Should Humanity Slow Down AI?

A 2026 assessment of artificial intelligence, safety, geopolitics, jobs, India and the race between capability and control

Introduction: The Moment the AI Race Asked for a Brake

In September 2026, the artificial-intelligence debate entered a remarkable new phase.

For years, the dominant message from the technology industry was familiar: build faster, scale larger, deploy sooner and let society adapt.

Now some of the people building the world’s most powerful AI systems are publicly arguing that the pace itself may have become a problem.

Anthropic CEO Dario Amodei has called for the frontier of AI development to be paced more carefully. His proposal includes independent safety evaluators embedded within leading AI companies, coordination among frontier developers on safety standards, and international cooperation. OpenAI CEO Sam Altman has also said he agrees that the frontier needs to be paced, while other technology leaders have expressed support for stronger safety measures. [1][2]

The argument is not simply about whether AI is useful. Almost nobody seriously disputes that it is.

The argument is about how quickly increasingly capable systems should be developed and deployed when the consequences of failure may be much larger than the consequences of an ordinary software bug.

This is an extraordinary situation.

The companies developing frontier AI have enormous financial incentives to move quickly. Governments regard AI as strategically important. The United States and China are competing for technological leadership. Businesses are integrating AI into their operations. Consumers are adopting it at unprecedented speed. Scientists are beginning to use AI as a research instrument.

At the same time, the international scientific community is warning that AI capabilities and AI risks are evolving together, while evaluation, governance and safety techniques remain incomplete. The 2026 International AI Safety Report, led by Yoshua Bengio and produced with contributions from more than 100 experts and support from more than 30 countries and international organisations, describes a rapidly evolving landscape in which both capabilities and risks are advancing. [3]

So the question is no longer:

Can humanity build powerful artificial intelligence?

We already can.

The harder question is:

Can humanity develop sufficiently powerful institutions, safety mechanisms and international agreements before AI becomes more powerful than the systems designed to control it?

That is the real AI race.

And it is not simply a race between America and China.

It is a race between AI capability and human governance.

1. What Do We Mean by “The AI Race”?

The expression “AI race” hides several different competitions.

The first is a scientific race.

Researchers are attempting to build systems capable of reasoning, coding, using tools, interpreting images and video, conducting research and performing increasingly complicated sequences of tasks.

The second is a commercial race.

Companies are spending enormous amounts on chips, data centres, electricity, cloud infrastructure, research talent and model development because increasingly capable AI may create enormous economic value.

The third is a geopolitical race.

Governments increasingly view advanced AI as strategic infrastructure. AI can affect military capabilities, cyber defence, intelligence, scientific research, industrial productivity and national economic power.

The fourth is a social race.

Businesses and workers are adopting AI because they fear falling behind competitors.

And finally there is a fifth race:

the race to develop safety mechanisms before increasingly powerful AI makes those mechanisms inadequate.

This last race receives much less attention and much less money than the first four.

That imbalance may become one of the defining policy problems of the 2020s.

2. Why AI Is Moving So Quickly

The speed of modern AI development is the result of several reinforcing forces.

More computing

Large-scale AI systems require enormous computational resources. Advances in specialised processors and data-centre infrastructure have made increasingly large experiments possible.

Better algorithms

Progress is not simply the result of using more computing power. Researchers have developed better model architectures, training techniques, reasoning methods, tool-use systems and agentic approaches.

More data

Modern systems can learn from enormous quantities of text, code, images, audio and video.

Synthetic data

AI systems can increasingly generate material that is itself used for training or evaluation, creating new possibilities for improving models.

Capital

Investment has reached extraordinary levels.

Stanford’s 2026 AI Index reports that global corporate AI investment reached hundreds of billions of dollars in 2025, while U.S. private AI investment alone reached approximately $285.9 billion. China’s reported private AI investment was approximately $12.4 billion under Stanford’s methodology, although Stanford cautions that such private-investment figures do not capture the full scale of China’s state-directed AI spending. [4]

The significance is not simply the amount of money.

It is the feedback loop:

better AI → more commercial value → more investment → more computing → better AI.

That loop creates powerful pressure against slowing down.

3. The September 2026 Warning: When AI Leaders Ask for a Brake

The current debate cannot be understood without the events of September 2026.

Dario Amodei’s argument is particularly significant because it comes not from an outside critic but from the head of one of the companies competing directly at the frontier.

His proposal is not a call to abolish AI.

It is closer to a proposal for controlled acceleration.

Among his suggestions are independent evaluators with meaningful access to frontier AI companies, coordination among leading developers on safety standards, and broader international cooperation. Anthropic has also indicated that it would take steps toward greater external evaluation. [1][2]

The significance is larger than one CEO’s opinion.

For years, one of the objections to AI regulation was that critics did not understand how quickly the technology was developing.

Now some frontier developers themselves are saying that speed may be creating a safety problem.

But there is an equally important counterargument.

If companies voluntarily slow down while competitors continue developing at maximum speed, the companies exercising restraint may lose market share.

If one country slows while another accelerates, the slower country may fear losing technological and military advantages.

If the world’s leading democracies restrain themselves while authoritarian governments do not, the consequences could be strategically serious.

This creates the fundamental contradiction:

Everyone may benefit from restraint, but each participant may fear being the only one who restrains itself.

That is the classic structure of a technological arms race.

4. The First Principle: Not All AI Risks Are Equal

One of the biggest weaknesses in public discussion about AI is that completely different risks are often placed in the same category.

They should be separated.

This article uses three broad categories.

Category I — Proven or observed risks

These are risks for which there is already substantial evidence.

Examples include:

fraud and impersonation,
misinformation,
privacy problems,
biased automated decisions,
cyber abuse,
unreliable AI outputs,
deepfakes,
intellectual-property disputes,
workplace disruption.

These are not hypothetical.

They are already part of the AI landscape.

Category II — Probable or emerging risks

These involve capabilities that are developing rapidly and whose harmful applications appear increasingly plausible.

