What the moratorium call was actually about

In March 2023 the Future of Life Institute published an open letter signed by thousands of researchers, engineers, executives, and notable signatories (Elon Musk, Yoshua Bengio, Steve Wozniak, others) calling for a six-month pause on AI training runs "more powerful than GPT-4." The letter argued that AI labs were locked in an out-of-control race to develop and deploy systems whose risks were poorly understood.

The reception was loud, polarized, and brief. Within months the conversation had moved on. By late 2024 the letter was already an artifact of an earlier discourse.

Three years later — with GPT-5, Claude Opus 4.X, Gemini Ultra, and a wave of open-weights models having shipped — it's worth looking at what the moratorium call actually proposed, what happened instead, and what business owners should take from the episode.

What was being asked

Three specific asks from the letter:

  1. A six-month pause on training runs more capable than GPT-4. No lab complied. Anthropic, OpenAI, Google DeepMind, Meta AI, Mistral, and others continued training increasingly capable systems through 2023 and 2024.

  2. Development of shared safety protocols. This happened partially — labs published responsible-scaling policies, frontier-model safety frameworks, and pre-deployment evaluations. The protocols vary in rigor and aren't externally verified.

  3. Regulatory intervention. This happened more substantially. EU AI Act passed in 2024 and entered enforcement phases. US executive orders established baseline requirements. UK launched the AI Safety Institute. Most major jurisdictions now have some regulatory framework.

The actual outcome: no pause, partial industry self-regulation, significant external regulation.

What businesses should take from the episode, in 2026

Three lessons applicable to operating decisions:

1. Capability projections were directionally right, timeline projections were uneven

The 2023 forecasts of where AI would be by 2026 were close on raw capability and off on commercial deployment. Models are roughly where the forecasters expected — multi-modal, agentic, deeply integrated into knowledge work. Deployment moved faster in some domains (coding, writing, customer support) and slower in others (regulated industries, physical-world applications).

For business owners: capability forecasts from credible sources tend to be useful for 18-month planning. Beyond that, the variance widens significantly.

2. Regulatory shifts move slower than capability shifts, but they catch up

The EU AI Act took 4 years from proposal to enforcement. The capability frontier moved more in those 4 years than in the previous 20. Businesses that built heavily on capabilities now subject to regulation are doing migration work; businesses that built more conservatively are not.

The practical implication: if you're building an AI-dependent product, watch the regulatory pipeline. Capabilities that are clearly going to be regulated (high-risk applications, content generation in regulated contexts, biometric processing) carry implementation risk beyond technical risk.

3. The provider concentration risk is real and growing

The 2023 letter predicted that capability concentration would create market dynamics resembling early oil or railroad. Three years in, the prediction is partially playing out. Three major foundation-model providers (OpenAI, Anthropic, Google) handle the majority of high-end commercial AI workloads. Open-weight models (Llama, Mistral, others) provide alternatives but at meaningful capability and operational cost.

For businesses building on AI: maintain optionality. Architect to swap providers. Don't tie unrecoverable product decisions to one vendor's roadmap or pricing.

What's gotten better

It's easy to read the moratorium retrospective as pessimistic. Several things did get better:

  • Model safety policies are real and increasingly substantive. Anthropic's responsible-scaling policy, OpenAI's preparedness framework, and Google DeepMind's frontier-safety framework all exist as accountable documents now.
  • Independent evaluation capacity grew. METR, Apollo Research, the UK and US AI Safety Institutes, and academic groups now have meaningful capability to evaluate frontier models before deployment.
  • Red-teaming practices matured. Model deployment now typically follows months of internal and external adversarial evaluation.
  • Catastrophic misuse cases have largely been avoided. The bioweapons, cyber, and disinformation harms specifically warned about in 2023 mostly haven't materialized in the worst forms predicted. Whether that's because of safety work or because the forecasts overestimated the harm potential at this capability level, both interpretations have evidence.

What's gotten worse

Equally honest about what hasn't gone well:

  • Concentration in a small number of labs is now a structural feature of the industry. The hope that open-weight models would prevent oligopoly hasn't fully played out.
  • Compute costs for training frontier models have grown faster than most observers expected. Only a small number of organizations can train at the frontier.
  • Deployment velocity has consistently outpaced safety-research velocity. Models ship before alignment understanding matures.
  • Job displacement in knowledge work has been faster than expected in specific domains (junior coding, customer support, basic writing, image creation, basic legal research).

What this means for the work we do

Marketing MIX uses AI tools daily across our practice. We also routinely turn down work that requires more AI capability than is currently dependable. The intersection of "what AI can do" and "what we'll stake a client's outcome on" is a moving target, but in 2026 the line is roughly:

  • Yes: AI for drafting, research, translation, code generation with senior review, content variation, brainstorming, customer support augmentation, semantic search, structured data extraction.
  • No: AI for autonomous decision-making in regulated contexts, AI for content that ships without human review, AI for any task where the failure mode would damage the client's reputation or legal standing.

This is roughly the same line we drew in 2024, with a wider "yes" column and a tighter "no" column. Both lines have moved; the gap between them is wider, not narrower, than three years ago.

Related

For our AI-search visibility work: /promotion-in-ai-search-and-llm-systems. For our broader strategy practice: /marketing-strategy-development. For commentary on the search-engine shifts AI has driven: /blog/googles-monopoly-is-under-attack-chatgpt-declares-war.

Written by Maksym Stepanenko, founder of Marketing MIX. Last reviewed: 2026-05-13.