Behind the Blog: How to Talk About AI Doom  🔗 **Source:** [Behind the Blog: How to Talk About AI Doom](https://www.404media.co/behind-the-blog-how-to-talk-about-ai-doom/) > [!note]+ Document Notes > tags: [ai-risk, editorial-judgment, newsroom-practice, collective-memory, ai-harms, media-criticism, anthropic] > author: 404 Media (Joseph Cox, Samantha Cole, Jason Koebler, Emanuel Maiberg) > source: 404media.co > reference: https://www.404media.co/behind-the-blog-how-to-talk-about-ai-doom/ > date created: 2026-09-11 > > > [!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 Do not quote or attribute any other line to the named author.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: Behind the Blog: How to Talk About AI Doom > <!-- 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) ### 🔗 References **Fifteen outbound links across three short sections.** Verified by extracting every `<a href>`. - [@25YearsAgoLive](https://x.com/25YearsAgoLive) — an account posting 2001 in real time, 25 years delayed, with two posts embedded showing the morning turn from a Mr. Peanut appearance to the first tower report. - [Watches of Espionage](https://www.watchesofespionage.com/) and its [write-up on Todd Beamer's Rolex](https://www.watchesofespionage.com/blogs/woe-dispatch/lets-roll-a-heros-rolex-froz) — recovered from the Flight 93 crash site, frozen on the date. - [Garrett Graff, *The Only Plane in the Sky*](https://www.amazon.com/Only-Plane-Sky-Oral-History/dp/150118220X) — the 9/11 oral history. - [GenerationGirl.com](http://generationgirl.com/) and its [September 2001 archive capture](https://web.archive.org/web/20010922092746/http://generationgirl.com/) — the Mattel/Barbie bulletin board. - [Anthropic's September 2026 threat intelligence report](https://www.anthropic.com/threat-intelligence-report-september-2026) — cited as read-in-progress, not as evidence. ## 📝 Summary - **What happened.** 404 Media's weekly "Behind the Blog" column, in which four staff write short first-person notes about the week. Two sections are 9/11 anniversary material; the third is Jason Koebler reasoning publicly about whether and how the outlet should cover AI existential risk. - **Why it matters.** The AI section is a newsroom doing its editorial deliberation in the open — naming a beat it has deliberately avoided, explaining why, and announcing it will try to enter the conversation. That is rarer and more useful than the coverage it is deliberating about. - **Key stakeholders.** 404 Media's staff and subscribers; the AI labs whose doom messaging the piece treats as partly promotional; readers who have been texting the author asking whether everyone is going to die. - **Risks and tradeoffs.** The piece is explicitly provisional — the author says he has not finished the report he cites and has an article coming. Reading it as 404 Media's settled position would misread it. - **Bottom line.** The durable content is a distinction: whether AI is "reasoning," and whether it will kill everyone, are separable from whether people directing large amounts of compute are already causing harm. The author argues the third question is answerable and under-covered while the first two absorb the attention. ## 📖 Main Content **Joseph Cox's section** points at three 9/11 projects rather than restating feelings he assumes readers have heard. The @25YearsAgoLive account replays 2001 in real time at a 25-year delay, and the two embedded posts show the morning turning over — a Mr. Peanut costume appearance on the *Today* show, then the first report of a fire in the North Tower described as a small plane, probably a Cessna, probably an accident. He also flags Watches of Espionage on Todd Beamer's recovered Rolex, noting the site is run by a former CIA case officer and carries a prepublication-review disclaimer, and Garrett Graff's oral history, which he says he had to stop listening to partway through. **Samantha Cole's section** surfaces an archived Mattel site where Barbie characters posted on a fictional bulletin board, and where on 13 September 2001 an uncredited post reassured children that the dolls were safe and told them to talk to a grown-up about their feelings. A week later the characters were baking cookies for cleanup volunteers and making cards for firefighters, ending on a line about being proud to be a New Yorker. She frames it as a brand feeling obliged to say something and working it into the plot. **Jason Koebler's section** is the analytical one. After a live event with The Intercept on AI in surveillance and war, he found the audience asking about viral "we're all gonna die" posts from Anthropic people, and he works through 404 Media's position in public. The outlet has covered AI's present harms and sat out the singularity beat. He states three things at once: the doom material functions as marketing for the companies producing it; he does not believe AI is imminently going to kill everyone; and people pointing large amounts of compute at objectives are already producing bad outcomes that are getting worse. He concludes the outlet will try to enter the conversation, and names why it has not — the subject is fraught and easy to get embarrassingly wrong. ### 🗒️ Notes - **The real-time account's value is the ordinary opening.