AI Rules Tighten as Models and Chips Scale

Governments are moving fast to set guardrails for artificial intelligence as the technology spreads through business and daily life. New rules in Europe and executive actions in the United States aim to make powerful systems safer and more transparent. Industry is also racing ahead, building larger models and more capable chips. The result is a new phase in AI: rapid growth paired with rising oversight.

The new rulebook

Europe has adopted the EU AI Act, which the European Commission has described as “the first comprehensive law on AI worldwide.” It uses a risk-based approach. Low-risk tools face light duties. High-risk systems, such as those used in hiring, education, or critical infrastructure, must meet strict requirements before they reach the market. Some practices are banned outright, including social scoring by public authorities. Real-time biometric identification in public places is heavily restricted.

In the United States, a 2023 presidential executive order directed federal agencies to develop safety tests, watermarking guidance for AI-generated content, and stronger privacy protections. It leaned on existing authorities to seek information from companies training the most powerful models. The National Institute of Standards and Technology (NIST) is expanding its technical guidance, including the AI Risk Management Framework, to help organizations evaluate and manage risk.

The United Kingdom convened the 2023 AI Safety Summit and launched an AI Safety Institute to test so-called frontier models. More than two dozen countries signed the Bletchley Declaration to cooperate on safety. In parallel, major AI developers in the U.S. and beyond made voluntary commitments to red-team models, disclose capabilities and limits, and invest in cybersecurity.

Why it matters

AI is moving from pilot projects to core operations. Banks use it to detect fraud. Hospitals explore it for imaging and triage. Schools test it for tutoring. Mistakes can be costly. Regulators say clear rules can reduce harm while keeping innovation on track.

Industry leaders have urged caution. “If this technology goes wrong, it can go quite wrong,” OpenAI chief executive Sam Altman told U.S. senators in May 2023. Geoffrey Hinton, a pioneering researcher, warned the BBC in 2023: “It’s hard to see how you can prevent the bad actors from using it for bad things.” Their comments underscored a shared concern about scale and misuse.

NIST and other standards bodies stress trustworthy AI principles: validity and reliability, safety, security, accountability, transparency, privacy, and fairness. These ideas now appear in laws, procurement rules, and investor checklists. They are also shaping how teams build systems, from data collection to model evaluation and deployment.

What changes for companies

The shift is practical. Organizations will need to show how they build and monitor AI. Many are creating cross-functional teams and updating documentation. Key steps include:

  • Map your AI footprint: Build an inventory of models and use cases. Flag high-risk applications early.
  • Classify and assess risk: Use structured reviews to weigh impacts on safety, rights, and compliance.
  • Data governance: Track data sources, consent, and quality. Reduce bias and label synthetic data.
  • Testing and red-teaming: Stress-test models for safety, security, and reliability. Document findings.
  • Transparency: Provide clear user disclosures. Publish model or system cards where appropriate.
  • Incident response: Set up channels to report failures or harms. Log and fix issues fast.
  • Vendor management: Add AI clauses to contracts. Ask suppliers for evidence of controls.
  • Content authenticity: Adopt watermarking or provenance tools for media when feasible.

For startups and small firms, the burden can feel heavy. Policymakers in Europe included sandboxes and support measures. U.S. agencies are updating grants and technical assistance. The goal is to help smaller players comply without stifling new ideas.

Timelines and enforcement

The EU AI Act rolls out in phases over the next few years. Bans on certain practices take effect first. Rules for general-purpose models and high-risk systems follow. National authorities will enforce the law, backed by an EU-level office focused on advanced models. In the U.S., the executive order drives agency rulemaking and guidance on a staggered schedule. Early priorities include testing protocols, critical infrastructure risk, and federal procurement standards.

Enforcement capacity remains a question. Regulators need technical talent, tools, and testbeds. Industry cooperation will be key, particularly for auditing complex systems. Cross-border coordination will also matter: many AI providers operate across markets, and their models update frequently.

The hardware factor

Compute power is the fuel of modern AI. As models grow, demand for advanced chips has surged. NVIDIA unveiled a new generation of AI processors in 2024, part of a roadmap to speed training and cut costs. AMD and custom silicon from cloud providers add to the mix. These advances enable bigger models and faster deployment, but they also raise policy questions. Safety proposals in the U.S. and Europe tie some obligations to model capabilities and the resources used to train them.

Cloud platforms now bundle AI governance features, from access controls to content filters and monitoring dashboards. That can lower barriers for compliance. It also concentrates influence in a handful of infrastructure providers. Energy use is another concern. Data centers powering AI workloads require new capacity and grid planning, prompting talks between technology firms, utilities, and governments.

Background and context

Global coordination is not new. The OECD adopted AI principles in 2019 that many countries still cite. NIST released its AI Risk Management Framework in 2023 to offer a common language for risk. Civil society groups have pushed for impact assessments and redress mechanisms. Developers have participated in public tests, including a large independent red-teaming exercise at DEF CON 2023, to probe model behavior.

Critics warn that rules can be vague or slow to adapt. Definitions of “high risk” and thresholds for advanced models are still being refined. Open-source advocates fear unintended barriers. Human rights groups want clearer limits on biometric surveillance. Businesses ask for harmonized standards to avoid a patchwork of audits and paperwork.

What to watch next

  • Technical standards: Work at NIST, ISO/IEC, and European standards bodies will translate legal text into tests and controls.
  • Model disclosures: Expect more reporting on training data, benchmark methods, and safety evaluations for advanced systems.
  • Case law: Early enforcement actions and court rulings will define boundaries, especially for biometrics and workplace uses.
  • Public-sector use: Government procurement rules could set de facto norms across industries.
  • Global forums: More summits and working groups will try to align risk frameworks and share research on frontier models.

The bottom line

AI is scaling fast, and so are the rules around it. The direction is clear: more testing, more transparency, and more responsibility, especially for high-impact use cases. Developers that bake in safety and documentation early will move faster later. Regulators face their own test: keeping pace with a technology that learns and changes every day, without shutting the door on the benefits it can bring.

MOST POPULAR