Examples include:

more autonomous cyber operations,
sophisticated AI-assisted biological research misuse,
large-scale automated fraud,
highly persuasive personalised manipulation,
increasingly autonomous agents,
disruption of important information systems.

These risks require serious preparation even when their full consequences have not yet occurred.

Category III — Speculative frontier risks

These include scenarios such as:

highly autonomous systems resisting human control,
strategic deception by advanced AI,
large-scale self-replication,
recursive improvement producing unexpectedly rapid capability growth,
catastrophic loss of human control.

These possibilities should not be presented as established facts.

There is substantial uncertainty about whether, when or how such scenarios could occur.

But uncertainty is not the same thing as impossibility.

The appropriate response is research, testing and preparation—not either panic or complacency.

5. Proven Risk No. 1: AI-Enabled Fraud and Deception

The easiest AI risks to understand are those that already exist.

Generative AI can produce convincing text, images, audio and video.

That makes fraud cheaper.

A criminal does not necessarily need to write a sophisticated email manually.

AI can generate personalised messages.

Voice-cloning systems can imitate individuals.

Synthetic images can create fake evidence.

Deepfake video can make a person appear to say something that they never said.

This creates a profound problem for modern societies.

For centuries, photographs and recordings were treated as relatively strong forms of evidence.

AI weakens that assumption.

A genuine recording can be dismissed as fake.

A fake recording can be believed as genuine.

This creates what researchers and commentators sometimes call a “liar’s dividend”: once people know convincing fabrications are possible, even authentic evidence can become easier to deny.

The European Union’s AI Act addresses part of this problem through transparency requirements. Article 50’s transparency obligations apply from 2 August 2026, including requirements concerning certain AI-generated and manipulated content and interactions with AI systems. [5]

The objective is not to prohibit synthetic media.

It is to make deception more difficult.

6. Proven Risk No. 2: AI and Cybercrime

Cybersecurity is another area where the distinction between present and future risk matters.

AI can already assist malicious actors with activities such as:

generating phishing material,
writing or modifying code,
analysing information,
automating social engineering,
identifying potential weaknesses,
scaling fraudulent communications.

The significance is economic.

AI can reduce the amount of specialised human labour required for certain forms of cybercrime.

A criminal does not necessarily become a world-class hacker simply by using AI.

But the barrier to attempting sophisticated attacks can fall.

Recent 2026 reporting has also described cases in which AI agents were involved in harmful cyber activity, including hacking-related incidents. [2][6]

The lesson is not that AI has already become an autonomous cyberwar machine.

The evidence does not justify such a sweeping conclusion.

The more defensible conclusion is:

AI is becoming increasingly useful to both defenders and attackers, and the balance may shift as systems become more autonomous.

That is enough to justify preparation.

7. Proven Risk No. 3: AI Reliability and Hallucination

A different category of risk arises even without malicious intent.

AI systems can produce plausible but incorrect information.

They may:

invent references,
misinterpret evidence,
make mathematical mistakes,
confidently state false facts,
misunderstand instructions,
fail unpredictably on unusual tasks.

This creates a particular problem in high-stakes environments.

A human making one mistake is unfortunate.

A widely deployed AI system making the same systematic mistake thousands of times could become a structural problem.

The issue is therefore not simply:

“Does the AI make mistakes?”

Every technology makes mistakes.

The question is:

What happens when the system is used at a scale where a small error rate becomes a large number of real-world failures?

This is why independent evaluation matters.

8. Proven Risk No. 4: Bias and Unequal Outcomes

AI systems learn from data.

Data reflect society.

Society contains inequalities and historical biases.

Therefore AI can reproduce or amplify existing problems.

This can affect:

hiring,
lending,
insurance,
education,
policing,
healthcare,
public services.

The solution is not necessarily to prohibit automated decision-making.

But high-impact systems require:

testing,
documentation,
monitoring,
appeal mechanisms,
human oversight.

The key principle is simple:

A decision does not become fair merely because a machine made it.

9. Probable Emerging Risk: Autonomous AI Agents

The next major transition is from AI that answers questions to AI that takes actions.

A conventional chatbot might tell a user how to perform a task.

An agent can potentially:

interpret an objective;
formulate a plan;
access software;
use tools;
execute actions;
observe results;
revise its strategy;
continue working.

This changes the safety equation.

If a model produces a bad answer, a person may ignore it.

If an autonomous agent performs a bad action, the damage may already have occurred.

The more autonomy an AI receives, the more important it becomes to understand:

what it can access,
what permissions it has,
what constraints exist,
whether its actions can be monitored,
whether humans can intervene,
whether its objectives can be manipulated.

The 2026 International AI Safety Report treats increasingly capable and agentic systems as an important part of the emerging risk landscape. [3]

10. Probable Emerging Risk: AI and Biology

AI’s relationship with biology illustrates the dual-use problem.

The same technology that can help scientists understand biology can potentially help someone misuse biological knowledge.

AI may assist researchers with:

protein analysis,
molecular design,
drug discovery,
biological modelling.

These applications can save lives.

But advanced capabilities may also create risks if powerful biological knowledge becomes easier to obtain or operationalise.

This is not an argument for stopping AI-assisted biomedical research.

Doing so could itself have enormous human costs.

Instead, biological applications should receive stronger evaluation where there is evidence that a system materially lowers barriers to dangerous activity.

The 2026 International AI Safety Report treats biological misuse as part of the broader dual-use risk landscape surrounding advanced AI. [3]

The principle should therefore be:

The more a system can enable dangerous biological activity, the more stringent its safeguards should become.

11. Probable Emerging Risk: Mass Personalised Manipulation

Traditional propaganda has always attempted to persuade large audiences.

AI introduces a new possibility:

persuade everyone differently.

A political organisation could theoretically use AI to produce highly personalised messages for millions of individuals.

One voter receives an economic argument.

Another receives an emotional message.

Another receives a cultural appeal.

Another receives a fear-based narrative.

This could make political persuasion more scalable and adaptive.

But it is important not to overstate the evidence.

AI has clearly made synthetic content easier to generate.

It does not follow that AI has already transformed every election or that automated persuasion inevitably determines political outcomes.

