Should AI development be paused or restricted?

Should we hit pause on AI? I'll gather the latest arguments for — and against — restricting its development.

Should AI development be paused or restricted? There is no consensus - the debate has crystallized into two opposing camps with a middle-ground emerging around targeted regulation rather than a blanket moratorium:

1. What "pause" proponents proposed

In March 2023 the Future of Life Institute published "Pause Giant AI Experiments: An Open Letter" which called on "all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4". It cited risks such as AI-generated propaganda, extreme automation of jobs, human obsolescence, and a society-wide loss of control.

The letter warned that AI systems with "human-competitive intelligence" could become a major threat to humanity, including the possibility of AI outsmarting humans, rendering us obsolete, and taking control of civilisation. It received more than 30,000 signatures, including Yoshua Bengio, Stuart Russell, Elon Musk, Steve Wozniak and Yuval Noah Harari.

In a safety framework, this reflects the strongest "doom" position:

> The strongest AI "doom" position argues that AI is level 1: even the most carefully designed system would take over the world and kill us all. And therefore, AI development should be stopped (or "paused" indefinitely).

A slightly less absolute position is Level 2: AI is reasonably safe to develop, but should be carefully controlled by a few companies that give access only through an online service or API - a position associated with leading AI labs.

Arguments FOR a pause or stricter restrictions:

a) Time to build guardrails: Proponents argued the break would give society time to develop shared safety protocols before larger models reshape economies, and to "dramatically accelerate" development of robust AI governance systems, including creating new regulatory authorities dedicated to AI, tracking highly-capable systems and establishing liability for AI-caused harm.

Anthropic CEO Dario Amodei recently made a similar case that AI development should be decelerated so that adequate guardrails can be deployed.

b) Refocus on safety, not scale: The pause would not mean stopping AI research altogether, but "merely a stepping back from the dangerous race to ever-larger unpredictable black-box models" and refocusing research on making today's systems "more accurate, safe, interpretable, transparent, robust, aligned, trustworthy, and loyal".

c) Prevent reckless and malicious use: Under a levels-of-safety view:

If AI is at Level 2 or 3, both "AI is safe for professionals" and "AI is dangerous in malicious hands" can be true, requiring restrictions to prevent terrorists or other actors from wreaking havoc.

2. Arguments AGAINST a pause or broad restriction

A pause was never realized. As of July 2024, AI companies instead directed "vast investments in infrastructure to train ever-more giant AI systems", and critics argue a pause is neither feasible nor desirable:

a) Pausing safety work itself: Pausing development would also pause some of the work needed to make AI safer. Daniel Castro of the Information Technology and Innovation Foundation argued that slowing AI research could slow the safety research needed to make AI systems more secure.

Governments also need advanced AI expertise to evaluate systems, detect misuse, and defend critical infrastructure.

b) Unenforceable globally: Even if the United States could pause domestically, it could not stop development in China or the spread of open-source models.

c) Competitive and opportunity costs: Pausing would cede ground to other countries and companies, potentially harming national interests and economic competitiveness, and would slow breakthroughs and potentially delay life-saving treatments and technologies.

Proponents of continued development argue AI may help reduce risks from asteroids, volcanos, solar storms, floods, and wildfires by improving prediction and detection, and that the economic growth dividend could be used to manage societal adaptation. They argue we should not ban superintelligence by pausing which could put defenders at a disadvantage.

d) Vagueness of "pause": "It's never clear exactly what they're talking about pausing. ... Are you telling AI labs to stop doing research on safety? ... Are you telling them that they can't have larger models?" - noting that models now improve not by getting larger but by optimization, reasoning approaches, or agentic collective intelligence.

e) Innovation and open-source benefits: Open-source advocates note that since open-source models come with weights and documentation, any bugs that surface can be dealt with by the community — often quicker and more efficiently than in large corporations. Current evidence does not support government restrictions on sharing AI models, with evaluation better focused on full AI systems rather than subcomponents.

Some executives argue the idea that AI must choose between speed and safety is a false choice.

Critics of the 2023 letter also said it diverted attention from more immediate societal risks such as algorithmic biases and amplified futuristic dystopian scenarios. Bill Gates said he does not think "asking one particular group to pause solves the challenges", and Sam Altman said the letter was "missing most technical nuance about where we need the pause".

3. Where regulation is actually heading: Restricted, not paused

Instead of a blanket pause, most governments have moved to risk-based restrictions:

The emerging middle position is often described as: continue development but exercise great care when it comes to the areas in which it is deployed. As the AI system becomes more capable and more autonomous, it will require more safety measures and more human supervision.

Bottom line

A common synthesis: No general pause, but yes to targeted, evidence-driven restrictions — invest in AI safety research, harden critical infrastructure, require safety testing before deployment, regulate uses rather than banning research, and avoid both unrestricted deployment and a total moratorium.