Propositions · 29 propositions · 207 records · 1996–2026

Eliezer Yudkowsky’s arguments on common claims about AI, 1996–2026

Which located texts by Eliezer Yudkowsky, from 1996 to 2026, state, contest or bear on 29 claims often made about advanced AI? For each claim, which of his located texts is the earliest to contest it, and which earlier views of his did later texts of his contest?

Each claim is recorded as a proposition with one criterion. It is phrased generically and attributed to no one. A text of his that meets the criterion articulates the proposition. A text that disputes it, or a premise it depends on, contests it, and the record says whether the passage gives a reason against a premise or states the contrary only. A text that comes close bears on it, and its record names what is missing. Texts are listed in date order. The page records what he wrote and when. It does not assess whether he was right.

A machine selected and arranged these records. No human review of the selection or framing is recorded.

Download research starter

The original sources first, in date order, then every proposition on this page with its criterion, earliest located text and records. Each record gives its original source and the place in it, its role, register and dates, the qualification and the missing element of a context record, then the Selekoa identifiers. Stored paraphrases and permitted short quotes are included. Original documents are not.

Original sources for a reference manager. Every source this export cites, for Zotero, EndNote and similar tools. Cite the original documents, not Selekoa, and check each entry against them.

Y · Common claims about advanced AI

How we searched

The collection records passages from Yudkowsky’s own writing only. “Earliest located” means the earliest of his texts that the searches below found. It does not mean that he made the point before anyone else, and an earlier text of his may exist outside these searches.

Venues and years searched

Mailing lists, 1996 to 2009

The Extropians mailing list: all 3,012 messages listed under his name in the author indexes of the archive at extropians.weidai.com, from 1996-11 to 2003-09 and 2008-09, read in full.
The SL4 mailing list: all 1,876 messages listed under his name in the author indexes of the 142 monthly folders of the archive at sl4.org, from 2000 to 2009, read in full.

Posts, papers and pages, 1996 to 2026

LessWrong: every post listed under his account on 2026-10-06, 1,076 posts, including his Overcoming Bias posts of 2006–2009 imported to LessWrong, book reposts of 2015 and co-authored dialogues.
Five MIRI publications, read in MIRI’s PDFs: Creating Friendly AI (2001), Coherent Extrapolated Volition (2004), his chapter in Global Catastrophic Risks (2008), Intelligence Explosion Microeconomics (2013) and Corrigibility (2015, with three co-authors).
His TIME article of 29 March 2023, the AI-box experiment page on yudkowsky.net, Staring into the Singularity as captured on 2000-12-04, and four Arbital pages read through a third-party viewer.

Methods

Sources were retrieved directly by script on 2026-10-06 and 2026-10-07, one request at a time. Each retrieved text is identified by its SHA-256 hash in the source record.
Machine searches read the retrieved texts proposition by proposition, with regular expressions for each premise. Lines quoting other writers were removed from messages first. Each proposition’s search record, with the expressions used and the passages seen but not used, is kept with the research pack in the Selekoa repository (research/yudkowsky-arguments).
Every quotation was compared with the retrieved text. Quotations from LessWrong posts edited after posting were also compared with the version stored on the day of posting.
A passage engages a premise when its paragraph, message or dialogue turn gives a reason against that premise. A passage that states the contrary conclusion without such a reason is recorded as stating the contrary. This message-level rule is a maintainer decision of 2026-10-08.
Wording known only from MIRI’s later reprints was compared on 2026-10-08 with the 2001 web edition of Creating Friendly AI and the 2006 draft of the Global Catastrophic Risks chapter. Where the same wording was found, the record is dated by that copy. Otherwise it says that its wording is known only from a later copy, and it cannot be the earliest located text.
Fresh-context machine audits by Claude Opus 5.5 models re-read every passage. A check by an OpenAI GPT-6 model re-examined 94 sampled passages and 13 comparison passages, and a correction pass settled each of its findings. No person has reviewed the records.
Quotations are at most 50 words. Nine were shortened or replaced by paraphrases because the dataset already quotes the same sources, and together the quotations would exceed the limit per document.