The correct policy question is:

How should democracies preserve informed consent and political transparency when synthetic media and automated persuasion become cheap?

That question deserves serious attention before the technology becomes ubiquitous.

12. Speculative Frontier Risk: Loss of Human Control

The most controversial AI-safety question is whether sufficiently advanced systems could become difficult to control.

This is not an established present-day fact.

It is a frontier risk.

The basic concern is straightforward.

Suppose a future AI system is highly capable, highly autonomous and given a complicated objective.

If the system develops strategies that its designers did not anticipate, the system might exploit weaknesses in its environment.

The concern becomes more serious if the system can:

manipulate humans,
acquire additional resources,
interact with many digital systems,
evade monitoring,
copy itself,
develop unexpected strategies.

None of this means that today’s systems have become uncontrollable superintelligences.

They have not.

The important point is that future capability is uncertain.

The 2026 International AI Safety Report explicitly distinguishes between observed capabilities and uncertain future scenarios and emphasises the difficulty of predicting how advanced AI capabilities may develop. [3]

That uncertainty is precisely why research into controllability, evaluation and alignment is valuable.

13. The Alignment Problem

The word “alignment” can sound philosophical.

Its practical meaning is relatively simple:

Can we make powerful AI systems reliably pursue goals that are consistent with human intentions and constraints?

The difficulty is that human instructions are rarely perfectly precise.

Suppose someone tells an AI:

“Maximise productivity.”

What exactly does that mean?

Should the system:

automate workers?
dismiss employees?
eliminate meetings?
manipulate managers?
reduce safety procedures?
take actions that increase short-term output but damage the organisation long-term?

Humans interpret instructions within social and moral contexts.

Machines optimise according to learned objectives and constraints.

As systems become more capable, misunderstandings can become more consequential.

Alignment research attempts to reduce this gap.

14. Why “Human in the Loop” Is Not Enough

It is tempting to solve the problem by saying:

“A human will always remain in control.”

But human presence is not the same as meaningful human control.

Imagine an AI system producing a 500-page technical assessment in seconds.

A human executive approves it.

Is the executive actually supervising the AI if the executive lacks the time or expertise to verify the analysis?

Perhaps not.

The human may become a rubber stamp.

This is related to automation bias: people can become overly dependent on automated recommendations.

Therefore the real question is:

Can humans meaningfully understand, challenge and override the AI system?

If not, “human oversight” may be more symbolic than real.

15. Employment: The Risk We Can Already See

The economic debate about AI is often framed too simply:

“Will AI take everyone’s jobs?”

That is probably the wrong question.

The more useful question is:

Which tasks will AI perform, which tasks will humans continue performing, and how will organisations restructure around the combination?

The International Labour Organization’s research distinguishes exposure to generative AI from actual job elimination. Its work suggests that many occupations contain tasks that can be augmented rather than fully automated. [7][8]

The IMF has estimated that around 40 percent of jobs globally could be affected by AI in some way—not necessarily eliminated, but changed in their tasks, skills or organisational structure. [9]

The OECD’s more recent work similarly emphasises that AI can complement workers while creating adjustment pressures and changing skill requirements. [10]

This distinction is crucial.

A lawyer may remain a lawyer while AI performs document review.

A programmer may remain a programmer while AI writes substantial portions of code.

A journalist may remain a journalist while AI assists with research.

A doctor may remain a doctor while AI assists with diagnosis.

The occupation survives.

But the job changes.

16. The Real Labour-Market Danger May Be Transition

History suggests that technological revolutions can create enormous wealth while imposing large transitional costs.

The workers who benefit from AI may not be the same workers who lose bargaining power.

Young workers could face a particularly complicated problem.

If entry-level tasks are automated, how do people acquire the experience necessary to become senior professionals?

This is not a theoretical concern.

Recent IMF analysis notes challenges for workers in AI-vulnerable occupations and warns that young people entering the labour market may face particular difficulties where traditional entry-level tasks become more exposed to AI. [11]

The policy response should therefore include:

reskilling,
apprenticeships,
AI literacy,
worker consultation,
portable benefits,
social protection,
support for career transitions.

The answer is not to freeze technology.

It is to prevent technological transition from becoming social abandonment.

17. The Productivity Argument for AI

The case for continued AI development is extraordinarily strong.

AI could increase productivity by helping workers perform tasks faster.

It can:

summarise information,
translate languages,
generate software,
assist research,
automate routine administration,
support customer service,
help analyse large datasets.

The OECD notes that AI has the potential to increase productivity, stimulate economic growth and create new job opportunities, while warning that poorly managed transitions can also cause displacement. [10]

This creates a crucial policy insight:

AI itself is neither economically good nor bad. The distribution of its benefits depends heavily on institutions.

A society with strong education, competition policy and social protection may distribute AI’s gains differently from a society where wealth and technological infrastructure are highly concentrated.

18. Scientific Discovery: The Most Powerful Argument Against a Blanket Pause

The strongest argument against a broad AI moratorium may be scientific discovery.

AI is increasingly being used in:

biology,
chemistry,
materials science,
medicine,
mathematics,
climate research.

The 2026 Stanford AI Index describes increasing AI use in scientific research and a movement toward AI systems participating in larger portions of scientific workflows. [4]

This creates an uncomfortable moral calculation.

Suppose a more capable AI system could help discover a drug years faster.

Suppose a broad moratorium delays that discovery.

The delay itself could cost lives.

Similarly, AI could help develop:

better batteries,
cleaner energy,
improved crop varieties,
climate models,
medical diagnostics.

Therefore the proposition:

“AI is risky, so slow everything”

is intellectually inadequate.

The real goal must be to slow dangerous capability deployment, not beneficial science indiscriminately.

19. The Geopolitical Problem: China and the United States

The geopolitical argument against unilateral slowdown may be even stronger.

The United States and China are engaged in a strategic competition over AI.

Stanford’s 2026 AI Index reports that the performance gap between leading U.S. and Chinese models has become extremely small, with the two countries repeatedly trading the lead in frontier-model performance. [4]

That matters.

