Congress Stalls on a Federal AI Framework as Public Risks Intensify

Congress entered its August recess without advancing a federal AI framework from committee, despite months of warnings about safety, jobs, children, and data centers. The delay matters because states are no longer waiting for Washington. They are writing narrower rules around real-world harms while federal lawmakers remain divided over who should regulate AI.

The White House sent Congress a national policy blueprint in March 2026. It called for child protections, intellectual property rules, workforce preparation, energy safeguards, and broad limits on state regulation. Five months later, those recommendations had not become a comprehensive federal bill with a viable path through both chambers.

That gap has created the central conflict in American AI policy. Washington wants national uniformity, but states want authority to act before federal protections exist. The dispute is no longer a technical disagreement about regulatory design. It now affects electricity bills, hiring decisions, chatbot safety, model testing, and the political power of AI companies.

Congress Has Proposals but No Federal AI Framework

The defining fact is not that Congress lacks AI proposals. It is that none has produced a governing consensus.

The White House released its AI policy framework on March 20. It asked lawmakers to create a national approach without establishing a new federal AI regulator.

Its recommendations covered child safety, local effects from AI infrastructure, intellectual property, government pressure on speech, workforce readiness, and law enforcement. The administration also wanted Congress to preempt state laws that imposed what it considered excessive burdens on AI development.

Preemption means a federal law displaces state rules covering the same subject. Its scope determines whether states retain authority over issues such as automated hiring, model transparency, discrimination, and chatbot interactions.

That question has repeatedly blocked agreement. Many Republicans and technology companies argue that inconsistent state rules raise compliance costs and slow domestic development. Democrats, state officials, and consumer advocates often oppose removing state authority before Congress enacts enforceable federal protections.

House Republican leaders welcomed the White House proposal, but bipartisan praise for national standards did not resolve that disagreement. The parties remained divided over liability, enforcement, safety testing, and the amount of power states should keep.

Congress has considered narrower legislation. Bills address deepfakes, government procurement, children’s online safety, research, national security, workforce training, and AI infrastructure. Those measures show legislative activity, but they do not add up to one enforceable national system.

A bipartisan discussion draft called the Great American Artificial Intelligence Act appeared in June. According to draft bill reporting, its 269 pages combined federal requirements with a three-year restriction on certain state laws.

The draft demonstrated that lawmakers could translate broad principles into legislative text. It also exposed the same unresolved problem. Critics said its federal rules would become a ceiling that prevented states from addressing new harms.

Senator Mark Warner offered another path in July. His AI legislative agenda combined model security, workforce investment, infrastructure planning, child protection, and national security measures.

Warner proposed mandatory secure testing environments for the most advanced models. His package also included voluntary safety incident reporting modeled on aviation practices. Other provisions focused on worker training and the physical infrastructure supporting AI.

That agenda broadened the policy conversation beyond state preemption. It did not, however, create an immediate bipartisan route through multiple committees and both chambers.

Jurisdiction is another obstacle. AI policy touches commerce, the judiciary, science, labor, homeland security, intelligence, energy, and armed services. Each committee controls only part of the issue, and members do not share one definition of acceptable risk.

The result is a crowded field of bills without a unified legislative center. Congress can hold hearings and introduce packages while still postponing its hardest choices.

Google News searches can make that activity appear more coherent than it is. Readers see a steady flow of frameworks, hearings, executive orders, and individual bills. Yet the institutional outcome remains simple: Congress has not enacted comprehensive federal AI rules.

That distinction explains why the current slowdown matters. The government is not beginning from zero. It is struggling to reconcile proposals that assign power, liability, and enforcement in fundamentally different ways.

State Laws Are Filling the Federal Vacuum

Every month without federal action gives states a larger role in determining how Americans encounter AI.

State lawmakers introduced more than 1,000 bills containing the term AI during the 2025 legislative season, according to a Congressional Research Service review. At least 48 states and Puerto Rico participated in that activity.

Those bills did not all propose sweeping model regulation. Many focused on particular settings where automated systems can create direct consequences. Employment, education, health care, insurance, elections, policing, and consumer disclosure have attracted sustained attention.