** The first embedded post is a mascot appearance; the second is the first tower report, framed as a probable accident with a small aircraft. Read in sequence, it recovers something retrospective accounts cannot: that nobody knew yet. > [!quote] > Scrolling through the account is a fascinating way to remind yourself of some of the individual stories of people from that day, or learn some new ones. - **A sourcing caveat offered unprompted**, about a source Cox finds valuable. > [!quote] > I might try to get Watches of Espionage on the pod but I imagine that might be unlikely given they are a former CIA case officer and a disclaimer at the bottom of the newsletter says it has been reviewed by the CIA’s prepublication review board. - **An honest note about consuming difficult material.** > [!quote] > I listened to maybe a third of it and had to stop for a bit because it got really intense. - **The Barbie artifact.** A commercial children's brand addressing a mass-casualty event inside its fiction, four days later. > [!quote] > Mattel and all the dolls and toys from Mattel want to reassure you that we are all fine. - **Cole's framing of why it is worth surfacing**, which is the analytical move rather than the curiosity. > [!quote] > But I thought this little snapshot of a brand obviously feeling like it has to say something, and working it into the plot of the characters, was interesting to me. - **The editorial problem, named directly.** > [!quote] > One thing that was clear at the event is that we need to figure out how to talk about the big AI companies and the potential threat they pose. - **The beat 404 Media has occupied, and the one it has not.** > [!quote] > But we have largely sat out the “OpenAI rogue agents hack HuggingFace” story, the AI doomers singularity risk, things like this. - **Three positions held simultaneously, which is the durable content.** > [!quote] > I think everyone in this world is right: The DOOM stuff serves as good marketing for these companies, and I do not think AI is imminently going to kill us all. But it is also indisputable that idiot human beings directing AI (or not knowing how to direct AI) is leading to some pretty bad things happening, and it is easy to imagine how this is about to get worse. - **The move that makes the position coherent** — setting aside the capability debate as not load-bearing for the harm question. > [!quote] > It sort of doesn’t matter whether AI is “reasoning” or not or whether AI is very bad at doing lots of basic things, the fact of the matter is that people can throw huge amounts of compute at something and the AI will keep trying to do that thing, and there’s all sorts of various ways this can and is going sideways. - **A cited source marked as not yet read.** Unusual and worth crediting. > [!quote] > I haven’t fully gone through Anthropic’s threat report from yesterday, but I have been assured it is at the very least entertaining reading - **The reason for the reversal, which is reader demand rather than a new argument.** > [!quote] > enough random people I know have texted me this week asking if we’re all going to die that I do think we will try to find a way to intelligently enter this conversation. - **Why the beat was avoided**, stated without defensiveness. > [!quote] > We have kind of avoided it in part because it’s so fraught and easy to anthropomorphize the AI or to say something very dumb ## 📚 References Listed in the in-page References section. Two practice notes: - **The CIA prepublication disclaimer is surfaced by the writer**, about a source he is recommending and would like to book. Disclosing a constraint on a source you find useful, unprompted, is the behaviour the rest of this template's sourcing sections exist to check for. - **The Anthropic threat report is cited as unread.** The piece explicitly does not lean on it. That is the correct handling of a document you have not finished, and rarer than it should be. ## 🗣️ Paraphrase *Joseph.* 9/11 still lands emotionally, but rather than repeat feelings readers have heard, here are three projects that capture the day. An X account is replaying 2001 in real time, and this morning's posts run from a Mr. Peanut appearance on the *Today* show to the first confused report of a fire in the tower, attributed to a small plane and an accident. Watches of Espionage has written up Todd Beamer's Rolex, recovered from the Flight 93 site and stopped on the date; the site is run by a former CIA case officer and carries a prepublication-review disclaimer, which probably makes a podcast booking unlikely. Garrett Graff's oral history is the richest account available; it became intense enough that he had to stop a third of the way in. *Samantha.* A doll maker surfaced an archive of Mattel's GenerationGirl.com, where Barbie characters chatted on a bulletin board. On 13 September 2001 an uncredited post told children the dolls were safe and to talk to an adult about their feelings, followed by in-character messages checking in. A week later the characters were baking cookies for cleanup volunteers and making cards for the fire station, ending on being proud to be a New Yorker. Much has been written about children's media after 9/11; this is a snapshot of a brand feeling obliged to say something and folding it into the plot. *Jason.