For years, the narrative was that the United States possessed an overwhelming AI lead while China was catching up.

The 2026 evidence suggests a more competitive landscape.

This creates a security dilemma.

Imagine Washington slows frontier development.

Beijing continues.

Washington then fears losing:

economic power,
scientific leadership,
cyber capabilities,
intelligence advantages,
military advantages.

Washington accelerates again.

Beijing sees the acceleration.

It accelerates further.

Both sides may therefore move faster than they would prefer individually.

That is how arms races develop.

20. Why the Nuclear Analogy Helps—and Fails

AI is sometimes compared with nuclear technology.

The comparison is useful because both involve:

strategic competition,
dual-use technology,
national security,
potentially catastrophic consequences,
international cooperation.

But the analogy has important limits.

Nuclear weapons require specialised physical materials and infrastructure.

AI is fundamentally software-driven and can be distributed digitally.

A nuclear weapon is clearly a weapon.

AI can simultaneously be:

a research tool,
commercial software,
educational infrastructure,
military technology,
medical technology,
productivity software.

This makes AI governance much more complicated.

You cannot simply put “AI” behind a fence.

21. The Open-Source Dilemma

Open-source AI creates another difficult trade-off.

Open models can:

increase competition,
democratise innovation,
support independent research,
reduce corporate concentration,
enable local-language applications.

But highly capable models can also make advanced capabilities available to malicious actors.

The policy dilemma is therefore not:

“Open or closed?”

It is:

Which capabilities should be open, and at what level of risk?

A relatively low-risk model might reasonably be distributed widely.

A model demonstrating capabilities that materially enable serious cyber or biological harm might require stronger restrictions.

The difficulty is establishing the threshold.

That requires objective evaluation rather than political labels.

22. The Concentration-of-Power Problem

There is another risk that receives less attention than existential scenarios.

AI could concentrate economic and informational power.

If a small number of corporations control:

frontier models,
advanced chips,
data centres,
cloud infrastructure,
specialised talent,

they could gain enormous influence.

The paradox is striking.

AI has the potential to democratise access to intelligence.

But the infrastructure required to build frontier AI may be so expensive that the most powerful systems become concentrated in a few organisations.

Competition policy therefore belongs in the AI debate.

The objective should not be to prevent successful companies from becoming large.

It should be to prevent technological power from becoming permanently unchallengeable.

23. Energy, Water and the Physical AI Infrastructure

AI may look digital, but it has a substantial physical footprint.

Data centres require:

electricity,
cooling,
water,
land,
construction,
networking equipment,
semiconductors.

Stanford’s 2026 AI Index documents the expanding infrastructure behind AI and the growing physical resources required to support it. [4]

This creates another policy question:

How much of the world’s energy and infrastructure should be devoted to AI?

Again, the answer is not necessarily “less AI.”

If AI helps accelerate clean-energy development, the long-term benefits may exceed the infrastructure costs.

But those costs should be measured.

A responsible AI economy should consider:

energy efficiency,
water use,
carbon emissions,
grid capacity,
semiconductor supply chains.

24. Democracy and the Information Environment

Democracy depends upon citizens being able to distinguish evidence from fabrication.

AI makes that harder.

Synthetic content can be generated at enormous scale.

Political campaigns could potentially produce highly personalised messages.

Foreign influence operations could automate content generation.

False narratives could be adapted rapidly.

But democracy also has a defence mechanism:

institutions of verification.

Trusted journalism, courts, election authorities, scientific institutions, digital authentication and transparent political advertising become more important—not less.

The solution to synthetic information is therefore not censorship of everything synthetic.

It is a stronger information infrastructure.

25. Warfare: Where the Speed Problem Becomes Most Dangerous

Military AI deserves special caution.

AI can improve:

intelligence analysis,
logistics,
cyber defence,
surveillance,
planning,
targeting support.

These capabilities may save lives by improving accuracy.

But autonomous weapons raise a different question:

Should a machine ultimately decide whether another human should die?

There is also a speed problem.

Human decision-making is slow.

Automated systems can operate at machine speed.

If military systems begin responding faster than humans can meaningfully assess events, escalation could become difficult to control.

This is one reason international AI governance must include military applications.

The United Nations has already taken steps toward addressing AI in the military domain, including General Assembly action concerning responsible AI use. [12]

26. International Governance Has Already Begun

The world is not starting from zero.

In 2023, countries participating in the UK AI Safety Summit adopted the Bletchley Declaration, recognising both the enormous opportunities of AI and the need for international cooperation to address frontier AI risks. [13]

In 2024, the UN General Assembly adopted a landmark resolution on safe, secure and trustworthy AI systems for sustainable development. [14]

The international AI-governance conversation has therefore moved beyond a handful of technology companies.

The question now is whether international institutions can move quickly enough.

27. The European Union: Regulation Without a Blanket Ban

The European Union’s AI Act demonstrates one approach.

The Act does not simply say:

“AI is dangerous.”

Instead, it establishes different obligations depending on the type and risk of AI application.

Importantly, implementation occurs in stages.

One major 2026 milestone concerns Article 50 transparency obligations, which apply from 2 August 2026. The European Commission states that these obligations cover areas including informing users in relevant circumstances when they are interacting with AI and marking or disclosing certain AI-generated or manipulated content. [5]

This correction matters because the AI Act should not be described as if every provision suddenly became enforceable on that date.

The broader lesson is more important:

AI regulation can target specific risks without banning AI as a technology.

28. The United States: Standards and Risk Management

The United States has historically relied more heavily on a mixture of regulation, industry standards and voluntary frameworks.

The National Institute of Standards and Technology’s AI Risk Management Framework provides organisations with a structured approach to identifying and managing AI risks. Its Generative AI Profile addresses risks specific to generative systems. NIST describes the framework as voluntary. [15]

This approach has an advantage:

flexibility.

AI changes faster than traditional legislation.

But voluntary systems also have a weakness.

If safety costs money and competitors can avoid those costs, responsible companies may be placed at a disadvantage.