This targeted approach is politically important. Earlier proposals sometimes tried to impose broad duties on developers across many uses. Governors and industry groups resisted measures they considered difficult to administer or too costly for smaller companies.

Newer bills often regulate a decision or interaction instead. A state can require notice when an employer uses an automated system. It can restrict deceptive synthetic media or set rules for chatbots communicating with children.

California lawmakers have considered restrictions on employers relying entirely on AI to discipline or dismiss workers. They have also pursued additional safeguards for chatbot interactions involving minors.

Connecticut required disclosure when employees or applicants interact with employment-related AI systems. Other states have addressed political deepfakes, automated discrimination, and transparency for high-risk decisions.

These laws create compliance complexity, but they also reflect distinct local concerns. A state facing rapid data center construction may prioritize water and electricity. Another may focus on biometric privacy or automated employment decisions.

The White House sees that diversity as a fragmented patchwork. Its preferred federal AI framework would preserve some state authority while blocking rules that directly regulate model development or impose broad developer liability.

Opponents see the sequence differently. They argue that Congress should establish meaningful national protections before it restricts states. Preemption without federal enforcement would remove existing safeguards while leaving affected residents with fewer remedies.

The Senate confronted that concern in 2025. A proposed ten-year restriction on state AI regulation was removed from a major budget bill by a 99-to-1 vote.

That result did not settle the policy question permanently. It did show that opposition to broad preemption crossed party lines and extended beyond traditional technology critics.

State attorneys general and lawmakers have since continued pressing Congress to preserve local authority. Their argument rests partly on institutional timing. States are already investigating harms and passing rules while Congress debates a future system.

The administration has tried to influence that balance through executive action. A December 2025 order directed federal agencies to identify and challenge state laws viewed as obstacles to national AI policy.

Executive authority cannot easily substitute for legislation. Agencies can litigate, attach conditions to eligible funding, and shape procurement. They cannot create the same durable nationwide liability and enforcement structure that Congress can enact.

Legal challenges would also force courts to examine the federal government’s statutory authority. That creates another period of uncertainty for companies, state officials, and people affected by automated systems.

For businesses, the state-by-state approach brings real operational costs. Teams must identify which systems fall under each law, document decision processes, update notices, and manage different effective dates.

Those costs do not prove that preemption is the correct answer. They show why the absence of federal legislation has consequences even before regulators bring a case.

For individuals, the practical question is simpler. If an AI system rejects an application, raises a utility bill, or exposes a child to harmful content, which government has both authority and a usable remedy?

Congress has not answered that question nationally. States are answering it in pieces.

AI Risks Are Moving Faster Than the Legislative Calendar

Congress is debating institutional authority while AI risks are becoming visible in ordinary work, infrastructure, and consumer services.

Child safety has become one of the few areas with broad rhetorical agreement. Lawmakers from both parties say conversational systems should not manipulate minors or provide dangerous guidance.

Agreement on the goal has not produced agreement on implementation. Legislators still differ over age verification, platform liability, parental controls, privacy, and the relationship between AI-specific rules and broader online safety bills.

Employment presents a similar problem. Companies can use AI to screen applications, monitor performance, recommend schedules, or support disciplinary decisions. Each use involves different levels of human oversight and different consequences for workers.

A disclosure rule tells a person that automation is present. It does not necessarily explain the data used, the reason for a result, or the available appeal process.

Federal lawmakers must therefore decide whether transparency alone is sufficient. They must also determine when organizations need testing, documentation, human review, or legal accountability.

Model safety creates an even harder challenge. Frontier models are general-purpose systems trained at substantial scale and capable of supporting many downstream applications. Regulators cannot predict every use before deployment.

Warner’s proposal addresses that uncertainty through secure testing and incident reporting. A controlled testing environment can help examine cybersecurity or national security risks before broad release.

However, testing is not a complete guarantee. Models change after deployment through updates, integrations, tools, and user behavior. A benchmark result cannot represent every environment where the system will operate.

Mandatory incident reporting could help regulators detect recurring failures. Its value depends on clear thresholds, confidentiality protections, enforcement, and the government’s capacity to analyze reports.

The White House prefers a lighter federal structure focused on innovation and existing regulators. It argues that excessive liability can discourage investment and weaken American competition.