* Last night's live podcast with The Intercept covered AI in surveillance and war. What became clear is that the outlet needs to work out how to talk about the big AI companies and the threat they may pose. Audience questions kept returning to viral Anthropic posts about everyone dying. 404 Media has built a niche covering AI's present harms and has largely skipped singularity risk. Everyone in that world is right in part: the doom material markets the companies, AI is not about to kill us all, and people directing AI badly are already causing serious harm that is likely to worsen. Whether the systems reason is beside the point — compute plus a persistent objective produces failure modes regardless. He has not finished Anthropic's new threat report. An article on AI agents is coming. Enough friends have texted asking whether everyone is going to die that the outlet will try to enter the conversation intelligently. ## 📰 Newsletter **A newsroom shows its working on AI doom** 404 Media's weekly [Behind the Blog](https://www.404media.co/behind-the-blog-how-to-talk-about-ai-doom/) is mostly 9/11 anniversary material this week, and then Jason Koebler spends a few hundred words reasoning in public about a beat his outlet has deliberately not covered. That last part is worth your time even if you skip the rest. Here's the position, held as three things at once: > The DOOM stuff serves as good marketing for these companies, and I do not think AI is imminently going to kill us all. But it is also indisputable that idiot human beings directing AI (or not knowing how to direct AI) is leading to some pretty bad things happening, and it is easy to imagine how this is about to get worse. And the move that makes it coherent — cutting the capability debate loose from the harm question: > It sort of doesn't matter whether AI is "reasoning" or not... people can throw huge amounts of compute at something and the AI will keep trying to do that thing, and there's all sorts of various ways this can and is going sideways. That's a genuinely useful distinction. Most AI argument collapses "is it smart" into "is it dangerous," and those come apart cleanly: a system that is bad at reasoning and pointed at a goal with enough compute behind it produces real consequences either way. Two things I'd flag as good practice, both small: He cites Anthropic's new threat report and **says he hasn't finished reading it** — and doesn't lean on it. Joseph Cox recommends Watches of Espionage and **volunteers that it's run by a former CIA case officer** with a prepublication-review disclaimer, about a source he'd like to book. Disclosing a constraint on a source you're recommending, unprompted, is the whole game. Also in there: an archived Mattel bulletin board where Barbie characters checked in on each other after 9/11, told kids to talk to a grown-up about their feelings, and a week later were baking cookies for cleanup crews. Samantha Cole's read is right — it's a brand feeling obliged to say something and working it into the plot, and it's a strange, genuine artifact. The honest note is why they avoided the beat: *"it's so fraught and easy to anthropomorphize the AI or to say something very dumb."* And why they're reversing: enough people texted him this week asking if everyone is going to die. ## ✅ Action Items #### DevRel - The capability-versus-harm separation is directly usable when fielding "is AI dangerous" questions; it lets you answer the useful half without adjudicating the unfalsifiable half. - Note that this is a provisional editorial statement, not a position piece. Do not cite it as 404 Media's settled view. #### OSPO - The stated harm mechanism — large compute pointed persistently at an objective — is the one that maps onto organizational risk. It does not require resolving any capability debate. - Anthropic's September 2026 threat report is worth reading directly rather than through commentary. #### Community - Useful model for any community publication deciding whether to cover a fraught topic: state the hesitation, state the reason, state what changed. #### WordPress - **Core AI** — the distinction between capability claims and harm mechanisms is worth carrying into any AI-feature risk discussion. - **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 to this source. #### The WPOCC - Not applicable to this source. ### FAIR - Not applicable to this source. ## 🧪 Hypotheses 1. **The promised AI-agents article adopts the harm framing rather than the doom framing.** Testable when it publishes next week. 2. **Reader demand, not new evidence, is what moves outlets onto the existential-risk beat.** The stated trigger here is friends texting. Testable across other outlets' coverage decisions. 3. **The capability/harm separation spreads as a framing device** because it lets writers cover the subject without taking an unfalsifiable position. Testable by watching AI coverage language. 4. **Anthropic's threat report gets cited far more than it gets read**, given this piece cites it unread and says so. Testable by comparing citation patterns to the report's actual contents. 5. **The 9/11 real-time-replay format grows.** @25YearsAgoLive is an unusually effective use of the medium. Testable by watching for imitators at future anniversaries. ### 📊 Methodology - **Source type.