Sources · Pausing frontier AI
This house would pause the development of frontier AI.
Fact bank v1 · 13 cards
- PAI-01True
“In 2023, more than 1,000 tech leaders and AI experts, including Elon Musk, signed a letter calling for a six-month pause on giant AI systems.”
Confirmed: in March 2023, more than 1,000 AI experts, researchers and backers signed an open letter. It asked for an immediate pause of at least six months on building "giant" AIs, so their dangers could be studied.
The Future of Life Institute organised the letter. Signers included Elon Musk, Apple co-founder Steve Wozniak, and engineers from Amazon, DeepMind, Google, Meta and Microsoft. It said AI labs were in an "out-of-control race". The group also put out a plan asking for things like outside safety checks on AI systems and rules on who can get huge amounts of computing power.
“more than 1,000 artificial intelligence experts, researchers and backers signed an open letter calling for an immediate pause on the creation of “giant” AIs”
House of Lords Library, UK Parliament·Artificial intelligence: Development, risks and regulation (In Focus briefing)Government / parliament · 18 Jul 2023 ↗ - PAI-02True
“Hundreds of top AI scientists and tech bosses, including OpenAI's Sam Altman, signed a statement warning that AI could cause human extinction.”
Confirmed: more than 500 business and science leaders, including people from OpenAI and Google DeepMind, signed a 23-word statement. OpenAI's boss Sam Altman was one of them. It said dealing with the risk of extinction from AI should be a global priority, like pandemics and nuclear war.
Signing a statement does not prove the risk is big. A 2024 international science report, led by AI pioneer Yoshua Bengio, said the real risk of losing control of AI is still hotly debated, because only a little research has measured it. Nature's editors argued that "doomsday" talk pulls attention away from the harm AI is causing today.
“should be a global priority alongside other societal-scale risks such as pandemics and nuclear war”
Nature·Stop talking about tomorrow's AI doomsday when AI poses risks today (Nature editorial)Peer-reviewed research · 27 Jun 2023 ↗ - PAI-03True
“The computing power used to train the top AI models is growing about four times every year.”
Confirmed: an international science report found that the computing power used to train the most advanced AI models has grown about 4 times a year since 2010.
Most of that growth comes from companies spending more money on more chips, not from chips getting better. The report says this kind of growth is a main reason AI has improved so fast. In December 2025 the UK's AI Security Institute said performance in some areas is doubling every eight months. In late 2023, the best models almost never finished software tasks that take an expert an hour (under 5% success). By mid-2025, they managed it over 40% of the time.
“the total compute used in training has increased by approximately 4x per year since 2010”
UK Department for Science, Innovation and Technology and AI Security Institute (GOV.UK)·International Scientific Report on the Safety of Advanced AI: Interim Report (chaired by Yoshua Bengio)Government / parliament · 17 May 2024 ↗Further sources for this card
- PAI-04True
“Most Americans, and even most AI experts, worry the government will not regulate AI enough.”
Confirmed: in 2024 Pew surveys, about six in ten US adults and 56% of AI experts were more worried that the government would not go far enough in regulating AI than that it would go too far.
Pew surveyed 5,410 US adults and 1,013 AI experts living in the US (people who wrote or spoke at AI conferences). The public is much more uneasy than experts: 51% of adults were more worried than excited about AI, compared with 15% of experts. Experts were more hopeful: 56% expected AI to be good for the US over the next 20 years, against 17% of adults. This report did not ask about a pause.
“Both the public and experts largely worry the U.S. government will not go far enough in regulating AI.”
Pew Research Center·How the US Public and AI Experts View Artificial IntelligenceIndependent research institute · 3 Apr 2025 ↗ - PAI-05Mostly false
“AI systems are already escaping human control and copying themselves.”
Not yet. UK government testers found no evidence of AI models trying to copy themselves, or secretly doing badly on tests, on their own. But AI is getting better at some of the steps that would be needed.
In the UK AI Security Institute's tests, success on simplified self-copying tasks went from 5% in 2023 to 60% in 2025. Models did best at early steps, like getting computing power and money, and struggled with later ones. The institute says they are unlikely to succeed in the real world for now. When told to, models could sometimes do worse on tests on purpose. In 2024, an international science report said experts broadly agreed that the AI systems known then posed no significant risk of loss of control. Experts disagree about the future.
“However, there's not yet evidence of models attempting to sandbag or self-replicate spontaneously.”
UK AI Security Institute·Frontier AI Trends ReportGovernment / parliament · 18 Dec 2025 ↗ - PAI-06Mostly false
“AI is going to wipe out 40% of the world's jobs.”
That mixes up "exposed" with "lost". The IMF found that almost 40% of jobs worldwide are exposed to AI, meaning AI could change them. In rich countries, it says about half of the exposed jobs may gain from AI.
For the other half, AI could do key tasks that people do now. That could mean fewer new hires and lower pay, and in the most extreme cases some of these jobs may disappear. About 60% of jobs are exposed in rich countries, 40% in emerging economies and 26% in low-income countries. A 2024 international science report said economists see a sudden, huge loss of jobs to AI as a fairly fringe idea for now.
“almost 40 percent of global employment is exposed to AI”
International Monetary Fund·AI Will Transform the Global Economy. Let's Make Sure It Benefits Humanity (IMF Blog, Kristalina Georgieva)Intergovernmental body · 14 Jan 2024 ↗ - PAI-07True
“Studies show AI tools make workers more productive, especially less experienced ones.”