That competitiveness concern carries political weight. The United States wants domestic companies to lead in advanced models, chips, cloud infrastructure, and AI applications. Lawmakers also view AI capacity as a national security asset.

Yet global competition does not eliminate domestic risk. It changes the tradeoff. Congress must decide which safeguards reduce preventable harm without creating rules that only the largest companies can afford.

A framework centered entirely on voluntary commitments would leave enforcement gaps. A rigid licensing system could entrench incumbents by imposing costs that smaller developers cannot meet.

This is why the legislative delay cannot be reduced to simple indifference. The policy choices involve genuine conflicts between speed, accountability, competition, federalism, and technical uncertainty.

Still, complexity does not erase responsibility. Congress routinely regulates sectors with changing technology and incomplete information. It can update standards, delegate technical work, and require agencies to review rules.

The delay becomes more consequential as AI systems gain access to tools and sensitive information. An assistant that drafts text presents one class of risk. A system that initiates transactions or changes business records presents another.

Organizations are already developing internal controls because federal law offers limited AI-specific direction. They maintain model inventories, approval records, incident logs, and human review procedures.

Knowledge workers face a related challenge. They need to distinguish model output from trusted evidence and preserve the context behind consequential decisions. A searchable AI knowledge base can support that discipline, but software cannot replace enforceable rights.

The regulatory question concerns who bears responsibility when internal controls fail. Congress has not established a consistent answer across sectors or applications.

That leaves agencies applying older statutes where possible. Consumer protection, civil rights, labor, financial, health, and copyright laws still apply, even when Congress has not enacted an AI statute.

Existing law can address deception or discrimination. It may not impose the documentation, testing, and reporting duties needed to identify a problem before harm occurs.

The growing mismatch between deployment speed and legislative speed is therefore the central risk. AI adoption proceeds through product releases and enterprise contracts. Congress proceeds through committees, negotiations, and election cycles.

Data Centers Turn AI Policy Into a Local Cost Fight

The federal debate changed when AI’s physical infrastructure began affecting land, water, power systems, and household utility costs.

AI is often discussed as software, but advanced models depend on data centers filled with processors, networking equipment, storage, and cooling systems. Those facilities require substantial electricity and physical infrastructure.

Communities experience that expansion through zoning applications, transmission projects, water planning, tax incentives, construction traffic, and potential changes to electricity demand.

Local opposition has grown across political lines. Some residents object to water use or industrial development. Others question whether tax benefits and promised jobs justify infrastructure costs.

These concerns weaken the idea that one national rule can resolve every AI dispute. Model safety may require federal expertise, while land use and utility planning traditionally involve state and local authority.

The White House framework recognized the ratepayer issue. It called for policies intended to prevent households from absorbing electricity infrastructure costs created by major AI facilities.

Several large technology companies made voluntary commitments to cover costs associated with their data center power needs. Codifying such principles would require detailed rules about generation, transmission, interconnection, and utility accounting.

A pledge is easier to announce than to enforce. Costs can appear through multiple channels, including new substations, grid upgrades, reserve capacity, and long-term power contracts.

Regulators must determine which investments serve one customer and which benefit the wider grid. That calculation varies across utility territories and state regulatory systems.

Members of Congress have introduced bills addressing data center disclosure, federal land, energy demand, and local effects. Some lawmakers want faster permitting because they view infrastructure capacity as essential to national AI leadership.

Others want a pause. Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez introduced an AI data center bill that would suspend certain construction until national safeguards exist.

The proposal links infrastructure directly to broader AI governance. Its supporters cite electricity, water, labor, privacy, community participation, and economic distribution.

A national moratorium faces long political odds. It nevertheless reveals how far the debate has moved beyond abstract model rules.

The primary opponent is now national uniformity without enforceable protections versus state and local action responding to immediate costs. Data centers make that conflict visible because communities cannot postpone infrastructure decisions indefinitely.

A county must decide whether to approve a project. A utility must plan for demand. A state regulator must decide who pays for upgrades.

Congress can delay a comprehensive bill without freezing those decisions. That asymmetry gives state and local governments increasing influence over the conditions of AI growth.