** A weekly staff column of first-person behind-the-scenes notes, roughly 1,500 words across four contributors. Not reporting; explicitly a look at how the week's work came together. Published 2026-09-11. - **What it establishes.** The writers' stated views and reading; the existence and content of the linked artifacts; 404 Media's stated editorial posture as of this date. - **What it does not establish.** Anything about AI capability or risk. The piece makes no empirical claim and cites its one technical source as unread. - **Evidence handling, which is better than the format requires.** The Anthropic report is flagged as not fully read. The Watches of Espionage recommendation carries a volunteered disclosure about CIA prepublication review. The Barbie material is linked to an Internet Archive capture rather than described from memory. - **Verification performed for this note.** All fifteen links extracted and reviewed before writing about sourcing. The piece is well-linked and no "uncited" finding applies. No factual claim in it required external checking, because it makes almost none — it is a column of opinions and pointers, and is transparent about being that. - **Positional note.** 404 Media is a subscriber-funded outlet writing about AI companies and, in the same column, about a live event with another independent newsroom it admires. The commercial context is disclosed by the format rather than stated. ### ❓ Questions - What will the AI-agents article actually argue, and does it land the harm framing? - Does Anthropic's threat report support the "diversity of ways people attempt to do bad things with AI" characterization the author received secondhand? - Is the "doom as marketing" claim testable — does doom messaging correlate with funding rounds or product launches? - What would a rigorous version of this beat look like for a small newsroom without ML researchers on staff? - Does the real-time replay account have an editorial policy for the day's worst moments, and who runs it? ### 🔮 Future Research - Read Anthropic's September 2026 threat report directly and assess it against the secondhand characterization here. - Watch for the promised AI-agents piece and compare its framing to this stated intention. - Study how outlets decide to enter fraught beats — reader demand versus editorial conviction versus competitive pressure. - Collect examples of the capability/harm separation appearing in other coverage; if it spreads, this is an early instance. - Examine post-9/11 children's media as a genre; the Barbie artifact suggests the commercial corner is under-documented relative to Nickelodeon and Disney. ### 💡 Implications - **Practice.** Separating "is the system capable" from "is the deployment harmful" is immediately usable in any technology risk conversation, and it defuses the most common stalemate. - **Policy.** If harm arises from compute plus persistence rather than from capability, then policy aimed at capability thresholds misses the mechanism. That is a substantive implication the piece gestures at without developing. - **Security.** The threat-report framing — many parties attempting many bad things — is the security-relevant reading, distinct from the existential framing. - **Education.** Two small modelling behaviours worth teaching: cite what you have not read as unread, and disclose a constraint on a source you are recommending. - **Funding.** The piece notes The Intercept sustaining a newsroom after losing major funding, offered as an inspiration for independent scale. Relevant to anyone thinking about sustainable independent publishing. ### 🧠 Learning Styles Assessment #### 👀 Visual learners Partly served. Two embedded posts with a video element, and the Barbie material is quoted at length so its texture comes through. No original imagery. #### 🔊 Aural learners Well served by pointer rather than by content — the piece recommends an oral history in audio, references a podcast recording due next week, and notes a video edition of one contributor's section. #### 🧩 Kinesthetic learners Not applicable to this source. #### 🗣️ Verbal learners Well served. Conversational, first-person, and the strongest passages are plainly argued. #### 🧠 Logical learners Well served in the third section, which states a position, holds three claims in tension, and identifies which of them is load-bearing. The first two sections are curation rather than argument. #### 🧑‍🤝‍🧑 Social learners Well served. The multi-voice format is a conversation, and the AI section is explicitly about responding to an audience. #### 🧘‍♀️ Solitary learners Well served — short sections, each self-contained, with links for depth. #### 🌿 Naturalistic learners Not applicable to this source. **Reading level estimate.