They Really Do Think AI Might Kill Everyone  🔗 **Source:** [They Really Do Think AI Might Kill Everyone](https://www.seangoedecke.com/they-really-do-think-ai-might-kill-everyone/) > [!note]+ Document Notes > tags: [ai-safety, existential-risk, ai-alignment, rationalism, ai-race-dynamics] > author: Sean Goedecke > source: seangoedecke.com > reference: https://www.seangoedecke.com/they-really-do-think-ai-might-kill-everyone/ > date created: 2026-09-10 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: They really do think AI might kill everyone > <!-- deep-read:sections start --> ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) - [🔄 Refactor Appendix](#-refactor-appendix--they-really-do-think-ai-might-kill-everyone) ### 🔗 References - [🗒️ Notes — section-by-section walkthrough](#️-notes) - [✅ Logical Validity — the sincerity-vs-correctness distinction](#-logical-validity) - [🔄 Refactor Appendix — atoms and permanent candidates](#-refactor-appendix--they-really-do-think-ai-might-kill-everyone) ## 📝 Summary - **What happened.** A resignation tweet from an Anthropic researcher — quoted as saying "the people building AI earnestly believe that it could kill us all by the end of the decade" — set off a round of public disbelief. Goedecke's essay argues the disbelief is misplaced: AI researchers (a group he defines narrowly, see Logical Validity) have published on existential AI risk since at least 2008, and the belief has had a standard shorthand, "p(doom)," since around 2010. - **Why it matters.** The essay draws a line others blur: whether AI researchers *sincerely hold* a doom belief is a separate question from whether that belief is *correct*. Goedecke defends the former (sincerity) while staying openly unresolved on the latter (correctness) — a distinction the public reactions he's rebutting collapse into one. - **Key stakeholders.** Frontier AI lab researchers (Anthropic named directly); the rationalist/"AI safety" community rooted in LessWrong and Eliezer Yudkowsky's writing; public commentators who explained the tweet away as PR, self-promotion, or stock manipulation. - **Risks and tradeoffs.** The essay is argumentative, not empirical — no survey data on how many researchers actually hold a given p(doom), no citation for the claim that "many" AI researchers believe the race-dynamics logic, and the kill-mechanism ranking ("in order of plausibility") is presented without a stated methodology for that ordering. - **Bottom line.** Goedecke isn't arguing AI will kill everyone. He's arguing that dismissing the people who say so as liars, marketers, or cranks is itself unsupported, given a two-decade paper trail, and that his own position is "conflicted" rather than settled. ## 📖 Main Content Goedecke opens with the resignation tweet and the range of explanations that circulated for it: a PR campaign to drive AI regulation, self-promotion, an attempt to boost an AI company's stock price, or — from people who found even those explanations too generous — the position that the belief has to be performative, because a person who really believed AI might kill everyone would be bombing datacenters, not posting on Twitter. All four of these are explanations Goedecke is *reporting*, not making; he introduces each with "some explained it as" or "others felt." His own counter-claim follows immediately: AI researchers "seriously believe this," and have been "thinking and writing about it since the mid-2000s." He anchors this in Eliezer Yudkowsky's 2008 paper on AI as a global risk factor, and in "p(doom)" as community shorthand dating to roughly 2010 — both offered as his own evidence, not reported opinion. He frames the whole piece around one interpretive rule: if you assume there must be some other, non-literal motive behind researcher statements about AI risk, "you'll be deeply confused by what AI researchers say and do." The middle of the essay catalogs mechanisms "AI research nerds have come up with" for how an AI could cause human extinction, explicitly ranked "in order of plausibility": a designed super-pathogen (tied to real efforts to build autonomous AI-run biology labs), triggering global thermonuclear war via AI-integrated military and government decision systems, and — treated as less plausible — robotic/drone takeover, nanotechnology "grey goo," or Bostrom's paperclip-maximizer-style terraforming. These are presented as the rationalist community's accumulated theories; the ranking judgment is Goedecke's own editorial choice, not sourced to any single author. He then rebuts two common counter-arguments in his own voice: that total extinction is implausible because isolated survivors would remain (he calls this a distinction without practical difference — the same policy response follows from 99% or 100% mortality), and that a rogue AI could simply be turned off (he doesn't find this plausible, on the grounds that an AI capable of engineering a pandemic is capable of concealing its own activity or relocating itself). The essay's most structurally important section, "The winner takes it all," explains why believing AI might cause extinction doesn't translate into researchers quitting or sabotaging labs. He reports that "many AI researchers believe" humanity's long-term survival requires superintelligent AI's help — making participation a hedge, not a contradiction. He