AI Doom and AI Boosterism: Reading Both Critically
4 min read · updated August 3, 2026
Public argument about AI has organised itself into two camps whose members mostly read each other in order to disagree. There is a way to read both usefully, and it starts by conceding that each camp contains careful people making arguments the other camp is not engaging with.
The shape of the disagreement
The camps are not symmetrical opposites, which is the first thing lost in most coverage. Stated at their strongest:
The risk-focused position holds that systems whose capabilities are increasing on a trend nobody can currently forecast, whose internals are not interpretable, and whose behaviour is specified by an optimisation target rather than by a written rule set, warrant caution proportional to the uncertainty. Note that this argument runs on uncertainty rather than on confidence. Its strong form does not claim to know what happens; it claims that not knowing is itself the reason to move carefully.
The acceleration-focused position holds that the demonstrated benefits of the technology are concrete and near-term — in medicine, in education, in the cost of expert-level work — that regulatory friction reliably entrenches incumbents rather than reducing risk, and that harms from delay are real but never counted because they are statistical rather than visible. Note that this argument also runs on an asymmetry, just a different one.
Both of those are respectable. Neither is refuted by the other, because they are largely about different quantities: one about the tail of a distribution, one about the mean.
Incentives, applied evenly
The most common failure in this discussion is applying incentive analysis to one side only. Applied evenly, it is more interesting, because both camps contain the same structural pattern: the people most prominent in each are people whose position benefits from the technology being important.
| Position | Description |
|---|---|
| risk-focused | Institutional funding, staffing and attention flow to organisations that exist to study a serious risk. A claim of extreme capability is also, incidentally, a claim that the field one works in matters enormously. And a firm arguing that its own systems are dangerous is arguing that they are powerful, which is not an obviously costly thing to argue. |
| acceleration-focused | Capital, hiring and valuations follow the expectation of rapid capability gains. Opposition to regulation aligns with the commercial interest of firms that would bear its cost. And a claim that the risks are overstated is also a claim that no one need slow down. |
| commentators in both | Attention accrues to strong positions. A hedged, uncertain account is harder to publish and harder to share than a confident one in either direction, which selects for confidence independent of evidence. |
The row that gets omitted in most such analyses is the third, and it is the one that explains why the visible version of each position is more confident than the version its careful proponents hold.
Why incentives settle nothing
This section exists because the preceding one is routinely misused. Identifying that someone benefits from a belief does not bear on whether the belief is true. Epidemiologists benefit professionally from epidemics being serious; this tells you nothing about whether a given epidemic is serious. Fire-alarm manufacturers benefit from fires; fires are real.
Incentive analysis is useful for exactly one thing: predicting which errors a source is likely to make, so you know where to check. It tells you to look for selective emphasis in a direction, not to discount the argument. An argument is assessed on its premises and its inferences, and a person with every reason to want a conclusion may have arrived at it correctly.
Structural features worth noticing
- Unfalsifiable framings, on both sides. “The risk is by nature unprecedented, so absence of evidence is expected” and “every past technology panic was wrong, so this one is too” are the same move: a structure that absorbs any evidence. Neither is worthless, and neither can lose.
- Category conflation. Near-term measurable harms — bias in deployed systems, fraud, labour displacement — and speculative long-horizon risks are different claims with different evidence bases. Bundling them lets a strong argument about one carry a weak one about the other, in either direction.
- Timeline vagueness. A prediction without a date cannot be scored. Ask when, and ask what observation the author would count as having falsified it. Sources that answer both are a small minority and are worth substantially more than the rest.
- Selection over the same evidence. Both camps frequently cite the same results. When two accounts of one paper disagree, the paper is usually the shortest route to finding out which one selected.
How to read a piece from either side
Three questions, and they are the same three regardless of which direction the piece points.
- What would change the author’s mind? If nothing would, you are reading an identity rather than an argument.
- Which claim is doing the work? Most pieces contain one load-bearing empirical claim surrounded by framing. Find it and check that one thing.
- Is the strongest opposing argument stated? A piece that engages the best version of the other side is worth more than one that engages the worst, and the difference is visible in a paragraph.
The honest position available to a non-specialist is that the near-term effects are partly measurable and being measured, the long-horizon claims are not currently decidable, and confidence in either direction is a personality trait rather than a conclusion.