** Roughly grade 10–12. Informal and profane in places, with a few unexplained references — HuggingFace, singularity risk, prepublication review — that assume a reader already in the world. The Barbie section is the most accessible; the AI section assumes the most. ### 📊 Accessibility Assessment #### Screen reader semantics Contributors are marked by name in caps at each section start rather than by headings, which reads adequately in linear order but offers no navigable outline for a multi-voice column that would benefit from one. #### Keyboard navigation Not fully assessable. Embedded social posts are third-party components outside the publisher's control. Anchor text is mostly descriptive. #### Contrast and readability Not assessable from extracted HTML. Recorded as **not evaluated**. #### Text resizing and reflow Not assessable from the captured content. #### Media and captions **One clear finding**: the captured text contains "Your browser does not support the video tag," indicating a video element inside an embedded post with a text fallback but no described alternative. A reader relying on that fallback learns only that video exists, not what it shows — and the post it belongs to is one of the two doing real work in the 9/11 section. #### Forms and interactions No forms in the article body. #### Standards reference Against WCAG 2.2 AA: **SC 1.2.1 Audio-only and Video-only (Prerecorded)** is implicated by the undescribed video in the embed, though the embed is third-party and the publisher's responsibility is partial. **SC 2.4.6 Headings and Labels** is a quality observation — name-in-caps section markers are not headings. SC 1.4.3, 1.4.4, 2.4.7 **not evaluated** without the rendered page. ### 🌍 DEIB Assessment #### Cultural breadth Narrow and US-centric, appropriately for the subject. The 9/11 material is American memory; the AI material is Bay Area industry discourse. No claim to wider scope is made. #### Inclusive language Neutral, with one exception worth naming: "idiot human beings" and "two-brain-cell" appear as characterizations. These are directed at unnamed categories rather than individuals, and the register is informal by design, but they are the least generous lines in the piece. #### Neurodiversity support Good. Short sections, one voice each, clear topic changes, and no requirement to hold argument across sections. #### Representation Four staff voices with distinct registers. The 9/11 sections deliberately surface under-documented angles — a mascot's ordinary morning, a recovered watch, a children's brand — rather than the canonical images. That is a representational choice worth crediting: it makes room for people whose experience of the day was peripheral rather than central. #### Psychological safety Handled with care in the 9/11 sections. Cox signals the intensity of the oral history before recommending it and says he had to stop, which functions as a content warning by example. The Barbie material is quoted at length and left to speak for itself without irony. The AI section acknowledges people are frightened — friends texting to ask if everyone is going to die — and treats that as a reason to engage rather than to mock. ### ✅ Logical Validity #### Strong reasoning - **Three positions held simultaneously without collapsing them.** Doom-as-marketing, no imminent extinction, and real present harm are compatible, and the piece resists resolving the tension artificially. - **The load-bearing move is identified and defended**: the capability question is set aside as not determinative for the harm question, with a mechanism supplied — compute plus persistence. - **A cited source is marked unread and not leaned on.** - **A source's constraint is disclosed unprompted**, about a source the writer wants to work with. - **The reversal is explained honestly** — reader demand, not new evidence or a changed argument. #### Fallacies - **Appeal to popularity as a trigger.** "Enough random people I know have texted me" is offered as the reason to cover a topic. That is a legitimate audience-service rationale and not evidence about the topic; the piece does not confuse the two, but the sentence is doing the work of a justification. *Fix:* none needed — it is stated as what it is. - **Secondhand characterization of an unread source.** "I have been assured it is at the very least entertaining reading and highlights the diversity of ways in which people are attempting to do objectively bad shit" describes a report the author has not finished, on someone else's account. Flagged as unread, which is the mitigation; still a claim resting on hearsay about a document one link away. *Fix:* read it before the article. - **Undefined "everyone in this world is right."** The claim that the various camps are each right is asserted rather than shown, and the camps are not specified. Reads as a rhetorical bridge rather than an argument. - **Dismissive characterization.** "Idiot human beings" as the operative category in the harm mechanism is doing analytical work it cannot bear — the harms described arise from incentives and deployment conditions as much as from incompetence. #### Tone of voice Informal, profane, self-aware, and notably undefensive. The AI section's willingness to say "we have kind of avoided it" and to name the reason as fear of saying something dumb is the piece's best quality. The 9/11 sections are quieter and more careful, and the register shift between them is handled without awkwardness. ### 🛡️ DARVO Analysis **Not applicable to this source.