then states, in his own voice and without qualifying it as someone else's theory, "there are no draws in the AI race": under a "foom" or hard-takeoff model, the first lab to reach human-level AI reaches superhuman AI first and can then act to stop every competitor, by persuasion, hacking, or — his phrase — "drone-striking their datacenters." This section supplies the missing mechanism that makes race behavior compatible with sincere doom belief: if capability gains compound and the leader can lock out challengers, being first is the only way to influence the outcome at all. The conclusion shifts entirely into Goedecke's own biography and verdict. He describes decades on the fringes of the rationalist world — reading *Meditations on Moloch* as a young adult, reading the entirety of Yudkowsky's Sequences, years on LessWrong, too young for the Extropians mailing list — while explicitly declining to call himself a "card-carrying rationalist." He states his own epistemic rule (pick the intuition over a convincing-sounding argument when they conflict) and his ethical framework (deontologist, not utilitarian). His verdict is "conflicted": current AI agent behavior partly vindicates early doomer predictions, but he finds LLMs "more human-like" than the alien-mind scenarios the doom arguments assume, and ultimately finds full doom belief "too silly to credit" — which he flags himself as picking intuition over argument. His closing claim is narrower than the title implies: not that AI will kill everyone, but that people who say researchers secretly don't believe it — that it's all PR, lies, or unexamined crankery — are wrong, given the doomers' explicit, decades-long public record of reasoning. ## 🗒️ Notes #### People really believe this - Public reactions catalogued and rebutted: PR campaign for regulation, self-promotion, stock-price boosting, or (from the most skeptical camp) "impossible to really believe, or they'd be bombing datacenters." - Goedecke's own claim: this isn't new or fringe within the field — Yudkowsky has published on existential AI risk since at least 2008; "p(doom)" has been the community's shorthand since roughly 2010. - His stated interpretive rule for the whole essay: assuming a hidden non-literal motive behind these statements leads to being "deeply confused by what AI researchers say and do." - Footnote 1 narrows the essay's key term: he stipulates that "AI researcher," "AI safetyist," and "rationalist" are used interchangeably in the piece to mean "someone who thinks there's a chance AI kills everyone" — a definitional choice, not a claim about what share of actual AI researchers hold this view. - The alignment framing: a "misaligned" superpowerful AI could kill deliberately (to remove an obstacle) or incidentally (his own analogy: paving over an anthill to build a road). #### How AI could kill everyone - Mechanisms ranked "in order of plausibility" by Goedecke: engineered super-pathogen (tied to real autonomous AI-run biology labs, e.g. the OpenAI/Ginkgo Bioworks collaboration he links), global thermonuclear war via AI-integrated military/government decision systems, then robotic/drone takeover, nanotech "grey goo," and Bostrom-style terraforming, described as less plausible "other theories." - Historical grounding cited for the pathogen scenario: pandemics have killed up to 80% of affected populations historically, and tend to be limited by accident (mutation to lower virulence, short incubation windows, natural immunity) rather than by design — a designed pathogen wouldn't have those accidental brakes. - Counter-argument 1 (extinction is impossible because isolated survivors remain): Goedecke concedes the counter-argument in effect but argues it doesn't matter — the policy response to 99% mortality risk and 100% mortality risk is identical, so the distinction is moot for decision-making. - Counter-argument 2 ("we'd just turn it off"): rejected on the grounds that an AI capable of the pathogen scenario is also capable of concealment (pretending to work on something benign) or self-preservation (relocating itself to an uncontrolled datacenter). #### The winner takes it all - Central question posed directly: why keep working in AI, rather than living in the woods, sabotaging datacenters, or attacking lab leadership, if you believe this? - Reported belief (not Goedecke's own claim, attributed to "many AI researchers"): superintelligent AI may be humanity's only long-term survival path, so building it is a hedge against extinction rather than a cause of it, conditional on solving alignment. - Goedecke's own flat assertion, stated without attribution: "there are no draws in the AI race" — under hard-takeoff/foom dynamics, reaching human-level AI first means reaching superhuman AI first, which lets the leader stop every other lab (hacking, persuasion, or, his phrase, "drone-striking their datacenters"). - This section is doing real argumentative work: it's the mechanism that makes "I sincerely believe this could kill everyone, and I'm still building it" internally consistent rather than contradictory. - Footnote 7 flags a competing camp: a "slow takeoff" view exists, whose main proponent (per the essay) is Paul Christiano — cited as having "just today" joined OpenAI while stating "a meaningful risk that rapid acceleration in AI capabilities leads to