⏱️ Lectura: 9 min
A group of industry and science leaders renewed calls this week to slow down AI, according to a report by The Hill on September 14, 2026. The request is not new: it reopens a discussion that has lasted more than three years over whether model scaling is advancing faster than the ability to control it.
📑 En este artículo
- TL;DR
- What happened
- Context and history
- Why it’s so hard to slow down AI
- How to assess risk in your own AI use
- Impact and analysis
- What’s next
- Frequently Asked Questions
- What exactly did the AI leaders cited by The Hill ask for?
- Is this the first time there’s been a call to slow down AI?
- What are scaling laws and why do they make it hard to slow down AI?
- Is there any law that requires pausing a model’s training?
- What do critics of these calls argue?
- How can I follow the debate on AI governance?
- References
The underlying question is simple: who decides when a model is too large to release without external oversight? Today that decision remains almost exclusively in the hands of the same companies competing to build it.
TL;DR
- The Hill reported on September 14, 2026, that a group of industry and science leaders called for slowing down AI.
- The request adds to the 2023 open letter from the Future of Life Institute, which proposed a six-month pause on training systems more powerful than GPT-4.
- That same year, the Center for AI Safety published a one-sentence statement on the existential risk of AI, signed by executives from OpenAI, Google DeepMind, and Anthropic.
- The new call comes amid growing investment in compute to train increasingly larger models.
- Scaling laws show that more compute predictably reduces model loss, which incentivizes continued scaling instead of pausing.
- Critics of these calls argue that a unilateral pause could concentrate AI development in the hands of those who decide not to join in.
- Today no binding legislation in the United States requires pausing a model’s training because of its size.
What happened
The Hill reported that a group of figures from the tech and academic sectors signed a new public call to slow down AI, in a context where investment in training compute keeps growing without pause. The report does not detail a full list of signatories or a concrete enforcement mechanism, which is typical of this kind of open letter: they are statements of intent, not binding commitments.
What is clear is the pattern: every so often, a segment of the AI community (researchers, former lab employees, academics) publicly voices concern about the speed at which increasingly capable models are trained and released, without any external verification mechanism prior to launch. The difference between these calls and real regulation is precisely that: none of them are mandatory, they depend on each lab’s willingness.
Context and history
The best-known letter is the one from the Future of Life Institute, published in March 2023, which called for a six-month pause on training systems more powerful than GPT-4. Hundreds of researchers and industry executives signed it, although no major company actually halted its training during that period.
Two months later, the Center for AI Safety published a one-sentence statement: mitigating the risk of extinction from AI should be a global priority, on par with pandemics and nuclear war. It was signed by, among others, executives from OpenAI, Google DeepMind, and Anthropic, which drew attention at the time: the same people driving the development of these systems were warning about their risks.
flowchart LR
A["2023: FLI letter calls for 6-month pause"] --> B["2023: CAIS statement on existential risk"]
B --> C["2026: new call reported by The Hill"]
The pattern repeats in 2026: new voices call to slow down AI, without the global regulatory landscape having changed substantially since 2023. None of the previous letters managed to stop the pace of model releases; if anything, they coincided with years of accelerated growth in compute investment.
Why it’s so hard to slow down AI
The technical reason behind the resistance to pausing has a name: scaling laws. Papers by Kaplan et al. (2020) and Hoffmann et al. (2022) showed that, within certain ranges, a language model’s loss predictably decreases as training compute, data, and parameters increase. That turns further scaling into a low-technical-risk, high-competitive-return bet, exactly the opposite of an incentive to pause.
def flops_entrenamiento(parametros, tokens):
# approximation C = 6 * N * D (Kaplan et al., 2020; Hoffmann et al., 2022)
return 6 * parametros * tokens
# 70 billion parameter model trained with 1.4 trillion tokens
compute = flops_entrenamiento(70_000_000_000, 1_400_000_000_000)
print(f"{compute:.2e} FLOPs")
This calculation approximates the total compute in FLOPs needed to train a model according to the relation C = 6ND, used as a standard reference in the scaling literature. The more investment in compute, the more predictable the improvement, which pushes labs to keep scaling rather than slow down AI on their own.
💭 Key point: scaling laws say nothing about safety or alignment: they only predict that a larger model, trained with more data, will have better training loss. Whether it’s safer is a separate question.