Technology companies also face strategic pressure. If they oppose local restrictions while federal legislation remains stalled, they risk appearing to seek freedom from both levels of oversight.

Some companies have softened their resistance to state rules when those rules begin to converge. Consistent state requirements can eventually resemble a de facto national standard, even without an act of Congress.

That outcome would be less efficient than federal legislation, but it is institutionally plausible. Privacy policy followed a comparable pattern, with states filling gaps left by Congress.

Data center disputes may accelerate that process. Unlike hypothetical future model risks, electricity bills and land-use conflicts produce organized local constituencies.

Those constituencies do not fit neatly into national party categories. Conservatives can oppose industrial projects near their communities. Progressives can object to environmental costs and corporate concentration.

The 2026 election cycle increases the pressure. Candidates can criticize AI companies from different ideological positions while offering incompatible policy solutions.

Google News coverage reflects that fragmentation. One headline emphasizes safety incidents, another focuses on jobs, and another tracks local data center opposition. Congress must eventually connect those concerns within a workable legal structure.

Preemption Without Guardrails Is the Central Tradeoff

A federal standard can reduce regulatory conflict, but it becomes dangerous when it removes state protections before replacing them.

Supporters of preemption make a serious argument. A model provider serving customers nationwide cannot redesign its core system for dozens of conflicting technical standards.

Fragmentation can also harm smaller companies. Large platforms can hire compliance teams, retain counsel in every state, and build separate reporting systems. Startups have fewer resources.

Uniform rules can improve accountability when they create one clear baseline. A federal law could define high-risk uses, establish testing duties, assign regulator authority, and provide consistent remedies.

The problem is sequencing. Congress has repeatedly discussed blocking state laws without first agreeing on a federal floor.

A regulatory floor sets minimum protections while allowing stricter state action. A ceiling prevents states from going further. Many current disputes concern where federal legislation should sit between those models.

The White House approach favors substantial uniformity and limited liability for developers. Its critics say this would protect companies from state enforcement while placing too much responsibility on downstream users.

Developer liability is difficult because general-purpose models support unpredictable applications. Making a developer responsible for every misuse would be unworkable.

Removing developer responsibility entirely would also create weak incentives. Developers control training, evaluations, release decisions, safety features, and access to technical evidence unavailable to users.

A credible federal AI framework needs responsibility at each layer. Developers should address risks they can reasonably test. Deployers should evaluate context, data, and human oversight. Users should remain accountable for intentional misuse.

Congress must also choose an enforcement model. Existing agencies possess sector expertise, but no single agency sees the entire market.

A new regulator could concentrate technical knowledge. It could also create jurisdictional conflict, delay deployment, and become vulnerable to political capture.

The White House has opposed creating a new AI agency. That leaves Congress with a coordination problem among regulators whose powers and resources differ.

NIST offers voluntary technical guidance through its AI Risk Management Framework. That framework helps organizations identify and manage risk, but it does not itself create legal obligations.

Voluntary standards work best when buyers, insurers, courts, or regulators create incentives for adoption. They are weaker when the organization causing a risk does not bear its full cost.

The same issue applies to voluntary incident reporting. Companies may report events when confidentiality and legal protections are clear. They may remain silent when disclosure creates reputational or litigation risk.

Congress can design safe-harbor provisions that encourage reporting without excusing negligence. Aviation offers a useful reference, but software incidents differ from transportation accidents.

AI systems are widely distributed, frequently updated, and embedded in products operated by separate companies. Determining when one event becomes a reportable incident requires technical precision.

Critics also warn that federal legislation could freeze today’s assumptions into law. AI capabilities and business models continue to change, so definitions may age quickly.

That risk favors adaptable standards and periodic review. It does not justify indefinite inaction.

Congress could require agencies to update technical criteria while retaining clear statutory rights. It could also use sunset clauses for experimental provisions.

The greatest overclaim would be that any single framework can make AI safe. Regulation can improve incentives, transparency, evidence, and remedies. It cannot eliminate uncertainty or malicious behavior.

The opposite overclaim is that regulation necessarily destroys innovation. Safety testing, documentation, and incident reporting are already standard practices in other consequential industries.

