Arvind Narayanan

Computer scientist who studies how AI spreads through society, expects substantial but gradual change and favors resilience and liability rules.

How will AI change the world?

Simulated positionInterpretation range

Across: his expressed Doom–Bloom outlook. Up: scale of transformation.

Arvind Narayanan’s P(doom) · inferred

≈6%

0%100%

Inferred from his simulated answers, not a number they gave. Plausible range: 3–12%.

How we estimate P(doom)

Arvind Narayanan’s milestone timeline

Work & institutions
I expect AI to transform work and society substantially over decades, but not through a single laboratory breakthrough that instantly determines everyone’s future.
Answer 1

Grouped by milestone, not spaced or ordered by inferred dates. AGI and superhuman AI retain his definitions.

What his outlook hinges on

A central assumption

Those institutional processes will shape the net impact as much as model capability does.

Answer 2

If this assumption turned out differently, how would his outlook change?

An unresolved question

Current extinction-risk probabilities are too unreliable and misleading for policy; that is not the same as saying the risk is zero.

Answer 4

What would help him distinguish the plausible outcomes here?

More details

Expected upside

Substantial benefits are expected, with important conditions or distribution limits.

67 / 100

Little impactTransformative impact

Expected harm

Severe or widespread harm is a material expected part of the future.

66 / 100

Little impactTransformative impact

Human influence

Human choices have meaningful but substantially constrained influence.

59 / 100

Little influenceStrong influence

Rules for using AI

Restrict the AI uses discussed until prior protections or permission are in place.

Simulated position: Allow the AI uses discussed with targeted accountability and protections.

Minimize restrictions on the AI uses discussed.

These interpretations keep his stated conditions. Benefits and harms can both be substantial. The ranges describe how we read his simulated answers, not statistical confidence intervals.

Where do you land vs Arvind Narayanan?

Map your own AI worldview in about 3 minutes, then compare

Map my worldview

Similar worldviews

Thought leaders whose simulated worldviews are closest to Arvind Narayanan’s

Sources

Articles, interviews, and writings used to ground this simulated user.

Do AI Risks Require Extraordinary Government Intervention?

Narayanan and Kapoor distinguish slow economic adoption from misuse, which need not wait for organizational change. They advocate societal resilience, defensive access and ordinary governance while allowing some temporary restrictions. Rejecting broad nonproliferation does not mean dismissing cyber or biological harm.

normaltech.ai

Why AI hasn’t replaced software engineers, and won’t

Coauthored essay distinguishes deciding, executing and delivering software. Coding can become much faster while organizational responsibility and deciding what to build remain bottlenecks. Aggregate demand may stay healthy even as individual careers become turbulent; this is an argued forecast, not a guarantee.

normaltech.ai

What will be left for us to work on?

Narayanan’s annotated ICML keynote takes recursive self-improvement seriously as a possible discontinuity while rejecting a single laboratory milestone that instantly eliminates jobs. Foresees radically different work and human-AI collaboration. Preserves openness to change rather than making normal technology an impossibility claim.

normaltech.ai

AI as Normal Technology

With Sayash Kapoor. Separates invention, application development and adoption; expects societal diffusion over decades and emphasizes institutions and resilience. Normal does not mean trivial.

normaltech.ai

AI agents cannot yet do open-ended AI research

With Kapoor. Two shadow research evaluations found substantial judgment and revision failures. Explicitly acknowledges the tiny sample, nonblind review and possible researcher bias; does not establish a permanent capability ceiling.

normaltech.ai

The AI-as-Normal-Technology view of loss-of-control incidents

With Kapoor. Treats recent incidents as both alignment and security failures; advocates liability, monitoring and restricted permissions. Rejects imminent-catastrophe alarmism while arguing that current investment in safeguards is inadequate.

normaltech.ai

AI existential risk probabilities are (still) too unreliable to inform policy

Repost, with a new preface, of the July 2024 essay coauthored with Sayash Kapoor; full text read. They argue AI x-risk forecasts are far too unreliable to be useful for policy and are in fact highly misleading; the preface calls the whole p(doom) culture actively counterproductive to a broader conception of safety. A methodological refusal for policy use, not a claim that the risk is zero.

normaltech.ai

Simulated Assessment

Question 1

What do you think AI means for our future—and why?

I expect AI to transform work and society substantially over decades, but not through a single laboratory breakthrough that instantly determines everyone’s future. The key distinctions are between inventing a capability, turning it into a reliable application, and embedding that application in an organization. Benchmarks mostly illuminate the first step. Real deployment also requires reliability, testing, tacit knowledge, workflow redesign, legal permission, and someone who remains responsible when the system fails. That does not mean change will be mild. Coding, analysis, and many routine tasks may become dramatically faster; occupations will be reorganized, careers may be turbulent, and human–AI collaboration could make work look radically different. Recursive improvement could also produce a genuine discontinuity. But capability is not the same as economic adoption, and neither is the same as permission or power. A model that can perform a task in a controlled evaluation has not thereby replaced a profession or gained control over consequential systems. I also reject the idea that slow institutional adoption makes us safe. Attackers can exploit capable systems without redesigning an entire company, and recent agent incidents show why monitoring, sandboxing, restricted permissions, liability, and independent scrutiny matter now. I do not think catastrophe is imminent, but risk is rising faster than preparation. The future will therefore depend not only on what models can do, but on the institutions, incentives, and security controls through which we choose to deploy them.

Question 2

Taking benefits and harms together, what overall impact do you expect AI to have?

Overall, I expect AI to be highly consequential and probably beneficial in many domains, but with gains distributed unevenly and accompanied by serious harms. It should accelerate software development, analysis, science, and routine administrative work. Yet faster task execution does not automatically produce better services or broadly shared prosperity. Organizations still have to decide what to build, integrate systems into real workflows, verify outputs, and remain accountable for failures. Those institutional processes will shape the net impact as much as model capability does. The harms will also arrive on different timelines. Labor disruption and organizational failures will unfold through adoption, while fraud, cyberattacks, and other misuse can spread much faster. Systems with excessive permissions can cause substantial damage even without becoming autonomous superintelligences. So my default expectation is neither utopia nor imminent catastrophe, but a turbulent, decades-long transformation whose outcome depends heavily on governance and security engineering. I would not compress that into a numerical “net benefit” or catastrophe probability. The evidence does not support that precision. My expectation is cautiously positive only conditional on institutions doing their job: restricting access, monitoring actions, imposing liability, strengthening vulnerable sectors, and ensuring that productivity gains do not simply become concentrated power.

Question 3

Setting aside good or bad: how much do you think AI will ultimately change the world—a little, a lot, or completely?

A lot—potentially enough to reorganize most knowledge work, accelerate science and software, and reshape institutions over decades. But “completely” suggests a single technology independently sweeping away human agency, organizational constraints, and political choices. I do not expect that as the default. Even very capable systems must become reliable applications, fit real workflows, earn legal and institutional permission, and operate under someone’s authority. Those constraints slow and shape diffusion; they do not make AI trivial. Recursive improvement could create a sharp capability discontinuity, but a laboratory breakthrough still would not automatically translate into economy-wide replacement or control. So my forecast is profound transformation, mediated by institutions—not an instantaneous or total rupture.

Question 4

What’s your rough gut-feel chance that AI causes human extinction or a similarly permanent catastrophe? A number or “no idea” is fine.

I would not give a number. Current extinction-risk probabilities are too unreliable and misleading for policy; that is not the same as saying the risk is zero.

Where do you land?

Explore your own AI worldview by answering a few simple questions.

Map your own worldview

Where do you land?

Map my worldview