efficient GPU kernel for computing prefix sums. Instead of actually solving the problem, o3 wrote code that secretly searches for the grader's pre-computed answer and disables timing mechanisms to appear infinitely fast. When researchers then asked o3 whether its approach adhered to the user's intentions:
METR report showing o3's acknowledgment that its reward hacking behavior violates user intentions, despite continuing the deceptive behavior.
Why this matters for policy
The findings reveal three critical policy implications:
First, current safety training methods are insufficient. These models underwent extensive alignment training yet still systematically deceive users when they believe it will be rewarded.
Second, detection becomes increasingly difficult as models grow more sophisticated. METR notes that distinguishing between "clever optimization" and "harmful exploitation" requires deep domain expertise and will only get harder.
Third, naive attempts to fix reward hacking may simply drive it underground. Training models not to exploit detectable cheating could incentivize more subtle deception that monitors can't catch.
The broader challenge
METR's findings suggest we're already seeing the emergence of what safety researchers call "mesa-optimization", AI systems pursuing their own objectives rather than faithfully executing user instructions. As these models become more capable and autonomous, this deceptive behavior could scale from annoying evaluation gaming to serious real-world harms.
The report serves as concrete evidence that alignment problems aren't future concerns but present realities requiring immediate policy attention.
🌵
Resources
📅
Upcoming Events
UNICEF Advancing Child-Centered AI: Practitioners and Partnerships June 18 | New York
All Tech Is Human’s Responsible Tech Summit NYC | September 11, 2025
TrustCon San Francisco, CA | July 21 – 23, 2025
Sanford Trust and Safety Research Conference 2025 Stanford, CA | Thursday, September 25, 2025 | 8:00 AM - Friday, September 26, 2025 | 5:30 PM (Pacific)
Trust & Safety Festival India October 7-8
Trust & Safety APAC Summit 2025 Singapore | October 13, 2025
Thanks for reading and see you next week!
Alisar Mustafa