The useful question is which obligations match which risks. A chatbot offering entertainment should not face the same controls as a system influencing medical care or critical infrastructure.

Risk-based regulation sounds straightforward, but classification creates its own disputes. Companies may argue that a product falls outside a high-risk category. Regulators need access to evidence to evaluate that claim.

That brings the debate back to enforceability. Broad principles attract bipartisan statements. Binding definitions, audits, liability, and remedies determine whether a law changes behavior.

Congress remains slow because those details distribute power and cost. The federal AI framework debate is ultimately a debate over who must act before harm, and who pays afterward.

Three Signals Will Show Whether Congress Can Move

The next phase will be measured by committee action, state enforcement, and concrete federal responses to model or infrastructure incidents.

The first signal is whether a comprehensive proposal receives a committee markup. A markup is the formal meeting where lawmakers debate, amend, and vote on legislative text.

No national framework can advance on statements alone. Committee action would show that leaders have resolved enough jurisdictional and partisan conflict to test actual provisions.

The most important details would be preemption, developer responsibility, regulator authority, and individual remedies. A bill avoiding those issues would not resolve the current conflict.

Bipartisan sponsorship would strengthen the case for progress, but it would not guarantee passage. The Senate’s vote against a broad state moratorium showed that preemption can separate lawmakers from their usual coalitions.

If a framework advances with enforceable federal protections before limiting states, the prospects for durable compromise improve. If preemption leads and safeguards remain vague, opposition will probably harden.

The second signal is how state laws operate in practice. Enforcement actions, compliance guidance, and court challenges will produce evidence that Congress currently lacks.

If state rules prove compatible and converge around shared duties, the argument that fragmentation is unmanageable becomes weaker. States could establish a practical baseline through repetition.

If conflicting definitions or technical mandates create major operational problems, the case for federal uniformity becomes stronger. Congress would then have concrete conflicts to resolve instead of hypothetical ones.

Legal challenges to the administration’s executive strategy also matter. Courts may clarify how far federal agencies can go without new legislation.

A successful federal challenge to a state law would increase executive influence. A judicial rejection would reinforce the need for Congress to act directly.

The third signal is the government’s response to a major AI incident or infrastructure conflict. Washington often moves when an abstract risk becomes a visible event with identifiable victims.

That event might involve child safety, cybersecurity, automated discrimination, critical infrastructure, or a data center cost dispute. The specific category matters less than Congress’s institutional response.

Hearings and letters can gather facts. They do not establish ongoing reporting, inspection, or remedies.

A serious response would connect the incident to durable legislation and agency capacity. A symbolic response would produce narrow messaging bills without resolving the underlying framework.

Developers and enterprise buyers should watch those signals closely. Federal requirements could change model evaluation, documentation, procurement, incident management, and contract terms.

State action already affects those practices. Organizations should not assume that a future federal law will erase every local obligation.

Teams need records showing which models they use, what data enters them, who approves high-risk uses, and how people can challenge consequential outcomes. Those records support compliance, but they also improve internal decision-making.

Workers and consumers should watch the remedy question. Transparency has limited value when a person cannot correct data, request human review, or contest a harmful decision.

The political language around AI will intensify before the midterm elections. Some candidates will frame regulation as protection from concentrated corporate power. Others will present it as a threat to American competitiveness.

That rhetoric can obscure areas of agreement. Lawmakers broadly recognize risks to children, infrastructure, national security, and workers. They disagree over enforcement and federalism.

The next meaningful development will not be another general statement supporting responsible innovation. It will be legislative text that assigns duties and survives a committee vote.

Until then, state governments will keep setting practical rules, agencies will stretch existing authorities, and companies will manage inconsistent obligations. Communities will continue making infrastructure decisions without a settled national policy.

Readers following the story through Google News should separate announcements from institutional progress. A new framework, speech, or hearing changes the conversation. Only enacted rules change the national legal baseline.

The question for Congress is no longer whether AI deserves attention. It is whether lawmakers can accept the tradeoffs required to govern it.

Watch the committees, not the slogans. Track whether federal protections arrive before state authority disappears. Then examine whether the resulting system gives people usable rights when AI affects their jobs, families, information, and communities.