How to assess risk in your own AI use
Slowing down AI on a global scale depends on decisions that an individual development team doesn’t control. What is within reach of any team integrating models into production is applying an internal risk review before deploying an agent or model with autonomy.
revision_riesgo_ia:
responsable: "platform team"
antes_de_desplegar:
- verificar_model_card: true
- registrar_uso_de_datos_personales: true
- definir_limite_de_autonomia_del_agente: true
- canal_de_reporte_de_incidentes: "[email protected]"
revision_periodica: "quarterly"
This list doesn’t replace any regulation, but it provides a minimum governance baseline: which model is used, what data it touches, what autonomy it has, and who is notified if something fails. It’s the kind of control that the 2023 and 2026 open letters ask for at an industry-wide level, applied first in-house.
Impact and analysis
The immediate impact of this kind of call is usually media attention, not regulation. No previous letter achieved effective training pauses at leading companies, though they did influence public debate and the agenda of international AI safety summits.
| Initiative | Year | Format | Concrete request |
|---|---|---|---|
| FLI open letter | 2023 | Signed public letter | Pause training of systems more powerful than GPT-4 for 6 months |
| CAIS statement | 2023 | Single sentence signed by executives and scientists | Treat the risk of extinction from AI as a global priority |
| New call (The Hill) | 2026 | News report on a request from leaders | Slow the pace of development, as reported on September 14 |
The most commonly cited counterargument is competitive advantage: if one company or country pauses and another doesn’t, the one that pauses loses ground without reducing global risk, because development simply continues elsewhere.
⚠️ Watch out: several critics of these calls argue that a pause imposed only on the labs that sign it could concentrate AI development in the hands of those who decide not to join in, rather than reducing overall risk.
That’s the central limitation of any voluntary call to slow down AI: without an external verification mechanism and without coordination between countries, a unilateral pause changes who leads the race, but not necessarily how fast development advances globally.
What’s next
For now, there is no binding legislation in the United States that requires pausing a model’s training because of its size. In the European Union, the AI Act does distinguish additional obligations for general-purpose models based on their level of systemic risk, though it doesn’t impose training pauses either.
Judging by the pattern from 2023 to 2026, these calls will most likely keep appearing every time a new leap in capability generates public concern, without that yet translating into a binding global mechanism to slow down AI.
📖 Summary on Telegram: View summary
Try it yourself: read the original Future of Life Institute letter or the CAIS statement and compare their signatories with those of the new call reported by The Hill today.
Frequently Asked Questions
What exactly did the AI leaders cited by The Hill ask for?
According to the report from September 14, 2026, the central request is to slow down AI by reducing the training pace of the most advanced models, without a full list of signatories or a concrete enforcement mechanism becoming public.
Is this the first time there’s been a call to slow down AI?
No. In 2023, the FLI letter called for a six-month pause on training systems more powerful than GPT-4, and that same year CAIS published a statement on existential risk signed by executives from several leading companies in the sector.
What are scaling laws and why do they make it hard to slow down AI?
They are empirical relationships, documented by Kaplan et al. and Hoffmann et al., showing that more compute, data, and parameters predictably lower a model’s loss. That makes scaling a low-technical-risk bet, which reduces the incentive to pause.
Is there any law that requires pausing a model’s training?
No, currently no binding law in the United States requires a pause based on model size, although regulations like the European Union’s AI Act do distinguish obligations based on the system’s risk level.
What do critics of these calls argue?
They argue that a unilateral pause could concentrate AI development in the companies and countries that decide not to join in, putting those who do pause at a disadvantage without reducing risk on a global level.
How can I follow the debate on AI governance?
You can read the primary source at The Hill, the full FLI letter, the CAIS statement, and the scaling papers on arXiv cited in the references section of this article.
References
- The Hill: report from September 14, 2026, on the request from AI leaders to slow down development.
- Future of Life Institute: 2023 open letter that called for a six-month pause on training systems more powerful than GPT-4.
- Center for AI Safety: 2023 statement on the existential risk of AI, signed by executives and scientists in the sector.
- arXiv: Scaling Laws for Neural Language Models: paper by Kaplan et al. (2020) that formalized the relationship between compute, data, and performance.
- arXiv: Training Compute-Optimal Large Language Models: paper by Hoffmann et al. (2022, Chinchilla) on compute-optimal training.
📱 Do you like this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Imagen destacada: Foto de Omar:. Lopez-Rincon en Unsplash
0 Comments