Everyone I recommend on AI is optimistic. Except Cory Doctorow.

Ernie Svenson · email · 2026-08-04 Reader 🖥️

AI Backlash AI Bubble Debate Centaur vs. Reverse Centaur Human-in-the-Loop Ironies of Automation VC Funding Incentive for Hype

Cory Doctorow deserves to be the one skeptical voice in an otherwise optimistic reading list because his critique is disciplined by tracing where money and accountability actually flow rather than by reflexive contrarianism, and that discipline yields two portable diagnostic tools — centaur vs. reverse centaur, and the accountability sink — for judging whether a given AI tool serves the person using it or the vendor selling it.

practical · mixed — Framed as a personal endorsement/recommendation (why to add Doctorow to your reading list), but the bulk of the piece is expository — explaining two of Doctorow's diagnostic tools (centaur/reverse-centaur, accountability sink) — in service of a practical payoff: a portable test readers can apply to their own AI tool choices.

Structure

Part	Role
Personal history with Doctorow (blogging since 2002, Boing Boing)	Establishes long-standing, firsthand familiarity rather than a reputation borrowed secondhand
"Why I trust him" — no financial stake, mechanism-tracing over prediction	States the explicit criteria by which Svenson judges Doctorow's commentary trustworthy
"Centaurs and reverse centaurs"	Introduces the first diagnostic tool: a question to ask of any AI tool
"The accountability sink" (radiologist example)	Introduces the second diagnostic tool and shows 'human in the loop' can function as liability insulation, not safeguard
"Why careful people still miss things" (Bainbridge 1983)	Explains the mechanism (vigilance decrement) behind why the accountability sink works even on diligent people
"Where to start" (pluralistic.net, new book, tone warning)	Converts the endorsement into an action — where and how to actually read Doctorow

Key terms

centaur / reverse centaur
A diagnostic pair for classifying any human-AI pairing: a centaur is a person who stays in charge and does better work with a tool's help; a reverse centaur is a person who serves the tool, scrambling to keep pace with software that sets the terms.
accountability sink
A structural arrangement in which a human formally 'in the loop' (e.g., signing off on an AI's output) absorbs legal/professional liability for a decision they did not really originate or fully control, while their presence functions as insulation for whoever sold or deployed the system.

Propositions

explicitNearly everyone Svenson recommends on AI is optimistic; Doctorow is the exception, arguing that most AI discourse is a sales pitch aimed at investors.conf: high
supportsPresentation: The Free-Lunch Guide to Idea Circularity · p6 — Doctorow's claim that most AI discourse is 'a sales pitch aimed at investors' is the same structural mechanism Cummins states directly: hype is a structural requirement of venture-funded business models, independent of the technology's actual merit.
explicitDoctorow's skepticism is not about predicting when things break but about tracing where money is flowing, because that's where the next problem starts.conf: high
explicitcontestedThe AI build-out is a bubble, in the sense that its viability is a function of money flow rather than a settled question of underlying value.conf: medium
contradictsAgents Over Bubbles · p1 — Doctorow's 'the AI build-out is a bubble' claim directly contradicts extract-agents-over-bubbles' thesis that 'AI infrastructure investment is justified demand, not a speculative bubble.'
refines ←The Questions You Should Be Asking About the AI Bubble · p13 — Splits Doctorow's single 'it's a bubble because viability = money-flow' claim into two separable questions — investor-level viability risk vs. macroeconomic transmission risk — arguing the latter is what actually determines economic damage.
explicitA centaur is a person helped by a machine who stays in charge and does better work than alone; a reverse centaur is a person serving a machine whose pace they must scramble to match.conf: high
explicitdevelopedWhether a given piece of software makes you a centaur or a reverse centaur is decided by how you use it, not by the software itself — the same tool can produce either outcome, which is testable by asking whether it lets you do better work or merely makes you keep up.conf: high
extends ←Can an AI-Powered Scribe Curb Physician Burnout? · p6 — Gallani's human-scribe anecdote is a concrete, pre-AI historical instance of the reverse-centaur dynamic Doctorow's framework describes: delegation whose promised time-saving gets reclaimed by the employer's productivity expectations rather than kept by the worker.
explicitcontestedIn an accountability sink, a human nominally 'in the loop' (e.g. a radiologist signing off on AI-flagged scans, or a lawyer signing a brief) absorbs the professional/legal liability for a call effectively made by the AI, while the vendor's terms of service ensure that liability lands on the human rather than the system's maker — the human functions as insulation for the vendor, not as a safeguard for the decision.conf: high
supportsWhen Employees Are Held Accountable for AI-Generated Decisions · p10 — Svenson's 'accountability sink' (radiologist/lawyer signs, liability transfers) is the same structural mismatch that proposition names directly: formal decision authority sits with the AI while social/legal accountability is transferred to the human, without transferring the epistemic access needed to bear it.
contradicts ←Can an AI-Powered Scribe Curb Physician Burnout? · p5 — Gallani frames physician sign-off as the safety measure that keeps AI outputs non-final; Doctorow's accountability-sink argument predicts that because the AI's outputs are usually correct, review vigilance structurally decays over time, undermining the very safety guarantee physician sign-off is meant to provide — the case never tests whether review stays alert once the tool becomes routine.
explicitWhen a system is right most of the time, the person checking it loses the ability to spot the times it isn't — attention drifts as a structural feature of the review task itself, not as a personal failing of the reviewer.conf: high
extendsPresentation: The Ironies of A^2 I^2 (Ironies of Automation and Artificial Intelligence in Incidents) · p11 — Svenson's 'attention drifts when the system is right most of the time' restates Reed's 'the more capable the AI, the poorer humans' compensatory behavior becomes — capability induces complacency,' and extends its application from software/incident-response operators to radiologists and lawyers.
implicit"I always review it" is not, on its own, an adequate safety plan, because the vigilance-decrement mechanism can let a fabricated citation slip past a careful lawyer, not only a sloppy one.conf: high
explicitcorroboratedThe attention-drift/vigilance-decrement problem is not new to AI — Lisanne Bainbridge described the same dynamic in a 1983 paper on industrial control rooms: automating routine work makes the human's residual job harder, not easier.conf: high
supports ←Human Oversight Is Not a Human in the Loop · p4 — Same vigilance-decrement/Bainbridge mechanism, here applied to AI-assisted work generally rather than industrial control rooms specifically.
implicitDoctorow's not being financially dependent on the AI industry's success (unlike many technology commentators) is itself part of what makes his skepticism trustworthy — a commentator's material interests shape which claims they are motivated to make.conf: medium
implicitBecause the AI build-out's viability is uncertain (a possible bubble), it is prudent not to anchor a core professional workflow entirely to a single AI vendor's platform.conf: medium