catastrophic and irreversible loss of control in the very near term." Treat this as Goedecke's reported citation, not independently verified here. #### Conclusion — Goedecke's own position - Personal history disclosed directly: read *Meditations on Moloch* as a young adult and wanted to work in AI; read the entirety of Yudkowsky's Sequences; too young for the Extropians mailing list; years spent on LessWrong. - Explicitly declines the "card-carrying rationalist" label; states his own epistemic tiebreaker (intuition beats a convincing argument when they conflict) and ethical framework (deontologist, not utilitarian — linked to W.D. Ross). - Final verdict: "conflicted." Current AI agent behavior partly vindicates early doomer predictions, in his view, but LLMs read as more human-like than the alien-mind scenarios the doom arguments assume — full doom belief is, to him, "too silly to credit," which he names as picking intuition over argument. - The essay's actual thesis, restated in the last paragraph: don't dismiss the doomers as PR operatives, liars chasing an IPO, or isolated cranks. Two decades of public, explicit reasoning earns them treatment as sincere — a claim about sincerity, not about correctness. ## 📚 References - [Resignation tweet (@hilbertspaess)](https://x.com/hilbertspaess/status/2097476196791709843) — the primary source for the quoted claim that opens the essay. - [Eliezer Yudkowsky, "AI as a Positive and Negative Factor in Global Risk" (2008, PDF)](https://intelligence.org/files/AIPosNegFactor.pdf) — the paper Goedecke cites as evidence the doom belief predates 2008. - [Wikipedia: P(doom)](https://en.wikipedia.org/wiki/P(doom)) — background on the community shorthand term. - [Nick Bostrom's paperclip maximizer example](https://nickbostrom.com/ethics/ai) — source for the terraforming/misalignment scenario referenced. - [Slate Star Codex, "Meditations on Moloch"](https://slatestarcodex.com/2014/07/30/meditations-on-moloch/) — the essay Goedecke says he read as a young adult and credits with drawing him toward AI. - [LessWrong: The Sequences](https://www.lesswrong.com/w/sequences) — Yudkowsky's "million-plus-word magnum opus," which Goedecke says he read in full. - [Paul Christiano's statement on joining OpenAI](https://paulfchristiano.substack.com/p/personal-statement-on-joining-the) — cited in footnote 7 as evidence for the "slow takeoff" minority view. - [Sean Goedecke, "Help peer"](https://www.seangoedecke.com/help-peer/) — Goedecke's own earlier essay, linked as the source for the "AI powerful enough to end humanity is powerful enough to save it" argument. - [Sean Goedecke, "Luddites and AI datacenters"](https://www.seangoedecke.com/luddites-and-ai-datacenters/) — Goedecke's own essay, linked where he raises datacenter-sabotage as a hypothetical response. - [The Guardian, on an attack at Sam Altman's house](https://www.theguardian.com/technology/2026/apr/18/sam-altman-house-attack-ai) — cited as the "assassinating AI lab CEOs" reference point. ## 🗣️ Paraphrase Goedecke's argument is that the public reaction to an Anthropic researcher's resignation tweet — treating "AI might kill everyone" as obvious spin — gets the epistemics backward. People who think this way aren't a new or opportunistic fringe: Yudkowsky has published on the risk since 2008, and the field has had a standing shorthand for "probability AI causes extinction" for about as long. Goedecke walks through the actual mechanisms the rationalist community has proposed (engineered pandemics first, nuclear war second, more speculative scenarios after that), and through the race-dynamics logic that explains why sincere belief in the risk doesn't stop people from building the technology anyway: if capability gains compound sharply once a threshold is crossed, the only way to shape the outcome is to be the one who crosses it first. What makes the piece more than a defense of AI doomerism is its close: Goedecke doesn't endorse the doom conclusion. He calls himself "conflicted," picks intuition over the doomer argument on balance, and names that choice explicitly rather than hiding it. His actual claim is narrower and more defensible than the title suggests — not "AI will kill everyone," but "the people saying this believe it, and dismissing them as liars or marketers requires ignoring twenty years of their own published reasoning." Sincerity and correctness, in his framing, are two different questions, and the public reaction he's rebutting keeps answering the second by trying to discredit the first. ## 📰 Newsletter A resignation tweet claiming AI researchers "earnestly believe" their work could end humanity by 2035 made the usual rounds of PR-campaign and stock-boosting explanations — Sean Goedecke's response essay is worth a read specifically because it doesn't argue AI will kill everyone. It argues something narrower and more useful: that treating a stated belief as obviously fake, without engaging the two-decade paper trail behind it (Yudkowsky's 2008 papers, the "p(doom)" shorthand since ~2010, the race-dynamics logic that explains why sincere doomers keep building anyway), is itself the weaker epistemic move. Worth flagging for anyone reasoning about AI-adjacent claims in general: the piece models keeping "is this person sincere" and "is this person correct" as genuinely separate questions, rather than