Policy Memo
Topline
This bill package would hamper American AI development. While certain bills have interesting implications, Warner’s framework would be unlikely to advance American competitiveness in the global AI field.
Plymouth Institute for Free Enterprise
Artificial Intelligence (AI) is the 21st century’s industrial revolution and is already fundamentally altering American life. Politicians are offering expansive and prescriptive frameworks for the still-emerging AI industry. While it is important to create an appropriate legal and regulatory framework for AI, many of these policies are overbroad retreads of age-old agendas of state expansion. This report examines Senator Mark Warner’s proposal, “A Framework for America’s AI Future,” which encompasses several bills, including:
This bill package would hamper American AI development. While certain bills have interesting implications, Warner’s framework would be unlikely to advance American competitiveness in the global AI field.
The Data Center Tax Accountability and Disclosure Act (DCTADA) is a combination of environmental mandates masquerading as thoughtful data center regulations. It would deny full and immediate expensing for equipment and machinery to most data centers. Full and immediate expensing or “100% bonus depreciation” was enacted temporarily in 2017 and made permanent in 2025. It allows companies, when calculating their taxable income, to deduct legitimate costs associated with equipment and machinery in the same year that they are placed in service, instead of spreading such deductions out across a prescribed “useful life” of the equipment. This is the proper tax treatment and consistent with how virtually all other legitimate business expenses are treated. By denying full and immediate expensing specifically to data centers, the DCTADA would discourage data center investments and would threaten America’s edge over China in the data center and AI development race.
The DCTADA carves out an exception that would allow full and immediate expensing for data centers that have a Gold or Platinum LEED Green Building certification. This would empower the U.S. Green Building Council (USGBC) to act as the arbiter of which data center projects can qualify for equal tax treatment and would push more companies to jump through the bureaucratic hoops established by USGBC. The carveout for “green” data centers suggests that the push to curtail AI and data centers in this legislation is being guided not by national security concerns, but by a desire to pick winners and losers and establish federal control over the industry.
The legislation would also impose an onerous reporting and disclosure regime over data centers. The reporting and disclosure would cover water access and usage, electricity access and usage, backup power usage, greenhouse gas emissions, etc. Companies that fail to comply with the reporting and disclosure requirements would be subject to stiff monetary penalties. This would further stifle data center development by diverting companies’ focus away from building, operations, and maintenance to the preparation of reports for government bureaucrats.
This bill is unnecessary because the private sector is already taking voluntary actions to use resources more efficiently and invest in carbon-free energy. Hyperscalers are signing long-term contracts and making investments to expand, restart, or preserve nuclear power plants and other carbon-free power projects to meet their growing data-center demand. An increasing number of newer facilities are also conserving water by adopting closed-loop cooling systems that reuse the same water within the data center. These voluntary initiatives avoid the unnecessary compliance costs associated with a government-directed “green energy” agenda that makes digital services and AI more expensive.
The AI AGENT Act requires large online platforms (defined as an online service with more than 50 million monthly customers or subscribers in the U.S.) to create pathways for AI agents to manage users’ online interactions, electronic commerce decisions, user-generated content, and account settings as if the agent were the user itself.
Platforms must maintain third-party accessible interfaces to allow an external user’s agent hosted and operated on another service to interact with the covered platform. Furthermore, the Act's definition of user data for social media platforms expressly includes “social graph data,” defined as the user's own content, their responses to others' content (explicitly excluding third-party users' content), public profile, and metadata. Companies that fail to comply with the Act face civil penalties for each user.
Section 3(g)(3) of the bill would make the following duties non-waivable by contract, terms of service, or any form of user consent. An AI agent under the Act must:
The Act requires that AI agent providers register with the Federal Trade Commission (FTC) before being allowed to access a large platform’s agent pathway. The Act imposes a duty on large platforms to provide access on fair, reasonable, and non-discriminatory (FRAND) terms, with the FTC empowered to review fees and access denials.
The Act is government overreach that would force companies to invest significant resources in creating products without demand instead of focusing on what their customers want.
Allowing AI agents to interact with platforms on a user’s behalf has strong merit and may become the standard way accounts are managed in the future. For example, agents may become the most powerful and effective vehicle for parental account management for minors. However, this bill puts the cart before the horse and creates an onerous regulatory environment for a future that might not exist. AI agents are not sufficiently consistent at meeting many of the requirements imposed by the Act. Likewise, many large online platforms have legitimate competition, privacy, and intellectual property constraints that may be violated by forced access to their services.
A better way would be to allow this kind of data management as an option provided to an end user for a specific kind of data or service rather than to force it upon online services with a sufficiently large userbase. At the very least, a cleaner version of this Act with less FTC involvement, less strict penalties for violations, a restricted industry domain, and explicit good-faith safe harbors for both platforms and agent providers would likely solve many of these detriments.
The SAFE AI Act prohibits federal agencies from purchasing or using an AI image- or video-generating model that an agency head, the Administrator of General Services, or a federal court has determined fails safeguard standards against generating child sexual abuse material (CSAM) or non-consensual intimate images of identifiable adults (NCII).