Arguments

Doctorow is worth trusting as the dissenting, skeptical voice on AI. strength: medium

Explicit premises:

He punctures hype and is rigorous.
He has no financial stake requiring him to soften his views (marked by his years at the EFF).
He traces the mechanism by which things could unravel rather than making date-bound predictions.
A good share of the time, the mechanisms he identifies do come unraveled.

Implicit premises:

A demonstrated track record of correctly identifying failure mechanisms is a reliable signal of analytical quality, independent of whether any specific prediction was confirmed on schedule.
Firsthand, decades-long familiarity with someone's output (since 2002) is itself credible evidence of their reliability, not merely anecdote.

Inferences:

If mechanism-tracing without date-setting is the standard for trustworthy AI skepticism, commentators who instead make falsifiable, date-bound collapse predictions should be weighted as more brittle — a comparison the piece never draws explicitly (e.g., against other AI-bubble commentators in this corpus).

The centaur/reverse-centaur distinction is a useful evaluative tool for any AI tool decision. strength: medium

Explicit premises:

A centaur stays in charge and does better work than alone.
A reverse centaur serves a machine that sets the pace.
The same software can produce either outcome, so the deciding factor is how it's used.

Implicit premises:

The user has enough agency over how a tool is used for the question to be actionable — the framework presumes usage is a free individual choice, not one structurally imposed by an employer's metrics, an SLA, or the tool's own interface incentives.
The question is answerable through the user's own self-reflection, without needing external verification or measurement.

Inferences:

If the same software can go either way, then at scale it is organizational policy and tool design — not just individual usage discipline — that determines which regime actually prevails; the piece stops at the individual heuristic and doesn't follow this to the organizational level.

The accountability sink explains why 'human in the loop' can fail as a safeguard even when the human is diligent. strength: high

Explicit premises:

The radiologist signs off but the AI made the call; the vendor's terms of service ensure liability lands on the signer.
Attention drifts when a system is right most of the time — a task-structural feature, per Bainbridge (1983), not an individual failing.

Implicit premises:

Legal and contractual liability rules — who is deemed to have 'signed off' — are what create the sink; it is not solely a psychological/attentional phenomenon.
The person placed in the loop typically lacks the bargaining power to renegotiate the vendor's terms of service that assign them the liability.

Inferences:

If the accountability sink is jointly structural (contractual allocation of liability) and cognitive (vigilance decrement), then adding more review steps without changing either the liability allocation or the task design will not fix it, and may create a false sense of safety — the piece gestures at this ('isn't much of a plan on its own') but does not develop it into a policy recommendation.