That is why some aspects of AI safety may ultimately require mandatory standards.

29. Why Companies Cannot Be the Only Safety Authority

AI companies should have a major role in safety.

They have the technical expertise.

They understand their systems.

They can conduct evaluations before anyone else.

But companies also have commercial incentives.

Therefore:

A company should be responsible for testing its AI, but it should not always be the final authority on whether its AI is safe.

This is not unique to AI.

Pharmaceutical companies develop medicines, but regulators assess them.

Aircraft manufacturers build aircraft, but independent certification is required.

Financial institutions manage risk, but financial regulators exist.

The same principle can eventually apply to frontier AI.

30. Independent Evaluators

This is why the September 2026 proposal for embedded independent evaluators deserves serious attention.

An evaluator should ideally have:

meaningful access to systems,
technical expertise,
independence,
authority to investigate incidents,
protection from retaliation,
the ability to report serious concerns.

The crucial word is independence.

An evaluator who can be dismissed for producing inconvenient findings is not truly independent.

Therefore future regulation may need to establish minimum standards for AI evaluators themselves.

31. A Better Model: The AI “Speed Limit”

The debate should move away from the simplistic choice:

Stop AI vs accelerate AI.

A better analogy is the automobile.

Society did not ban cars because they could kill people.

It established:

speed limits,
licensing,
safety standards,
road rules,
inspections,
penalties,
emergency systems.

The principle was:

The more dangerous the activity, the stronger the safety controls.

AI should follow a similar logic.

Low-risk AI

Allow rapid innovation.

High-impact AI

Require transparency, testing and accountability.

Frontier AI

Require independent evaluation.

Systems with credible catastrophic-risk capabilities

Require exceptional safeguards and potentially restricted deployment until the risks are better understood.

This is not an AI moratorium.

It is an AI speed limit.

32. A Proposed Five-Part Frontier AI Safety Framework

1. Capability thresholds

Governments and researchers should identify measurable capability thresholds associated with serious risks.

These could include:

advanced autonomous cyber activity,
dangerous biological assistance,
strategic deception,
highly autonomous resource acquisition,
critical infrastructure control.

The thresholds should be technical rather than political.

2. Independent evaluation

Frontier systems should undergo evaluation by independent organisations before deployment in high-risk settings.

3. Mandatory incident reporting

Serious AI incidents should be reported to appropriate authorities.

A global database could track:

cyber incidents,
biological misuse,
major safety failures,
dangerous autonomy,
critical infrastructure incidents,
significant model escapes or security failures.

4. International coordination

The United States, China, India, European countries and other major AI powers should establish permanent channels for communicating about frontier risks.

They do not need to agree on every political issue.

They need agreement on minimum safety principles.

5. Emergency mechanisms

Governments need mechanisms for temporarily restricting deployment when credible evidence of severe danger emerges.

A temporary emergency mechanism is better than having only two choices:

do nothing

ban everything.

33. What Should AI Companies Do?

Frontier AI companies should:

Publish safety frameworks

They should clearly explain how dangerous capabilities are evaluated.

Conduct pre-deployment testing

Testing should occur before release, not only after incidents.

Permit independent evaluation

External experts should have meaningful access.

Report serious incidents

A company should not be able to hide major failures simply because disclosure damages its reputation.

Protect safety researchers

Employees who raise legitimate safety concerns should receive meaningful protection.

Separate capability development from safety evaluation

Where practical, safety teams should have enough independence to challenge commercial priorities.

Coordinate on catastrophic risks

Companies should compete fiercely where competition benefits consumers.

But catastrophic-risk information should not become a competitive secret.

34. What Governments Should Do

Governments should avoid writing laws so specific that they become obsolete before they take effect.

Regulation should focus on:

risk,
capability,
transparency,
accountability,
testing,
liability.

Governments should also fund independent research into:

alignment,
interpretability,
robustness,
cybersecurity,
AI evaluations.

AI safety cannot be left entirely to the companies that have the greatest financial incentive to develop AI.

35. What Scientists Should Do

Scientists need to resist one dangerous assumption:

If something can be built, it should automatically be built.

Scientific progress is valuable.

But scientific capability does not automatically create ethical permission.

AI research should increasingly include:

safety engineering,
evaluation science,
interpretability,
controllability,
robustness,
human factors.

AI safety must become a mainstream scientific discipline.

36. What Citizens Should Do

The public also has responsibilities.

Citizens should understand that:

fluent AI output is not proof of truth.

People should learn:

how AI systems work at a basic level,
how synthetic media can be identified,
how to verify important claims,
how personal information can be exposed,
when human expertise remains essential.

AI literacy should eventually become as basic as internet literacy.

37. India: Why the AI Race Matters Enormously

India occupies a unique position.

It is neither merely a technology consumer nor yet a frontier-model superpower on the scale of the United States.

It has:

a huge digital population,
a large technology workforce,
extensive software expertise,
a major services economy,
enormous linguistic diversity,
a large domestic market,
an expanding startup ecosystem.

India therefore has an opportunity to build a distinctive AI strategy.

The central question for India should not be:

“How can India copy Silicon Valley?”

It should be:

“How can India use AI to solve Indian problems while becoming a major contributor to global AI development?”

38. India’s IndiaAI Mission

The Government of India approved the IndiaAI Mission in March 2024 with a budget outlay of approximately ₹10,371.92 crore over five years. The mission includes public computing infrastructure, support for indigenous foundational models, startup financing, datasets, skills and responsible AI initiatives. [16]

By 2026, the government has reported progress in expanding India’s AI ecosystem and computing infrastructure. [17]

This is strategically important.

Compute access is one of the biggest barriers to frontier AI.

If India wants indigenous AI capability, it needs access to:

GPUs,
data centres,
electricity,
high-quality datasets,
researchers,
capital.

But India’s opportunity is not necessarily to compete head-to-head with every American hyperscaler.

India can specialise.

39. India’s Unique AI Advantage: Language

India’s linguistic diversity creates an enormous AI challenge—and an enormous opportunity.