The Act also prohibits offering tools primarily designed to circumvent, remove, or tamper with a model's technical controls, and creates a private right of action for an individual whose likeness is contained in any CSAM, an individual injured by a developer failing to implement sufficiently robust technical controls to prevent their likeness from being so used, and a developer whose technical controls are subverted.
The SAFE AI Act attempts to address AI-generated CSAM and NCII by limiting the federal government's ability to purchase services from vendors whose models produce them. This governmental apparatus would wield power as a consumer rather than a regulator to address a problem within a market.
The Act's construction may effectively bar government access to open-source and open-weight models, because their safeguards can be removed by parties the original developer does not control and because anyone who contributes to the open-source model is swept into its developer liability. Furthermore, the Act lets an agency head or the GSA Administrator convert non-compliance with voluntary NIST guidance into a de facto procurement bar, with no stated process, notice, or opportunity to appeal. This turns voluntary guidance into binding regulation without the procedural safeguards that would accompany actual rulemaking.
This bill would establish a new tax on AI to fund the National Workforce Transition Fund that would attempt to help workers adapt to labor market changes resulting from AI. Somewhat ironically, workers would not need to demonstrate that they were impacted by AI in order to qualify for assistance from the fund.
The new tax on AI is arbitrary and harmful. AI is the latest iteration of countless technological innovations—from tractors to computers to industrial robotics—that have increased productivity while changing the nature of work. Penalizing investments in AI by precluding companies from accounting for the full cost of their investments when they incur the costs will distort investment in one class of capital—namely the one that currently has the most potential for productivity gains. This would set a dangerous precedent and violate the principle of a neutral tax system.
The AI tax would simultaneously reduce productive investments while using the distorted tax revenues to fund ineffective federal job training programs and windfall grants to employers. Federal job training programs have an unfailing record of failure, and there’s little reason to believe that AI-related training programs administered through the same lethargic federal bureaucracy will suddenly prove responsive and effective. Moreover, the subsidies for employers to retain, redeploy, or "responsibly transition" workers invite political favoritism and encourage businesses to seek government assistance for workforce decisions they would have undertaken more efficiently without federal funding. With authorized spending for housing, childcare, transportation, health benefits, and virtually any other support the Secretary determines appropriate during a worker's transition, what is presented as an AI workforce initiative could function as an open-ended social spending program.
One positive component is improving the federal government’s measurement of AI and its effects on workers and industries. Such measures could help inform the private sector to pursue the most productive investments and the most valuable skills to teach workers.
This section would authorize approximately $2.4 billion in new federal spending to expand AI doctoral education through grants to colleges and universities. Federal funding comes with strings attached: federally specified program curricula and design. This would presume the Congress of 2026 to be the authority on AI, and capable of writing in law what AI PhD students should study for the foreseeable future. That is a red-tape recipe for how to cede the AI race to China instead of how to win it.
Most workers in America recognize that they will have to learn something about AI, and most are motivated to do so because of the opportunities the technology brings. Similarly, colleges and universities that want to attract students and provide them with an education that leads to successful and high-paying careers will—and already are—incorporating AI into existing curriculum and building new AI education programs. The potential productivity gains of AI are enough to incentivize institutions to develop effective AI education programs and for students to enroll in them. Rather than spending billions of dollars on federally prescribed AI education, policymakers should look to reduce existing regulations and accreditation requirements that unnecessarily impede the creation of responsive new education programs.
The Act would require that AI developers provide “access to the frontier artificial intelligence model, including model’s weights, configuration files, runtimes, or software libraries necessary to operate the frontier artificial intelligence model,” which creates a risk of leaking trade secrets.
An AI risk board might be a useful tool if used for the purpose of educating government about changes in the industry. Instead, the Act has the government dictate to industry what its safety standards should be, as if the government understands AI models better than the engineers who create them. The government can regulate its own use of frontier AI models and what bureaucrats should and should not be able to do with them, but it should not regulate which frontier AI models are available to the public or not. Rather than a standing board with no sunset, a task force covering 18 months could provide a clear timeline and recommendations. A task force of federal members drawn from the relevant agencies, plus nine private-sector members—three each of AI providers, users, and researchers, with one of each appointed by the Secretaries of Defense, Commerce, and Homeland Security—could provide the government with input that would produce timely and relevant recommendations for AI safety.
While many are concerned about the cybersecurity risks that AI models create, it is important to recognize that many of these novel risks are transient and reflect the lack of widespread adoption or knowledge of AI. New models can possess powerful offensive attack capabilities that should be taken seriously. These capabilities also come with superior cyber defense abilities. The government is completely unsuited for the task of thoughtfully weighing the balance of this aspect of new models. Letting a governmental entity have the power to preclear powerful models is not only an overreach into the free market, but would also invite disaster. Such an ill-suited governmental body is unlikely to address legitimate concerns while slowing down and gating access to new models. Importantly, as AI matures, the methods for strengthening security will mature as well, and many of the issues that this risk board seeks to address will be naturally mitigated.
BOTTOMLINE: Congress should reject “A Framework for America's AI Future.” While there are several useful concepts within this package, the framework misses several key principles that should be included in federal regulations on AI.