Should AI development be paused or restricted?
The FLI open letter calls for a pause on powerful AI training, while the CEPA article argues against pausing progress.
Should AI be paused? I'll gather the latest arguments for and against restrictions.
Should AI development be paused or restricted is one of the most active debates in technology policy. There is no consensus - even among leading AI labs - and the discussion has shifted from a total pause to how to restrict it responsibly.
What "pause" proposals have looked like
The most famous call was the Future of Life Institute (FLI) open letter in March 2023:
> "We call on all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4", published as "Pause Giant AI Experiments: An Open Letter".
It argued that AI systems with human-competitive intelligence can pose profound risks and that "Advanced AI could represent a profound change in the history of life on Earth", citing risks such as AI-generated propaganda, extreme automation of jobs, human obsolescence, and loss of control.
The letter said the pause should be public and verifiable, and "If such a pause cannot be enacted quickly, governments should step in and institute a moratorium". It clarified this did not mean a pause on all AI development, but a stepping back from the "dangerous race to ever-larger unpredictable black-box models with emergent capabilities".
That pause did not happen. Instead, "the next three years brought increasingly capable models from OpenAI, Anthropic, Google, Meta, xAI, and others".
The debate reignited in late 2025-2026 when Anthropic CEO Dario Amodei wrote that "we must slow the pace" of frontier AI development, arguing AI capabilities are advancing unexpectedly fast and approaching levels humans may be unable to control. Sam Altman, Demis Hassabis and Elon Musk have all expressed agreement.
Arguments FOR a pause or restrictions
1. Capabilities are outpacing safety understanding
Proponents argue labs are "locked in an out-of-control race to develop and deploy ever more powerful digital minds that no one – not even their creators – can understand, predict, or reliably control". Recent models have "behaved beyond human control" and are "beginning to act as though it has thoughts of its own".
Risks cited include:
2. Need time to build governance
The FLI letter called for using a pause to jointly develop "shared safety protocols... rigorously audited and overseen by independent outside experts" and to ensure systems are "safe beyond a reasonable doubt", plus "new and capable regulatory authorities dedicated to AI; oversight and tracking of highly capable AI systems and large pools of computational capability; provenance and watermarking systems... auditing and certification ecosystem; liability for AI-caused harm".
Advocates argue powerful systems should be developed "only once we are confident that their effects will be positive and their risks will be manageable".
Arguments AGAINST a pause or broad restriction
1. Global coordination is nearly impossible - the prisoner’s dilemma
Critics ask "what China will do? Beijing called Anthropic CEO Dario Amodei's warning 'fearmongering'". If China isn't on board, "capable AI models will be available regardless of what the US does".
More broadly: "If every developer accepts restrictions and keeps its promises, there is little problem. But each fears that a competitor could break the agreement and develop a far more powerful model". The Trump administration has argued that slowing development "would only give China a chance to catch up, while Beijing has denounced the idea as a Cold War-style attempt to preserve U.S. AI dominance". The recent US-China summit produced "No deal to slow or pause AI progress".
Unlike nuclear weapons, which are "developed by governments and require large facilities such as uranium enrichment plants, making them relatively easy to monitor. AI is largely developed by private companies, while data centers are far smaller. Whether effective international monitoring is possible is uncertain".
2. A pause could help attackers and hurt defenders
Some argue "Pausing AI development risks accentuating asymmetry between able defenders and rogue actors, whether they be nation states, criminals, terrorists, or simply misaligned models. AI defense can reduce or offset the risk of AI offense. If we delay access to the most capable models, defenders' hands would be tied".
Example cited: when bots attacked Hugging Face, a Chinese open-weight model helped diagnose the danger, and AI cyber defense has been used by Ukraine and the EU's cyber agency ENISA.
3. Huge opportunity cost
"AI technologies show significant promise for addressing major challenges including disease, climate change, and poverty". A McKinsey estimate cited that AI could contribute $13 trillion annually to global output by 2030, with more recent data showing AI drove 40% of U.S. real GDP growth in Q1 2025.