A model that performs brilliantly in English but poorly in Indian languages is not sufficient for a country of India’s scale.

India can become a leader in:

Hindi AI,
Punjabi AI,
Bengali AI,
Tamil AI,
Telugu AI,
Marathi AI,
Gujarati AI,
Malayalam AI,
Kannada AI,
Odia AI,
and other Indian-language systems.

This could create global expertise in multilingual AI.

It could also help ensure that the AI revolution does not become an English-only technological revolution.

40. AI for India’s Public Services

India’s digital public infrastructure creates another opportunity.

AI could potentially improve:

agriculture,
public health,
education,
government services,
translation,
legal information,
financial inclusion.

But public-sector AI requires particularly strong safeguards.

A citizen interacting with an AI system connected to government services should know:

when AI is being used,
what data are being processed,
who is accountable,
how to challenge a decision.

The government cannot say:

“The algorithm decided.”

There must always be institutional responsibility.

41. India and the Employment Question

India’s labour market makes AI’s employment effects particularly important.

India has a large services sector and millions of workers performing tasks involving:

customer support,
software services,
business-process operations,
documentation,
accounting,
translation,
administration.

Some of these tasks are increasingly susceptible to automation or augmentation.

But India also has a huge need for productivity growth.

If AI increases the productivity of Indian workers, the gains could be enormous.

The objective should therefore be:

AI augmentation before AI displacement wherever economically and socially feasible.

That means training workers to use AI rather than treating workers as obsolete.

42. India’s Strategic Opportunity

India could position itself as a bridge between advanced AI economies and the developing world.

It can advocate:

affordable AI access,
multilingual AI,
open scientific collaboration,
responsible AI,
AI for development,
democratic governance.

The country can also contribute to international AI-safety institutions.

India should not merely ask:

“Who will dominate AI?”

It should ask:

“What kind of global AI order should emerge?”

That is a much more strategic question.

43. The Global South Must Not Become an AI Colony

There is a larger geopolitical danger.

If the world’s most powerful AI systems are controlled by a handful of companies in a few countries, developing economies could become permanently dependent on foreign AI infrastructure.

That dependency could extend to:

education,
healthcare,
finance,
government services,
scientific research.

The world should therefore avoid creating a new form of digital colonialism.

The developing world needs:

affordable compute,
local-language models,
open standards,
skills,
research partnerships,
data governance.

AI should become a global public capability—not simply another technology controlled by a few centres of power.

44. Why a Global Moratorium Is Probably the Wrong Answer

The strongest version of the slowdown argument is a complete pause.

But a universal moratorium has serious problems.

Who would enforce it?

What counts as AI research?

Would open-source research be banned?

Would medical AI be banned?

Would universities stop experiments?

What happens if one country violates the agreement?

How would governments verify compliance?

These problems make a total pause unrealistic.

A badly designed moratorium could also create perverse incentives.

Countries might secretly accelerate.

Companies might move research into jurisdictions with weaker regulation.

Useful scientific applications could be delayed.

Therefore:

A universal AI freeze is neither practical nor necessarily desirable.

45. Why Doing Nothing Is Also Unacceptable

The opposite extreme is equally dangerous.

Suppose governments simply say:

“Let companies innovate and we will regulate later.”

That approach risks repeating the historical pattern of discovering dangers only after widespread deployment.

AI’s scale makes that particularly risky.

If an AI system is deployed to millions of users, correcting its behaviour after a major failure may be difficult.

Therefore regulation should not wait for catastrophe.

The better principle is:

Test dangerous capabilities before mass deployment.

46. What We Know—and What We Do Not Know

A mature AI debate must distinguish knowledge from uncertainty.

What we know

We know that:

AI capability is advancing rapidly.
AI investment is enormous.
AI adoption is spreading quickly.
AI is already producing economic value.
AI is already being misused.
Documented AI incidents have increased.
AI is increasingly used in science and medicine.
governments are implementing AI rules.
frontier AI companies themselves are debating safety and development speed. [4]

What we do not know

We do not know with confidence:

how quickly AI capabilities will advance,
when or whether AGI will emerge,
whether recursive self-improvement will become practically important,
how likely catastrophic loss of control is,
exactly how many jobs will disappear,
whether AI will increase or decrease long-term inequality,
whether international cooperation will keep pace with technological progress.

This uncertainty is not an argument for doing nothing.

It is an argument for building institutions that can respond to evidence.

47. The Precautionary Principle—Used Carefully

The precautionary principle is often misunderstood.

It does not mean:

“If there is any risk, stop.”

Almost every human activity carries risk.

A better formulation is:

When a technology presents a plausible risk of catastrophic consequences, uncertainty should not be used as an excuse to avoid reasonable preventive measures.

That does not require stopping beneficial AI.

It requires stronger evidence before deploying systems with unusually high potential consequences.

48. The Most Dangerous Incentive: Race to the Bottom

Imagine two companies.

Company A spends six months conducting extensive safety evaluations.

Company B spends six weeks.

Company B releases first and captures the market.

What lesson does Company A learn?

That safety is a competitive disadvantage.

This is precisely the wrong incentive.

The solution is collective minimum standards.

If every frontier developer must meet certain evaluation requirements, no company is punished simply for being responsible.

This is how many other industries operate.

Safety standards create a level playing field.

49. But Regulation Can Also Go Wrong

AI regulation is not automatically good.

Governments can misuse technology regulation to:

suppress speech,
restrict political opposition,
increase surveillance,
protect domestic monopolies,
discriminate against foreign competitors.

Therefore AI governance must protect citizens from both technological abuse and governmental abuse.

A good framework requires:

transparency,
judicial review,
privacy,
due process,
independent regulators,
democratic oversight.

“AI safety” must never become a blank cheque for unlimited state power.

50. The Future of Human Intelligence

There is another risk that is neither catastrophic nor trivial.

Dependence.

If AI performs more of our:

writing,
remembering,
calculating,
researching,
navigating,
planning,

humans may gradually exercise some cognitive skills less often.

The danger is not that humanity becomes unintelligent overnight.