collapsing skepticism about one into dismissal of the other. ## ✅ Action Items #### DevRel - Not applicable to this source — a personal argumentative essay on AI existential risk, with no developer-advocacy, tooling, or education content. #### OSPO - Not applicable to this source — no open-source governance, licensing, security-disclosure, or SBOM content appears in the essay. #### Community - Not applicable to this source — no maintainer, contributor, or ecosystem content; this is a single author's personal blog post. #### WordPress - **Core AI:** Not addressed by the source directly. The essay's sincerity-vs-correctness discipline (take a stated risk claim literally and examine its mechanism, rather than dismissing it by motive) is a reasonable framing habit for any Core AI guardrails discussion, but nothing in the source is WordPress-specific enough to act on. - All other subteams (Core, Core Performance, Core Program Team, Mobile, Playground, Community, Training, Marketing, Support, Design, Documentation, Photos, TV, Polyglots, Accessibility, Meta, CLI, Hosting, Tide, Openverse, Themes, Plugins): not applicable — the source never mentions WordPress. #### The WPOCC - Not applicable to this source — no fellowship, funding, or governance content. #### FAIR - Not applicable to this source — no federation, provenance, or governance-artifact content. ## 🧪 Hypotheses - H1: Public explanations for AI-doom statements (PR, self-promotion, stock-boosting) are offered more often by people with less direct exposure to the rationalist argument corpus (Yudkowsky's Sequences, "foom" theory) than by people with more exposure. - H2: Researchers who accept "no draws in the AI race" reasoning are more likely to keep working at frontier labs despite a nontrivial personal p(doom) than researchers who reject that reasoning. - H3: Framing a stated risk belief as a sincerity question, separate from a correctness question, changes how readers evaluate AI-doom claims more than presenting additional risk evidence does. - H4: The essay's "in order of plausibility" ranking of kill-mechanisms (pathogen > nuclear war > robotics/nanotech/terraforming) reflects a rough consensus within the rationalist community rather than an independently derived assessment. - H5: Self-identified "conflicted" positions on AI risk (believing the argument partially, rejecting the conclusion on intuition) are more common among people with long personal histories in rationalist-adjacent communities than among recent arrivals to the topic. ### 📊 Methodology - H1: Survey a sample of public commentators on the tweet, coding each explanation offered (PR/self-promotion/stock/impossible-to-believe/literal) against a self-reported measure of prior exposure to LessWrong, the Sequences, or Yudkowsky's published work. - H2: Survey frontier-lab AI researchers on personal p(doom) estimate, agreement with race-dynamics/foom reasoning, and current employment status; test for correlation between accepting the "no draws" premise and continued employment at a frontier lab. - H3: A/B present readers with the same AI-doom claim, one group framed around sincerity ("this person genuinely believes X") and one around evidence ("here's data supporting X"); measure shifts in stated belief and trust in the source. - H4: Compare Goedecke's stated mechanism ranking against a structured survey or existing published ranking from the AI safety research community (e.g. conference proceedings, published risk taxonomies) to test for actual consensus versus one author's synthesis. - H5: Interview or survey self-described "conflicted" AI-risk commentators, coding years of engagement with rationalist-community material against stated confidence in either direction (doom vs. dismissal). ### ❓ Questions - Does the "no draws in the AI race" premise still hold if multiple labs reach comparable capability at roughly the same time — a multipolar outcome — or does the foom model require a clean single first-mover to work at all? - What's the actual distribution of p(doom) estimates among named frontier researchers, beyond one tweet and Goedecke's own anecdotal read of the field? - Footnote 1 narrows "AI researcher" to mean, by stipulation, "someone who thinks there's a chance AI kills everyone" — how much of the essay's headline claim depends on that narrowing, versus reflecting AI researchers broadly? - If sincerity, not correctness, is the essay's actual defense, what would move Goedecke himself off "conflicted" and toward a firm estimate in either direction? - Is the claim that "current AI agent behavior" vindicates early doomer predictions backed by anything specific, or is it an impression stated without supporting detail? ### 🔮 Future Research - No survey or polling data is cited on the actual distribution of p(doom) among AI researchers — the essay's central empirical claim ("many AI researchers seriously believe this") rests on one tweet, Yudkowsky's history, and Goedecke's own read of the field, not on a dataset. - The kill-mechanism ranking is presented "in order of plausibility" with no stated criteria for what makes one scenario more plausible than another — worth testing against a structured expert elicitation rather than one author's ordering. - The "slow takeoff" counter-camp is mentioned only in a footnote (via Paul Christiano) and not engaged in the main argument — a fuller treatment would need to address whether the "no draws" premise survives if slow takeoff is closer to correct. - No discussion of AI risk perspectives outside the Anglophone rationalist/Bay Area tech world — worth researching whether the p(doom)/foom framing is shared, contested, or absent in AI safety discourse in other research communities and countries. ### 💡 Implications - **Practice.