Narratives

Svenson's own history reading Doctorow since 2002, when a Boing Boing link (Doctorow was then co-editing it) was an early sign his own blog was working.
→ Grounds 'why I trust him' in a personal, decades-long relationship to Doctorow's output rather than in borrowed reputation — makes the trust claim feel earned rather than asserted.
A radiologist signs off on scans an AI has already flagged; when the call is wrong, it's the radiologist's signature — not the AI's — on the record.
→ Makes the abstract 'accountability sink' mechanism concrete and portable, and sets up the direct analogy to lawyers signing briefs that follows immediately.
Svenson mentions he has a copy of Doctorow's new book, The Reverse Centaur's Guide to Life After AI, and is 'partway in.'
→ Signals ongoing, current engagement (not a stale recommendation) and functions as a soft epistemic hedge — he's vouching for the book's ideas via this piece before having finished it himself.

Questions addressed

✅ Why trust a skeptic when everyone else on the recommended list is optimistic?
Because Doctorow's skepticism is disciplined by mechanism-tracing and freedom from financial stake, not by reflexive contrarianism.
✅ How can you tell whether a given AI tool helps you or exploits you?
The centaur/reverse-centaur question: does it let you do better work, or does it just make you keep up? — offered as a heuristic, not a rigorous decision framework.
✅ Why does 'I always review it' fail as a safety plan for AI-assisted work?
Because the accountability sink (liability assignment via signature) combines with vigilance decrement (attention drift when a system is usually right) to let errors through even diligent review.
⭕ Once you accept the accountability sink and vigilance-decrement mechanisms are real, what should you actually do differently in practice?
Not addressed — the piece names the mechanisms and points readers to Doctorow's own site/book, but stops short of any concrete practice or policy change beyond 'be more careful about who benefits.'

Gaps

incompleteThe centaur/reverse-centaur test presumes the user has free choice over how a tool is used, but many real deployments (metrics-driven review quotas, AI-embedded workflow software, employer mandates, interface defaults that nudge pace-setting) structurally impose the reverse-centaur mode regardless of individual intent. The piece never addresses who actually controls the answer to its own diagnostic question, or what a worker with no control over tool configuration is supposed to do with it.
Read next: extract-mosseri-ai-tailwind-authenticity#g4 raises the mirror-image gap (no criteria for which workflows to hand to AI at an organizational level); also relevant: workplace-power and algorithmic-management literature on who actually sets the pace of AI-mediated work.
incompleteThe claim that Doctorow's identified mechanisms 'come unraveled' 'a good share of the time' is asserted with no track record, count, or falsifiability check — an ironic gap given that the piece's own stated standard for trustworthy commentary is that it traces mechanism rather than merely asserting outcomes.
Read next: A systematic retrospective checking specific Doctorow mechanism-claims (enshittification predictions, AI-bubble timing claims) against what actually happened.
illogicalThe piece treats Doctorow's freedom from financial stake in AI's success as a mark of trustworthiness, but doesn't address that Doctorow, like any commentator with a public brand and a new book to sell, has a reputational/commercial interest running in the opposite direction: his identity and income depend on being the reliable hype-puncturer. This is the same blind spot the corpus already flags for a different AI skeptic.
Read next: extract-marcus-mythos-overblown#g4 flags the identical gap for Gary Marcus ('does not disclose his own potential bias as a long-standing AI skeptic... His brand depends on debunking AI hype') — read together as a general pattern: skeptic-brand incentives are a symmetric case of the same structural-incentive logic the corpus applies to hype (vc-hype-incentive).
uninformedThe lawyer-specific accountability-sink claim (liability lands on whoever signs, regardless of who really made the call) is asserted without engaging actual bar-discipline opinions or case law on AI-assisted legal work — there are now numerous documented sanctions cases for AI-fabricated citations that would let the claim be checked against how courts and bars are actually allocating fault (e.g., whether reliance on an AI tool is treated as aggravating, mitigating, or irrelevant).
Read next: Bar ethics opinions and case law on AI-hallucinated citations and attorney sanctions.

Closes gaps elsewhere

Adam Mosseri: AI is a tailwind for authenticity · g4 — Partially closes the gap by supplying a portable heuristic question for classifying a workflow as worth handing to AI or not — 'does this let me do better work, or does it just make me keep up with it?' — though this is a single self-applied diagnostic question, not the risk-tiering framework or organizational-level criteria the original gap asked for.

Everyone I recommend on AI is optimistic. Except Cory Doctorow.

Structure

Key terms

Propositions

Arguments

Narratives

Questions addressed

Gaps

Closes gaps elsewhere

Related documents