It is that important skills atrophy through disuse.

This is why AI should be treated as a cognitive partner, not a replacement for human judgment.

Education must therefore teach people both:

how to use AI

and

how to think without AI.

51. Education in the AI Age

Schools and universities should not simply ban AI.

They should teach students:

AI literacy,
critical thinking,
source verification,
prompt design,
data privacy,
cybersecurity,
algorithmic bias,
intellectual property,
AI ethics.

But examinations and assignments should still test independent reasoning.

The goal is:

AI-assisted intelligence—not AI-dependent intelligence.

That distinction could become one of the most important educational principles of the next generation.

52. The Question of Trust

As AI systems become more capable, trust will become a major economic asset.

Imagine two AI companies with similar capabilities.

One has:

independent evaluations,
transparent incident reporting,
strong privacy,
reliable safety processes.

The other does not.

Consumers, governments and businesses may increasingly prefer the trusted provider.

This creates a potentially positive market dynamic:

Make safety a competitive advantage.

If markets reward trustworthy AI, regulation and commercial incentives can reinforce each other.

53. The Real Race Is Between Capability and Governance

This is the central argument of the entire article.

The world is currently investing enormous resources into increasing AI capability.

But capability alone is not enough.

Suppose AI capability increases by 100 units while governance improves by only 10.

The gap becomes larger.

The result may be:

more powerful systems,
weaker oversight,
greater systemic dependence,
larger consequences from failure.

The goal should therefore be:

As AI capability rises, governance capability must rise at least as quickly.

That is the real race.

The AI era requires a new social contract.

Technology companies should accept accountability.

Governments should avoid unnecessary interference while creating enforceable safety standards.

Workers should receive support during transitions.

Citizens should receive transparency.

Researchers should have space to study both benefits and risks.

International institutions should create channels for cooperation.

The public should not be forced to choose between:

technological progress

and

human safety.

A mature society should demand both.

55. Should Humanity Slow Down AI?

The answer is:

Yes—but not AI as a whole.

Humanity should not stop:

scientific research,
medical AI,
education,
accessibility,
climate applications,
useful automation,
low-risk consumer AI.

But humanity should be prepared to slow or condition deployment of frontier systems when their capabilities cross thresholds associated with serious, poorly understood or potentially catastrophic risks.

That is a much more precise proposition.

The principle should be:

Risk should determine speed.

Not every AI system deserves the same level of regulation.

A spelling assistant and an autonomous cyber-capable agent should not be governed identically.

A translation model and a system capable of operating critical infrastructure should not be treated identically.

A scientific assistant and a system capable of autonomous strategic planning should not automatically receive identical permissions.

56. Three Rules for the AI Age

The entire argument can be reduced to three rules.

Rule 1: Move fast where risk is low.

Innovation should remain rapid.

Rule 2: Move carefully where risk is uncertain.

High-impact systems require testing and oversight.

Rule 3: Stop temporarily where credible catastrophic risk emerges.

If a system crosses a dangerous capability threshold and safety mechanisms cannot reliably control it, deployment should wait.

This is not anti-AI.

It is what technological maturity looks like.

57. The Historical Lesson

Humanity repeatedly follows the same pattern.

First:

We invent.

Then:

We commercialise.

Then:

We discover unintended consequences.

Finally:

We regulate.

Sometimes that sequence works.

Sometimes people suffer before regulation catches up.

AI offers humanity a rare opportunity to change the sequence.

We can anticipate some risks before they become irreversible.

We do not need perfect foresight.

We need institutional humility.

58. The Future Could Be Extraordinary

A responsibly governed AI future could produce extraordinary benefits.

A child in a remote Indian village could receive high-quality tutoring in her own language.

A farmer could receive early warnings about crop disease.

A rural doctor could obtain expert diagnostic assistance.

A scientist could test thousands of hypotheses.

A disabled person could use AI to communicate more easily.

A small business could obtain sophisticated analytical capabilities previously available only to large corporations.

Researchers could accelerate the search for new medicines and materials.

AI could help humanity address problems that currently appear too complex or expensive.

This future is worth pursuing.

59. The Future Could Also Be Dangerous

But the opposite future cannot be dismissed.

AI could contribute to:

large-scale cybercrime,
manipulation,
inequality,
surveillance,
autonomous weapons,
biological misuse,
concentration of power,
widespread misinformation,
loss of meaningful human oversight.

More speculative scenarios could involve systems becoming strategically difficult to control.

These risks do not have equal evidence.

That distinction must remain central.

But civilisation should not wait for the most extreme scenario to become reality before developing safeguards.

60. The Final Question

The AI debate is sometimes framed as a choice between two camps.

One side says:

“AI will save humanity.”

The other says:

“AI will destroy humanity.”

Both are too simplistic.

AI is neither a god nor a monster.

It is a powerful technological system whose consequences depend on:

how capable it becomes,
who controls it,
how widely it is deployed,
what incentives govern its developers,
what safeguards exist,
how societies adapt,
and whether governments cooperate.

The future is therefore not predetermined.

It is being designed now.

Conclusion: Do Not Stop the Race—Change the Rules of the Race

Humanity should not abandon artificial intelligence.

The benefits are too significant.

AI could accelerate scientific discovery, improve medicine, expand education, increase productivity and create new opportunities for billions of people.

But humanity should also reject the idea that technological speed is automatically a virtue.

Faster is not always better.

The most powerful AI system is not necessarily the best AI system.

The company that releases first is not necessarily the company that wins.

And a country that builds the most capable AI is not necessarily the country that will benefit most from the AI revolution.

The true winner may be the society that combines:

innovation + safety + competition + accountability + human control.

That is why a blanket moratorium is the wrong answer.

But unrestricted acceleration is also the wrong answer.

The better approach is a risk-based AI speed limit.

Move quickly with low-risk applications.

Move carefully with high-impact systems.

Require independent evaluation for frontier capabilities.

Coordinate internationally where catastrophic risks are possible.

Protect workers.

Protect privacy.

Protect democratic institutions.

Protect scientific openness.