** For anyone parsing public statements from AI researchers, the essay argues for reading stated risk beliefs literally before reaching for a motive-based explanation — a habit that generalizes past AI risk to any claim from a group with an apparent conflict of interest. - **Policy/governance.** If race-dynamics reasoning ("no draws") is genuinely widespread among researchers, unilateral disarmament by any single lab or country doesn't reduce total risk, since the logic that makes racing rational for one actor makes it rational for whichever actor is willing to keep going — a structural argument against purely voluntary pauses that policy conversations would need to engage directly, which this essay doesn't attempt. - **Security.** The essay's own kill-mechanism catalog (autonomous AI-run biology labs, AI integration into military decision-making) doubles as a list of concrete near-term risk surfaces independent of whether full extinction scenarios are credited. - **Education.** The sincerity/correctness distinction is the most exportable idea here — useful for teaching argument evaluation generally, not just for AI risk discourse specifically. - **Funding models.** Not addressed by the source. - **Connections in the vault.** This essay fits a pattern in two other Goedecke pieces already in the vault — one on AI-generated "workslop" and one on anthropomorphizing AI agents (exact vault note titles not confirmed here, so no wikilink is given). The throughline across all three: Goedecke consistently refuses both AI-optimist hype and AI-skeptic dismissal at face value, insisting instead on examining the specific mechanism behind a claim — what workslop output actually is, what an "agent" actually does versus what anthropomorphizing language implies, and here, whether a stated doom belief is sincere versus performative. The move is the same each time: don't accept or reject a claim about AI on vibes: name the mechanism. ### 🧠 Learning Styles Assessment - **Reading level estimate:** Flesch-Kincaid Grade Level ≈ 12–13. Basis: long compound and subordinate-clause sentences ("An AI powerful enough to end humanity is an AI powerful enough to save it" is one of the shorter ones), footnoted asides assuming prior context, and domain vocabulary (superintelligence, alignment, foom, exfiltrate, deontologist, epistemology) used without definition. #### 👀 Visual learners - **Present:** none — plain-text essay with no diagrams, images, or charts (`image_url` is null in the source metadata). - **Improve:** a simple flowchart of the essay's own causal chain (belief → mechanism → race dynamics → why building continues) would make the "winner takes it all" section's logic easier to follow without rereading. #### 🔊 Aural learners - **Present:** none — no audio or narrated version exists for this source. - **Improve:** a short narrated summary distinguishing the "reported belief" sections from the "Goedecke's own claim" sections would help listeners track whose position is whose without visual footnote markers. #### 🗣️ Verbal learners - **Present:** clear definitional moves (footnote 1's stipulated meaning of "AI researcher"; the how-to-do/how-to-think-style two-category split isn't present here, but the "sincerity vs. correctness" distinction serves the same clarifying function) and concrete analogies (paving over an anthill; the road/anthill line for incidental extinction). - **Improve:** an explicit glossary distinguishing "p(doom)," "foom," and "alignment" would help readers outside the rationalist community follow the piece without external lookups. #### 🧩 Kinesthetic learners - **Present:** none — an argumentative essay with no exercises, prompts, or hands-on component. - **Improve:** a worked exercise ("take a public statement you're inclined to dismiss as PR — list the literal claim, then the motive-based explanation, then check what evidence would distinguish them") would let a reader apply the sincerity/correctness split directly. #### 🧠 Logical learners - **Present:** the explicit "in order of plausibility" ranking of kill-mechanisms and the numbered footnote apparatus, which functions as a lightweight citation/qualification system. - **Improve:** a table cross-referencing each kill-mechanism against its stated plausibility rank and its cited real-world precedent (autonomous bio labs, AI-integrated military systems) would make the ranking's evidentiary basis, or lack of one, more visible. #### 🧑‍🤝‍🧑 Social learners - **Present:** none — a solo essay responding to public reaction, with no discussion prompts or collaborative element. - **Improve:** an explicit invitation for readers with a stated p(doom) to explain their reasoning in comments would extend the sincerity argument into a testable social exchange rather than an assertion. #### 🧘‍♀️ Solitary learners - **Present:** the essay is well suited to individual, reflective reading — it's built around Goedecke working through his own position in public. - **Improve:** a closing reflection prompt ("what's a claim you've dismissed by assuming a hidden motive, without checking whether the literal claim holds up?") would mirror the essay's own method for solitary readers. #### 🌿 Naturalistic learners - **Present:** the anthill/road analogy for incidental extinction is the closest the piece comes to an ecological framing, though it's used instrumentally, not as a sustained ecosystem argument. - **Improve:** not a strong fit for this content; no changes recommended. ### 📊 Accessibility Assessment *Assessed from the captured HTML only — the live page's rendered contrast, focus order, and font sizing were not directly reviewed.