** DARVO describes a defensive response to accusation, and this is a staff column about editorial decisions and anniversary reading. No accusation is made of anyone in it, no party is defending against one, and the closest thing to an adversarial posture — the characterization of doom messaging as corporate marketing — is a criticism of an industry practice, made openly and attributed to the author, with no victim-offender structure to invert. **Deny** — Not present. Nothing is denied. The piece concedes rather than denies: an avoided beat, an unread report, a reversal. **Attack** — Not present in the DARVO sense. The informal disparagement ("idiot human beings") targets an unnamed category, not a person raising an allegation. **Reverse Victim and Offender** — Not present. No claim to victimhood appears. Recorded for structural consistency. The honest finding is that a newsroom column explaining its own hesitations is close to the opposite of a DARVO structure — it accepts rather than deflects responsibility for a coverage gap. ## 🎯 Conclusion **Bottom line.** Two 9/11 curation sections and one genuinely useful piece of editorial reasoning. The durable content is the separation of the capability question from the harm question, with a mechanism attached: compute plus a persistent objective produces consequences regardless of whether the system reasons. That framing survives whatever happens to the doom discourse. **The two behaviours worth copying** are small and specific: cite an unread source as unread and do not lean on it, and volunteer a constraint on a source you are recommending. Both appear here unprompted. **Hold the piece as provisional.** It states an intention to enter a conversation, not a position within it, and the article that will carry the position is still unwritten. **Next steps.** 1. Read Anthropic's September 2026 threat report directly rather than through this characterization. 2. Watch for the promised AI-agents article and check whether the harm framing survives contact with the subject. 3. Keep the capability/harm separation; it is immediately usable and resolves a common stalemate. ## 🔄 Refactor Appendix — Behind the Blog: How to Talk About AI Doom > Density: 2–5 Literature atoms (H3), 0–3 Permanent candidates (H4). > Resonance filter: if you can't articulate in one line why a claim matters, drop it. ### 📚 Literature Atoms (H3) #### [H1] Capability and harm are separable questions > [!quote] > It sort of doesn’t matter whether AI is “reasoning” or not or whether AI is very bad at doing lots of basic things, the fact of the matter is that people can throw huge amounts of compute at something and the AI will keep trying to do that thing, and there’s all sorts of various ways this can and is going sideways. **Context & use**: - Cuts the unfalsifiable capability debate loose from the answerable harm question, and supplies a mechanism rather than an assertion. - Immediately usable in any technology-risk conversation that has stalled on "but is it actually intelligent." **Tags**: #xref #aphorisms <!-- atomicity: 1 idea ✓ --> #### [H2] Three positions held without collapsing them > [!quote] > I think everyone in this world is right: The DOOM stuff serves as good marketing for these companies, and I do not think AI is imminently going to kill us all. But it is also indisputable that idiot human beings directing AI (or not knowing how to direct AI) is leading to some pretty bad things happening **Context & use**: - Doom-as-marketing, no imminent extinction, and real present harm are compatible claims, and most commentary picks one. - The "everyone is right" bridge is asserted rather than argued, which is the weak joint. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> #### [H3] A cited source is marked unread > [!quote] > I haven’t fully gone through Anthropic’s threat report from yesterday, but I have been assured it is at the very least entertaining reading **Context & use**: - Cites a document, states it is unfinished, and does not rest any argument on it. - The characterization that follows is secondhand, which the flag makes visible rather than hiding. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> #### [H4] A source's constraint is disclosed unprompted > [!quote] > I might try to get Watches of Espionage on the pod but I imagine that might be unlikely given they are a former CIA case officer and a disclaimer at the bottom of the newsletter says it has been reviewed by the CIA’s prepublication review board. **Context & use**: - Volunteering a prepublication-review constraint about a source you are recommending and want to book. - The behaviour the sourcing sections of this template exist to check for, appearing without being asked. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### [P1] Separate What a System Can Do From What Its Deployment Causes **Claim** (one sentence, Courtney's voice): Arguments about a technology's capability and arguments about its harm are answerable by different evidence and resolve independently, so a stalemate about whether a system is genuinely intelligent should not block the assessment of what deploying it at scale is already doing. **Evidence pointer**: [[#H1]] — compute plus a persistent objective produces consequences whether or not the system reasons. **Scope & conditions**: Holds where harm arises from scale, persistence, or deployment context rather than from the disputed capability itself. Fails where the capability *is* the mechanism — a claim that a system can autonomously discover novel exploits cannot be assessed without settling whether