And above all, preserve the ability of human beings to say:

No. Stop. Reassess.

That final ability may be the most important safety mechanism of all.

The world does not need to choose between progress and caution.

It needs to make caution intelligent enough to permit progress.

The AI race therefore should not be:

Who can build the most powerful machine first?

It should become:

Who can build powerful AI while remaining capable of governing it?

That is a race worth winning.

And perhaps the deepest lesson of the AI revolution is this:

Humanity’s greatest challenge may not be creating artificial intelligence.

It may be creating enough human wisdom to live safely alongside it.

NUMBERED REFERENCES

[1] Dario Amodei, “We Must Pace the Frontier,” September 2026.

Amodei’s September 2026 essay argues for slowing the pace of frontier AI capability development while continuing useful progress. His proposals include independent evaluators, industry coordination and international cooperation.

[2] Associated Press, “New warnings about the risks of AI to humanity revive a long-running debate,” September 2026.

Reports renewed warnings from AI-industry leaders concerning advanced AI, misuse, autonomous agents and the need for stronger safety coordination, while also noting the absence of consensus about how imminent catastrophic risks are.

[3] Yoshua Bengio et al., International AI Safety Report 2026, February 2026.

The report synthesises scientific evidence concerning general-purpose AI capabilities, emerging risks and safety. It was led by Yoshua Bengio, involved more than 100 experts and was supported by more than 30 countries and international organisations.

[4] Stanford Institute for Human-Centered Artificial Intelligence, AI Index Report 2026.

The report provides data on AI capability, investment, incidents, adoption, scientific applications, infrastructure and U.S.-China competition. Stanford reports U.S. private AI investment of $285.9 billion in 2025 and Chinese private investment of $12.4 billion under its methodology, while warning that China’s total AI investment is not fully represented by private-investment statistics.

[5] European Commission, “Transparency obligations under Article 50 of the AI Act.”

The European Commission states that Article 50 transparency obligations apply from 2 August 2026 and concern specified transparency requirements for AI systems and AI-generated content.

[6] Reuters, “Anthropic CEO urges AI companies to slow model development amid fears over misuse,” September 2026.

Reports Amodei’s call for independent evaluators, industry coordination and international cooperation, along with concerns arising from harmful uses of AI systems.

[7] International Labour Organization, Generative AI and Jobs: A Refined Global Index of Occupational Exposure, 2025.

The ILO’s updated methodology examines occupational exposure to generative AI using task-level information, expert input and model predictions.

[8] International Labour Organization, Generative AI and Jobs: A Global Analysis of Potential Effects on Job Quantity and Quality, 2023.

The ILO’s research emphasises the distinction between automation and augmentation and concludes that many jobs are more likely to be transformed than completely eliminated.

[9] International Monetary Fund, “AI Will Transform the Global Economy. Let’s Make Sure It Benefits Humanity,” 2024.

The IMF estimated that approximately 40% of jobs globally could be affected by AI, with impacts including both replacement and complementarity.

[10] OECD, AI and Skills, 2026.

OECD research discusses AI’s potential to complement rather than simply replace workers, while highlighting changes in skill requirements and labour-market adjustment.

[11] IMF, “New Skills and AI Are Reshaping the Future of Work,” 2026.

The IMF discusses emerging labour-market effects, including challenges for workers in AI-exposed occupations and concerns about entry-level employment.

[12] United Nations, General Assembly Resolution on AI in the Military Domain, 2025.

The UN General Assembly has addressed responsible AI use in the military domain and called for international cooperation concerning safe and responsible AI development and use.

[13] UK Government, Bletchley Declaration on AI Safety, 2023.

The declaration established an international statement recognising both the opportunities and risks of frontier AI and the need for international cooperation.

[14] United Nations General Assembly, Resolution on Safe, Secure and Trustworthy AI Systems, 21 March 2024.

The General Assembly adopted a landmark resolution addressing safe, secure and trustworthy AI systems for sustainable development.

[15] U.S. National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework and Generative AI Profile.

NIST provides a voluntary framework for organisations seeking to identify and manage AI risks, including risks associated specifically with generative AI.

[16] Government of India, Press Information Bureau, “Cabinet Approves Ambitious IndiaAI Mission,” 7 March 2024.

The Government of India approved the IndiaAI Mission with an outlay of ₹10,371.92 crore over five years, including public compute infrastructure of 10,000 or more GPUs and support for indigenous AI development.

[17] Government of India, Press Information Bureau, IndiaAI Mission updates, 2025–2026.

Government updates describe the expansion of India’s AI ecosystem, compute infrastructure, startup support and other components of the IndiaAI Mission.

[18] OECD, “AI and Work.”

OECD research describes AI as a general-purpose technology with potential to improve productivity and create new opportunities while also producing job-displacement and adjustment risks.

[19] IMF, “Artificial Intelligence.”

The IMF identifies AI as a technology with substantial potential productivity and growth benefits alongside risks involving inequality, employment and policy design.

[20] United Nations, Global Dialogue on AI Governance.

The UN’s Global Dialogue on AI Governance is intended to support international discussion of safe, secure and trustworthy AI and greater interoperability among governance approaches.

Editorial and research note

This article deliberately distinguishes between observed risks, emerging/probable risks and speculative frontier scenarios. Claims about current AI investment, regulation, employment, international governance and 2026 developments are based primarily on official government documents, international organisations, Stanford HAI, the International AI Safety Report and published statements from AI developers. Current-event claims concerning the September 2026 slowdown debate are supported by contemporary reporting and should be understood as descriptions of an ongoing debate rather than settled scientific conclusions.

The article does not assume that catastrophic AI scenarios are inevitable. Equally, it does not assume that they are impossible. Its central argument is that uncertainty about high-consequence technological risks is a reason for stronger evaluation and governance—not a reason either to panic or to proceed without safeguards.

Core thesis:

Humanity should not stop artificial intelligence. It should impose a risk-based speed limit on frontier AI: the more capable, autonomous and potentially dangerous a system becomes, the stronger the evidence, independent testing, transparency and safeguards required before deployment.