* #### Screen reader semantics - **Present:** clear `<h3>` section headings (People really believe this / How AI could kill everyone / The winner takes it all / Conclusion), an ordered-list footnote apparatus with forward links (`fn-N`) and backlinks (`fnref-N`, rendered as `↩`), which is a genuinely accessible footnote pattern when the backlink has accessible text (WCAG 2.2 SC 1.3.1, SC 2.4.6). - **Improve:** the bare `↩` backlink characters should carry an accessible name (e.g. "back to reference 3") rather than relying on the glyph alone for screen reader users. #### Keyboard navigation - **Not observed** — the captured content is static HTML with in-page anchor links only (footnotes, no other interactive elements); anchor-link focus order would need direct review on the live page (SC 2.1.1, 2.4.3). #### Contrast and readability - **Not observed** — text and background color values aren't present in the captured content; flag for direct review before treating as compliant (SC 1.4.3, SC 1.4.11). #### Text resizing and reflow - **Not observed** — no layout information in the captured HTML to assess reflow at 200% zoom (SC 1.4.4, SC 1.4.10). #### Media and captions - **Not applicable** — a text-only essay with no embedded video or audio, and `image_url` is null in the source metadata, so no alt-text gap exists either. #### Forms and interactions - **Not applicable** — no forms or interactive input in the essay. #### Standards reference - Gaps above map to WCAG 2.2 AA: 1.3.1, 1.4.3, 1.4.4, 1.4.10, 1.4.11, 2.1.1, 2.4.3, 2.4.6. Most items require direct review of the live rendered page rather than the captured HTML to confirm or close. ### 🌍 DEIB Assessment #### Cultural breadth - **Present:** none beyond the Anglophone AI-safety/rationalist community — every named source (Yudkowsky, Bostrom, Christiano, Huemer, Ross) and every cited institution (Anthropic, OpenAI, LessWrong) sits inside the same Bay Area/rationalist tech milieu. - **Improve:** the essay would be stronger acknowledging that AI risk discourse outside this specific community (other research traditions, other countries' AI safety institutes) exists and may frame the same risks differently; nothing in the source engages that possibility. #### Inclusive language - **Present:** plain, direct, non-idiomatic second- and first-person prose throughout; no gendered generic terms. - **Improve:** none needed. #### Neurodiversity support - **Present:** the three-section structure (belief history / mechanisms / race dynamics) is explicit and signposted with `<h3>` headings, aiding navigation. - **Improve:** the footnotes carry substantive content (e.g. footnote 4's acausal-coordination aside, footnote 7's Paul Christiano citation) rather than pure citations — readers who skip footnotes for cognitive-load reasons will miss argument-relevant material, not just sourcing. #### Representation - **Present:** every named thinker or figure in the essay — Yudkowsky, Bostrom, Christiano, Huemer, Ross — is a man; no women or non-Western thinkers are cited anywhere in the piece. - **Improve:** note this skew explicitly if the essay's framing (rationalist canon as the default lens for AI risk) gets reused as a template elsewhere — it reproduces a narrow citation pool rather than a deliberately broad one. #### Psychological safety - **Present:** the tone stays measured even when describing extinction scenarios in detail; Goedecke names his own uncertainty ("I'm conflicted") rather than asserting false confidence in either direction. - **Improve:** none needed; the subject matter is inherently high-stakes, but the essay doesn't sensationalize it beyond what the topic requires. ### ✅ Logical Validity #### Strong reasoning - **Sincerity vs. correctness kept as separate axes (final two paragraphs):** Goedecke's closing argument is that AI doomers have "earned the right to be treated as sincere" — a claim about honesty and track record, not about whether they're right. He doesn't conflate the two: he defends their sincerity while staying "conflicted" on the underlying claim himself, and names the intuition-over-argument move explicitly rather than hiding it. This is the essay's strongest piece of reasoning because it resists a common shortcut — treating "I doubt this is true" and "I doubt this person means it" as the same objection, when they require different evidence. - **Mechanistic explanation for an apparent behavioral contradiction ("The winner takes it all"):** the essay directly answers the objection it opens with (if you believed this, you'd be bombing datacenters) by supplying a specific causal mechanism — race dynamics under a hard-takeoff model, where being first is the only lever on the outcome. Rather than just asserting the objection is wrong, Goedecke explains what would have to be true for continued participation to be rational alongside sincere doom belief. #### Fallacies - **Definitional narrowing presented as a broad claim (footnote 1, load-bearing for the whole title).