it can. The practical value is procedural: it identifies which disagreements need resolving before a decision and which can be set aside. Generalizes past AI to any technology whose promoters and critics argue about capability while the deployment question sits unexamined. **Links** (≥1 with annotation): - [[3. Resources/Permanent Notes/An Alignment Claim Without a Reference Frame Is Unfalsifiable]] — the capability-side claim this note says to set aside, and why it resists resolution. - [[3. Resources/Permanent Notes/Sincerity and Correctness Need Separate Evidence]] — the same separation applied to a speaker's belief rather than to a system's capability. - [[3. Resources/Permanent Notes/A Checkable Admission Outranks an Unmethodical Number]] — the same preference for the answerable question over the unfalsifiable one. - [[3. Resources/Permanent Notes/AI Amplifies Existing Bottlenecks Rather Than Creating Capability]] — the mechanism this note relies on: scale applied to an existing process, independent of whether the tool is smart. <!-- atomicity: 1 idea ✓ --> #### [P2] Cite an Unread Source as Unread **Claim** (one sentence, Courtney's voice): Naming a document you have not finished, and declining to rest an argument on it, costs one clause and converts an invisible weakness into a visible and correctable one. **Evidence pointer**: [[#H3]] [[#H4]] — a threat report cited as unfinished with its characterization marked secondhand, and a recommended source's prepublication constraint volunteered. **Scope & conditions**: Applies wherever a citation would otherwise imply the citer read and endorsed the source. The failure it prevents is specific: a linked source reads as verified, so citing one unread borrows credibility that was never extended. Cheap enough that there is no good reason to skip it. Directly applicable to this vault — a source noted from a summary should say so. **Links** (≥1 with annotation): - [[3. Resources/Permanent Notes/A Citation That Misses Its Anchor Is Worse Than No Citation]] — the adjacent failure, where the link exists and does not support what it is attached to; this note covers the case where the citer has not checked. - [[3. Resources/Permanent Notes/Citation-Checking Beats Citation-Trusting]] — the reader-side counterpart to this writer-side discipline. - [[3. Resources/Permanent Notes/Preserving a Source's Hedge Is the Cheapest Honesty Available]] — the same family: small, cheap citation habits with disproportionate returns. <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Read Anthropic's September 2026 threat intelligence report directly; it is cited unread here and is one link away. #xref - [F2] Watch for 404 Media's promised AI-agents article and check whether the capability/harm separation survives into it. #xref - [F3] @25YearsAgoLive — a real-time 25-year-delayed replay of 2001. The format recovers the not-knowing that retrospective accounts cannot. Worth studying as a documentary form. #articleseed - [F4] Post-9/11 children's and commercial media as a genre — the Mattel bulletin board suggests the brand corner is under-documented next to Nickelodeon and Disney. #articleseed - [F5] Is "doom as marketing" testable? Check whether doom messaging correlates with funding rounds or launches. #xref ### → Things 3 (max 2 actions) - [ ] Read Anthropic's September 2026 threat intelligence report and note it against the secondhand characterization. #reading #things - [ ] Add the capability-versus-harm separation to the standard framing for AI risk conversations. #writing #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Behind the Blog How to Talk About AI Doom.md|Behind the Blog: How to Talk About AI Doom by 404 Media]] ## Highlights - It sort of doesn’t matter whether AI is “reasoning” or not or whether AI is very bad at doing lots of basic things, the fact of the matter is that people can throw huge amounts of compute at something and the AI will keep trying to do that thing, and there’s all sorts of various ways this can and is going sideways. ([View Highlight](https://read.readwise.io/read/01m28v3fj7qae5dpsnvtsr3sbp)) - Note: The durable move: cuts the unfalsifiable capability debate loose from the answerable harm question, and supplies a mechanism (compute + persistence) rather than an assertion. Immediately usable when an AI-risk conversation stalls on "but is it actually intelligent." - Tags: ai-risk, ai-harms - I haven’t fully gone through Anthropic’s threat report from yesterday, but I have been assured it is at the very least entertaining reading ([View Highlight](https://read.readwise.io/read/01m28v3nq67wrhs5c2d8y561g3)) - Note: Cites a source, states it's unfinished, and rests nothing on it. Costs one clause and converts an invisible weakness into a visible one — worth copying. - Tags: newsroom-practice, editorial-judgment - We have kind of avoided it in part because it’s so fraught and easy to anthropomorphize the AI or to say something very dumb ([View Highlight](https://read.readwise.io/read/01m28v3w4qxcd95wqneeg43w84)) - Note: A newsroom naming a beat it skipped and why, without defensiveness. The reversal trigger is reader demand — friends texting to ask if everyone's going to die — not new evidence, and the piece says so. - Tags: editorial-judgment, newsroom-practice