AI Camps: a calm map of the debate
Before you read. If you have seen headlines saying AI will save the world, or end it, you may have wondered who to believe. This series doesn't tell you what to think. It introduces the main schools of thought, quotes their leading voices accurately, and shows where they agree, where they differ, and what evidence would settle the question. Nobody here is a villain. Almost everyone in this debate wants the same thing, a good future, and disagrees about how to get there.
Why a map helps
AI discussions can feel like a shouting match. That is partly because people are answering different questions without noticing. Three questions sit underneath most disagreements:
- How fast will AI become much more capable? Estimates range from "within a few years" to "not this century", or "the question is poorly posed".
- If it becomes very capable, how hard will it be to keep it safe and beneficial? Views range from "manageable, like other technologies" to "extremely hard, and we may get only one chance".
- What should we do now? Answers range from "build faster, because delay costs lives" to "slow down or stop until we understand it".
Most "camps" are simply different combinations of answers to these three questions.
The camps at a glance
| Camp | In one sentence | Typical view on speed | Typical view on danger | Typical policy stance | Read more |
|---|---|---|---|---|---|
| Accelerationists (including "e/acc" and techno-optimists) | Technology, including AI, is the main engine of human progress, and slowing it causes real harm. | Fast, and good if fast | Manageable. Fear itself is the bigger risk | Build, compete, deregulate, expand energy | Accelerationists |
| Those most worried about catastrophe (often called "doomers") | Building AI much smarter than us, with today's methods, is likely to end in disaster. | Possibly fast | Very high | Stop or pause frontier development through international agreement | Doomers and p(doom) |
| Safety-focused | AI could be enormously beneficial and seriously dangerous, so safety science and governance must keep pace. | Uncertain, possibly fast | Real and worth significant effort to reduce | Invest in safety research, testing, transparency and oversight; build carefully | Safety-focused |
| Sceptics (of hype, of doom, or of both) | Today's AI is impressive but limited, and both utopian and apocalyptic claims run ahead of the evidence. | Slower than claimed, or unclear | Speculative long-term risk; focus on present harms | Address concrete harms now (bias, misinformation, concentration, labour) | Sceptics |
These are positions, not tribes. Many people hold a mix. For example:
- Dario Amodei of Anthropic writes enthusiastically about AI's benefits (Machines of Loving Grace, Oct 2024), yet puts a 25% chance on things going "really, really badly" (Axios, 17 Sep 2025).
- Gary Marcus, one of the best-known sceptics of current methods, wrote in 2022 that newer hybrid approaches gave him "some optimism about AI" for the first time in 40 years (Nautilus, 10 Mar 2022).
Three big ideas the camps argue about
- The singularity, explained. Where the idea of a runaway "intelligence explosion" came from (I. J. Good, Vernor Vinge, Ray Kurzweil) and why many researchers doubt it.
- Abundance. The hopeful case that AI could help make health, knowledge, energy and opportunity far more widely available, and the conditions that would have to be met.
- Risks and the conditions they need. For each major risk, what would have to happen for it to become real, and what early signs would tell us.
Where most people actually are
Loud voices sit at the edges, but the evidence suggests most experts are somewhere in the middle:
- In the largest survey of AI researchers (2,778 respondents, late 2023), about 68% thought good outcomes from advanced AI were more likely than bad ones. The median respondent still put about a 5% chance on "extremely bad outcomes (e.g. human extinction)" (Grace et al., JAIR, 2025).
- The International AI Safety Report 2026, written with input from experts nominated by more than 30 countries and organisations, says AI progress to 2030 could plateau, continue or accelerate. It also says current systems "lack the capabilities" to pose loss-of-control risks, although "they are improving in relevant areas" (executive summary, 3 Feb 2026).
In other words, the typical expert position is hopeful, uncertain, and in favour of taking safety seriously. That is also the stance of this platform.
How to read this series
- Quotes are exact and linked. When we paraphrase, we say so. When we could only confirm something through a secondary report, we say that too.
- Predictions are tracked. Where a camp makes a claim that can be checked over time, we link to its entry in our public tracker, Where are we now?. There you can see its current status and the evidence behind it.
- Numbers are opinions unless stated otherwise. A forecast of "10%" or "2029" is someone's judgement, not a measurement. We explain how to weigh such numbers in Doomers and p(doom).
What this means for you
You don't need to pick a camp. A reasonable, well-informed person can believe all of the following at once:
- AI is already useful.
- It may become far more capable.
- Some risks deserve serious work.
- Many confident predictions, hopeful or fearful, will turn out wrong.
Holding that balance isn't fence-sitting. It is where the evidence currently points.