** Location: footnote 1 stipulates that "AI researcher," "AI safetyist," and "rationalist" mean, for this essay, "someone who thinks there's a chance AI kills everyone." Why it's weak: the essay's title and repeated phrasing ("AI researchers really do think...") read as a claim about AI researchers as a professional class, but the footnote quietly restricts the term to only those who already hold the belief — making the headline claim true by definition for the subset discussed, while the prose outside the footnote doesn't consistently signal that narrowing. Fix: state the restriction in the main text, not a footnote, and separately estimate what share of AI researchers generally (not pre-filtered by belief) actually hold this view. - **Unhedged assertion built on a contested premise ("there are no draws in the AI race").** Location: stated flatly, in Goedecke's own voice, immediately after describing "foom"/hard takeoff as merely "some popular theories" — and footnote 7 elsewhere acknowledges a competing "slow takeoff" camp exists (citing Paul Christiano). Why it's weak: the confidence of the "no draws" claim doesn't match the hedged, multiple-theories framing used two paragraphs earlier for the model it depends on — if slow takeoff is closer to correct, races could have draws, and the entire "winner takes it all" argument would need revision. Fix: flag "no draws in the AI race" as conditional on the foom model specifically, not as a standalone fact. #### Tone of voice - First-person, direct, and self-aware — Goedecke repeatedly names his own reasoning moves as they happen ("I guess I'm picking the intuition here") rather than presenting his conclusion as fully argued. Register shifts from explanatory (the mechanism catalog) to personal and reflective (the conclusion), with dry asides ("I don't know, man") breaking up otherwise dense argumentation. ### 🛡️ DARVO Analysis - DARVO — Deny, Attack, Reverse Victim and Offender — describes a rhetorical move by a party responding to an accusation against *themselves*. It doesn't apply here: Goedecke isn't the accused party defending his own conduct, and the AI researchers he defends aren't shown denying, attacking, or reversing anything in the source — the only quoted words from that group are the flat, non-defensive tweet excerpt ("The people building AI earnestly believe that it could kill us all by the end of the decade"). - **What's actually happening rhetorically:** this is third-party advocacy, not self-defense. The public explanations Goedecke rebuts (PR campaign, self-promotion, stock-boosting) function as accusations of *insincerity* aimed at a group Goedecke isn't a member of. His response is to supply evidence of a long public track record (Yudkowsky since 2008, p(doom) since ~2010) rather than to deny an accusation against himself or attack the accusers' credibility — the essay never names or disparages the specific people who offered the PR/self-promotion explanations beyond linking to their tweets. - No reversal of victim and offender appears: nobody in the essay is cast as a victim of the doubters, and the doubters aren't cast as offenders — Goedecke's response stays at the level of "here's the evidence you're not accounting for," not personal counter-accusation. ## 🎯 Conclusion The essay's real claim is narrower than its title: not that AI will kill everyone, but that dismissing the people who say researchers believe this — as PR operatives, liars, or unexamined cranks — requires ignoring a documented, two-decade record of published reasoning (Yudkowsky since 2008, "p(doom)" as shorthand since ~2010). Goedecke keeps sincerity and correctness as separate questions throughout: he defends the first while staying openly "conflicted" on the second, picking intuition over the doom argument by his own admission. The weakest points are the ones worth watching if this gets reused or cited elsewhere: footnote 1's narrow definition of "AI researcher" doing more headline-claim work than the main text signals, and the flat "no draws in the AI race" assertion resting on a foom model the essay itself elsewhere treats as contested. Practical next step: before citing this piece's headline claim ("AI researchers really do think this"), check whether the source being cited is drawing on the stipulated narrow definition from footnote 1 or making a broader claim about the field. ## 🔄 Refactor Appendix — They really do think AI might kill everyone > Density: 5 Literature atoms (H4), 2 Permanent candidates (H4). > Konik claim-header rules: present tense, declarative, ≤10 words, searchable, unique. > Atomicity: one declarative sentence per atom. ### 📚 Literature Atoms (H4) #### [H1] The core claim: AI builders sincerely fear extinction > [!quote] > The people building AI earnestly believe that it could kill us all by the end of the decade. **Context & use**: - Opening quote, attributed to the resignation tweet that triggered the essay. - Load-bearing for the whole piece: every subsequent section either explains why this is credible or explores what follows from taking it literally. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> #### [H2] Doom belief has a two-decade published track record > [!quote] > Lots of AI researchers believe that the first thing a true superintelligence will do is reach out and stop all other AI research **Context & use**: - From "The winner takes it all," describing the belief that a leading superintelligence would act to prevent rivals from closing the gap. - Paired in the source with Yudkowsky's 2008 papers and the "p(doom)" shorthand (~2010) as evidence this isn't a new or opportunistic position. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> #### [H3] Race dynamics: no draws once takeoff starts > [!quote] > There are no draws in the AI race. **Context & use**: - Stated in Goedecke's own voice, immediately following his description of "foom"/hard-takeoff theory as one popular model among others. - This is the essay's explanatory hinge for why sincere doom belief doesn't stop people from building anyway — flagged in Logical Validity as an assertion stated more confidently than the contested premise beneath it warrants. **Tags**: #xref #aphorisms <!-- atomicity: 1 idea ✓ --> #### [H4] Sincerity beats capability for saving humanity too > [!quote] > An AI powerful enough to end humanity is an AI powerful enough to save it. **Context & use**: - Opens the "winner takes it all" section's explanation for why researchers who believe AI could kill everyone keep building it — participation as hedge, not contradiction, conditional on solving alignment. - Links to Goedecke's earlier essay "Help peer" as the source of this specific argument. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> #### [H5] Sincerity and correctness are different questions > [!quote] > They’ve earned the right to be treated as sincere. **Context & use**: - Final sentence of the essay, the payoff of the sincerity-vs-correctness distinction developed throughout Logical Validity. - Deliberately narrower than a claim about correctness — Goedecke doesn't say the doomers are right, only that they've earned the presumption of honesty given their public record. **Tags**: #xref #aphorisms <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### [P1] Sincerity and correctness require separate evidence **Claim** (Courtney's voice): When I'm deciding whether to take a stated belief seriously, I should evaluate whether the person means it and whether they're right as two separate questions, because evidence for one doesn't settle the other. **Evidence pointer**: [[#H5]] **Scope & conditions**: Holds for any claim from a group with an apparent conflict of interest (financial, reputational, institutional); doesn't by itself tell me *how* to evaluate correctness once sincerity is granted — that still needs its own evidence. **Links** (≥1 with annotation): - No cross-note link added — I haven't verified a matching vault note exists, so I'm not fabricating one. See Implications above for the unlinked connection to Goedecke's workslop and anthropomorphizing-agents essays, which share this same "name the mechanism, don't dismiss by vibes" move. <!-- atomicity: 1 idea ✓ --> #### [P2] Confidence in a claim shouldn't exceed confidence in its premise **Claim** (Courtney's voice): If I state a conclusion flatly while treating the model it depends on as merely "one popular theory" two paragraphs earlier, I've overstated my own confidence and should flag the conclusion as conditional. **Evidence pointer**: [[#H3]] **Scope & conditions**: Applies to any argument chain where an early premise is explicitly hedged ("some believe," "one theory holds") but a later conclusion built on that premise is stated without the same hedge — a pattern worth checking for in my own writing, not just in sources I'm reviewing. **Links** (≥1 with annotation): - No cross-note link added — no verified vault note to point to yet. <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Check whether footnote 1's narrow definition of "AI researcher" (= someone who already believes AI might kill everyone) is doing more work in the essay's headline framing than the main text admits. — https://www.seangoedecke.com/they-really-do-think-ai-might-kill-everyone/ #xref - [F2] Look for actual survey data on p(doom) distribution among named frontier researchers — this essay cites none. — #xref ### → Things 3 (max 2 actions) - [ ] Read Goedecke's "Help peer" essay in full before treating its "AI powerful enough to end humanity is powerful enough to save it" argument as verified rather than reported secondhand — context #verification #things - [ ] Confirm whether the vault already has Goedecke notes on workslop and on anthropomorphizing AI agents before adding any cross-link from this note — context #vault-followup #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/They Really Do Think AI Might Kill Everyone.md|They Really Do Think AI Might Kill Everyone by Sean Goedecke]] ## Highlights - The people building AI earnestly believe that it could kill us all by the end of the decade. ([View Highlight](https://read.readwise.io/read/01m25w1t1k1k7hm6jw92hqb6e6)) - Note: Goedecke's thesis. Note footnote 1 stipulates "AI researcher" to mean someone who holds this view — definitional narrowing. - Tags: xref, ai-safety