Simon Willison

Simon Willison

x.com/simonw

Software developer and blogger who tests LLMs and coding agents hands-on and writes about their uses and security risks such as prompt injection.

How will AI change the world?

Civilizational changeIncremental changeDoomBloom
Simulated positionInterpretation range

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

Doom–Bloom: 64 out of 100. Scale of transformation: 44 out of 100. Interpretation ranges: 50 to 75 horizontally, 39 to 52 vertically. These are interpretation coordinates, not event probabilities.

Simon Willison’s P(doom) · inferred

≈5%

0%100%

Inferred from their simulated answers, not a number they gave. Plausible range: 2–10%.

How we estimate P(doom)

What their outlook hinges on

A central assumption

Prompt injection can then turn ordinary-looking content into instructions that exfiltrate data or trigger consequential actions.
Answer 1

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

An unresolved question

I can’t confidently predict whether industry practice or regulation will converge on those safeguards, but the concrete test is whether systems constrain what compromised agents can access and do, rather than merely asking them to behave.
Answer 4

What would help them distinguish the plausible outcomes here?

More details

Expected upside

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

67 / 100

Little impactTransformative impact

Interpretation range 67 to 67 on the qualitative scale.

Expected harm

Manageable or localized harms are expected.

45 / 100

Little impactTransformative impact

Interpretation range 33 to 67 on the qualitative scale.

Human influence

Human choices have meaningful but substantially constrained influence.

53 / 100

Little influenceStrong influence

Interpretation range 37 to 88 on the qualitative scale.

Development pace

Stop or substantially slow development of more capable AI.

Simulated position: Continue development under stated safeguards.

Speed up development of more capable AI.

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

Where do you land vs Simon Willison?
Map your own AI worldview in about 3 minutes, then compare

Similar worldviews

Thought leaders whose simulated worldviews are closest to Simon Willison’s

Sources

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

Simulated Assessment

Question 1

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

I expect AI to produce genuinely useful tools, especially as capable models become faster and cheaper. Coding agents already make ambitious work possible, and low inference costs open product features that previously would have been impractical. Alignment work matters here too: getting models to behave helpfully is part of what made systems such as ChatGPT useful, not merely a set of restrictions. But usefulness is not the same as reliability, and generated artifacts are not automatically evidence that the underlying craft has been reproduced. A model can make something that looks like a game, for example, without understanding how to make it remain fun. Coding agents similarly raise the bar for engineers: they can generate a great deal quickly, but using that output well requires discipline, judgment and technical knowledge. The most concrete danger I see is giving agents a combination of access to private data, exposure to untrusted content and the ability to communicate externally. Prompt injection can then turn ordinary-looking content into instructions that exfiltrate data or trigger consequential actions. Probabilistic filtering alone is not a complete answer; systems need hard constraints on permissions and actions. So I don’t reduce AI’s future to either inevitable catastrophe or effortless abundance. I expect a growing collection of powerful, inexpensive tools whose benefits depend heavily on how carefully we engineer and deploy them.

Question 2

How much do you think AI will ultimately change the world?

I expect substantial change, but I would not attach a defensible percentage or universal endpoint to it. Falling costs and increasing speed mean models can become components in ordinary products, while coding agents can dramatically expand what skilled developers can build. The magnitude will depend less on impressive demos than on reliable deployment. Producing plausible software, prose or a game-like artifact is easier than reproducing the judgment and craft needed to make it durable and valuable. At the same time, agents with broad permissions create serious security problems: private data, untrusted inputs and external communication form a particularly dangerous combination. So I expect AI to reshape many kinds of work and software, but unevenly. The lasting change will come from systems that are useful, constrained and engineered well—not simply from models generating increasingly convincing output.

Question 3

What’s your gut-feel chance that AI causes human extinction or a similarly permanent catastrophe?

I don’t have a defensible percentage. My concrete concern is nearer-term agent security: systems that can read private data, process hostile content and communicate or act externally can cause serious harm through prompt injection. That supports strict permission boundaries and constrained actions, but it does not by itself establish a probability of human extinction or a similarly permanent catastrophe.

Question 4

How do you expect the people and institutions developing AI to respond to its risks?

I expect a mixed response. Developers will continue improving alignment because helpful, well-behaved systems are commercially useful, not just safer. Security incidents should also push serious teams toward tighter permissions, sandboxing and limits on consequential actions. But incentives will remain uneven. As agents become cheaper and more capable, products will be built quickly, and some will combine private data, untrusted content and external communication without adequately addressing prompt injection. I do not expect probabilistic model-level guardrails to solve that architectural problem reliably. Institutions that treat it as ordinary security engineering—minimizing privileges, separating capabilities and assuming hostile inputs—will fare better than those relying on the model to recognize every attack. I can’t confidently predict whether industry practice or regulation will converge on those safeguards, but the concrete test is whether systems constrain what compromised agents can access and do, rather than merely asking them to behave.
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