Confirmed in a large real-world study: 5,179 customer support agents solved 14% more issues per hour on average with an AI assistant. Newer and less skilled workers improved by 34%.
The most experienced and skilled workers gained very little. The study also found customers were happier and staff were more likely to stay in their jobs. It looks at one kind of job, so on its own it does not show how much AI will help the whole economy. It was later published in the Quarterly Journal of Economics (2025).
“increases productivity, as measured by issues resolved per hour, by 14% on average, including a 34% improvement for novice and low-skilled workers”
National Bureau of Economic Research·Generative AI at Work (NBER Working Paper 31161, Brynjolfsson, Li and Raymond)Independent research institute · 1 Nov 2023 ↗ - PAI-08Mixed
“AI will supercharge the whole economy within a few years.”
Experts disagree, and many expect smaller effects. One well-known estimate found AI would raise overall productivity by no more than 0.66% over 10 years, or by less than 0.53% once harder tasks are counted. Some economists expect much bigger effects within 5 to 10 years.
Economist Daron Acemoglu worked this out from how many work tasks AI can affect and how much it saves on each one. He warns that early studies looked at easy tasks, so gains on harder tasks may be smaller. A 2024 international science report said many economists expect modest effects over the next 10 years, but these forecasts are very uncertain because nobody knows how fast AI will improve.
“these macroeconomic effects appear nontrivial but modest—no more than a 0.66% increase in total factor productivity (TFP) over 10 years”
National Bureau of Economic Research·The Simple Macroeconomics of AI (NBER Working Paper 32487)Independent research institute · 1 May 2024 ↗ - PAI-09True
“AI has already predicted the shape of almost every known protein, a big step forward for science.”
Confirmed: DeepMind's AlphaFold tool predicted the shapes of around 200 million proteins, nearly every protein known to science, and made them easy to look up.
Nature reported this in July 2022. AlphaFold is a science tool built for one job, not a chatbot. The 2023 pause letter asked to pause only "giant" AI systems, so debaters can argue about whether tools like this would be slowed down.
“DeepMind’s AlphaFold tool has determined the structures of around 200 million proteins.”
Nature·'The entire protein universe': AI predicts shape of nearly every known protein (Nature news)Peer-reviewed research · 28 Jul 2022 ↗ - PAI-10True
“Free open AI models that anyone can download are only a few months behind the best models from big companies.”
Confirmed: the UK's AI Security Institute found, using outside data, that the gap between open models and the best closed models had shrunk to between four and eight months.
Open models can be downloaded, changed and shared by anyone. That helps new ideas and outside checking. But the institute says their safety protections can be removed quickly and cheaply, while closed models are easier to watch over. It also says it is unclear whether the gap will keep shrinking.
“this gap has narrowed over the last two years and is now between four and eight months according to external data”
UK AI Security Institute·Frontier AI Trends ReportGovernment / parliament · 18 Dec 2025 ↗ - PAI-11Mostly true
“Lawmakers who studied the idea concluded that a pause on AI could not be made to work.”
Mostly right, but it was one group. In 2023, a committee of UK MPs said it was "unconvinced" a pause could be made to happen, and urged the government to speed up making rules for AI instead. That is a judgement, not a proven fact.
BCS, the UK's professional body for IT, warned it would be an "asymmetrical pause", meaning bad actors would ignore it and get ahead. On the other side, EU law now makes companies tell the European Commission within two weeks once they know a model's training will pass a set amount of computing power. The EU says big training runs take a lot of planning, so companies know this before training ends.
“We are unconvinced that such a pause is deliverable.”
House of Commons Science, Innovation and Technology Committee, UK Parliament·The governance of artificial intelligence: interim report (Ninth Report of Session 2022–23)Government / parliament · 31 Aug 2023 ↗Further sources for this card
- PAI-12True
“The EU already has legal rules for the most powerful AI models.”
Confirmed: since 2 August 2025, the EU AI Act has set rules for companies that make general-purpose AI models (AIs that can do many different kinds of tasks). The most powerful ones, which the law says carry "systemic risk", have extra duties.
A model counts as systemic risk if training it used more than 10^25 computing operations (a 1 followed by 25 zeros). Training a model that big is thought to cost tens of millions of euros. The EU can add other models too. Only a handful of companies make such models today. They must test for risks, reduce them, report serious incidents and keep the model secure. Fines can reach 3% of yearly worldwide sales or €15 million, whichever is higher. These rules do not pause or ban training.
“Under the AI Act, providers of general-purpose AI models with systemic risk have additional obligations.”
European Commission·General-Purpose AI Models in the AI Act – Questions & AnswersOfficial statistics · 9 Sept 2025 ↗ - PAI-13Mostly false
“AI companies' safety guardrails already stop people from misusing their AI models.”
Not reliably. The UK's AI Security Institute found ways around the safety guardrails of every AI system it tested. The best guardrails are getting harder to beat, though.
In one test, an expert got past a leading model's guardrails in 10 minutes. For a model released six months later, it took more than 7 hours, about 40 times the effort. That was on one of the best-protected topics. Progress is uneven: some systems still fall for simple, well-known tricks, and the guardrails on open models that anyone can download can be removed cheaply.
“we’ve managed to find vulnerabilities in every system we’ve tested”
UK AI Security Institute·Frontier AI Trends ReportGovernment / parliament · 18 Dec 2025 ↗
The judge is an AI model. It can only rule on claims these cards cover. Everything else stays unverified, and it never rules on opinions. See every motion's cards.