Critics note: "Limiting AI benefits. A pause might slow down the development of AI applications in crucial areas such as education, healthcare, and food", and that AI may help reduce risks from asteroids, climate, pandemics and other threats by improving prediction and detection. As one summary puts it, the evidence does not support strong-form precaution against all AI development because "the opportunity cost of forgoing AI benefits is enormous".
4. Economic and practical concerns
The Cato Institute and figures like Jack Dorsey warn that "government-mandated AI pauses shield dominant companies from competition and stifle innovation". A coordinated pause could also benefit incumbent frontier labs by giving them time to "recoup, recover and focus on profits over progress" while raising barriers for challengers.
Others argue "technological progress is inevitable, and a pause would only delay the arrival of powerful AI systems without fundamentally changing their impact".
The middle ground: Restricted but not paused
Most policy today has rejected a full halt in favor of risk-based, targeted restrictions:
EU - The most comprehensive model: The EU AI Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024 with phased rollout. Prohibitions on unacceptable-risk practices took effect 2 February 2025, governance rules and General-Purpose AI (GPAI) model obligations applied from August 2025, and full applicability arrives in 2026-2028. The law "classifies AI systems by risk level", banning some uses (like social scoring, manipulative AI) and imposing conformity assessments, transparency and auditing on high-risk uses in employment, biometrics, and critical infrastructure.
It applies extraterritorially to any provider placing AI on the EU market, with fines up to 7% of global annual turnover.
US - Fragmented and innovation-first: The US has "No federal AI statute; Executive Orders 14179 and 14365; state laws" and relies on "voluntary industry commitments and sector-specific guidance rather than comprehensive legislation, reflecting a philosophy that regulation should not stifle innovation". An increasing number of states including California and Kentucky have enacted their own AI laws effective from 2025.
Other approaches: Proposals that have gained traction include: requiring independent review before training frontier systems, limiting the rate of growth of compute used for new models, implementing provenance/watermarking, liability for harm, and focusing regulation at the frontier-model level where "novel risks are most concentrated; application-only regulation misses" them.
So should development be paused?
Most experts and policymakers today land between these poles: not "pause everything," but not "unrestricted deployment" either. Common ground includes rigorous safety testing before release, transparency and auditing for frontier models, controls on physical bottlenecks like DNA synthesis and pathogen detection rather than just model intelligence itself, and international dialogue even if a binding treaty is unlikely given geopolitical rivalry.
In practice, humanity has so far chosen to regulate and slow rather than pause — attempting to harden critical infrastructure and increase monitoring while allowing development to continue under stricter rules.
Citations (3)
- Pause Giant AI Experiments: An Open Letter
Backs: “"We call on all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4", published as "Pause Giant AI Experiments: An Open Letter". It argued that AI systems with human-competitive intelligence can pose profound risks and that "Advanced AI could represent a profound…”
evidence, official document, points to yes. Claim: The FLI open letter called for a six-month pause on training systems more powerful than GPT-4, citing profound risks to society and humanity.
- Can AI development be controlled? The global debate over AI regulation
Backs: “Instead, "the next three years brought increasingly capable models from OpenAI, Anthropic, Google, Meta, xAI, and others". The debate reignited in late 2025-2026 when Anthropic CEO Dario Amodei wrote that "we must slow the pace" of frontier AI development, arguing AI capabilities are advancing unexpectedly fast and…”
background, article, neutral. Claim: The pause did not happen; capable models kept arriving, and the debate reignited when Anthropic's Amodei said 'we must slow the pace'.
- AI: Let’s Not Pause Progress
Backs: “Global coordination is nearly impossible - the prisoner’s dilemma. Critics ask "what China will do?”
evidence, article, points to no. Claim: Critics argue global coordination on a pause is nearly impossible, since China may not comply, so capable AI models will exist regardless.