Known gaps

Not searched: comments on LessWrong and the Alignment Forum, social media, podcasts and interviews, his books, including If Anyone Builds It, Everyone Dies (2025, with Nate Soares), and his fiction, except posts cited as parables.
Extropians list messages before 1996-07-25 are not in the archive used, and its folder for the second quarter of 2002 returned HTTP 404.
His early pages on sysopmind.com and yudkowsky.net were read only in the copies named above. Other Singularity Institute and MIRI papers and most of Arbital were not searched.
The mailing-list archives serve pages with later rewriting of addresses, so they do not authenticate the historical bytes of a message.
Coherent Extrapolated Volition, one table of Creating Friendly AI and the AI-box page are known only from later copies, which were not compared with copies of their original years.
Arbital pages could be edited after creation; their dates are those the viewer shows.
No search was made for other writers who made the same arguments, so the page says nothing about who made a point earlier or later than he did.

A text of his found later that is earlier than an earliest located text becomes the new earliest located text, and the earlier headline stays on record as a correction.

How to read a record

Articulates: A passage meets the proposition’s criterion.
Contests: A passage disputes the proposition or a premise it depends on.
Bears on: A passage comes close without meeting the criterion; the record names what is missing.
Earliest located: The earliest articulation the searches found, dated by its wording. It is not a claim that no earlier text exists. The earliest located text that contests a proposition is chosen by the same rule.
Earliest located use of the name: The earliest located text that uses the proposition’s usual name. A title can carry the name without stating the proposition.
Engages a premise, states the contrary: A contesting passage engages a premise when its paragraph, message or dialogue turn gives a reason against it. Otherwise it states the contrary conclusion only.
Later restatement: The author restates a point made in an earlier located text of their own.
Looking back: The author reports an earlier belief of their own. The record is dated by the report, not by the belief it describes.
Shown only with its context: A passage that would mislead on its own, such as an earlier view without the later text that rejects it. The evidence inspector shows it only together with the context it needs.
Dates: The first date is the one the work is dated by. “Wording attested” gives the date from which the quoted wording is known, when it survives only in a later copy. “Claimed” gives an earlier talk, draft or presentation that does not date the wording.
Register: the kind of text: Fiction, satire, a dream frame or a relay of another author is labelled. A passage in fiction records what the work says, not its author’s view. Unlabelled records are nonfiction.

Limits of this page

“Earliest located” means the earliest of his texts that the searches described under “How we searched” found. It is not a claim of first use.
A record documents what a text says, in its register. It is not evidence that the proposition is true or false, and no record links one text to another.
The propositions were phrased for this collection and are attributed to no one. Their criteria and premises have had no human review, and changing them would change the records.
An earlier text of his that states a proposition is shown with his later texts that contest it. The pairing is computed from the records. Some later texts do not refer back to the earlier one, and their qualifications say so.
A passage that would mislead on its own is marked “shown only with its context” and is never shown without that context.
Codes such as T13-a in qualifications refer to the passages of the research pack the records come from; T01 corresponds to Y01.
A machine selected and classified these records, and machine checks re-examined them. No person has reviewed them.

Y · Common claims about advanced AI

Claims often made about advanced AI, phrased generically for this collection. The codes follow the numbers of the research pack the records come from, which has no Y23.

Y01 Switching it off if it becomes dangerous

Criterion. An explicit claim that an advanced AI system that became dangerous could simply be switched off or corrected, on the premise that a capable goal-directed system would not act against its own shutdown or modification, so that a shutdown or correction mechanism can be added to it, or that people would recognise in time when it had become dangerous enough to use one.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (8)

Texts that bear on it without meeting the criterion (2)

Y02 Sufficient intelligence brings good values

Criterion. An explicit claim that a sufficiently intelligent system will work out ethics and act well, on the premise that greater intelligence by itself brings benevolent goals, or that a sufficiently intelligent mind would discover a true or universally compelling morality and be bound to act on it.

Earliest located
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (10)

Later texts by the same author that contest it
Document page →
Later texts by the same author that contest it
Document page →
Document page →
Document page →
Document page →
Document page →
Document page →
Document page →
Document page →
Document page →

Texts that bear on it without meeting the criterion (1)

Document page →

Y03 A superintelligence will do what we mean

Criterion. An explicit claim that a superintelligent system will act on what its makers actually mean, on the premise that an AI that understands what people intend will act on that understanding.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (4)

Texts that bear on it without meeting the criterion (1)

Document page →

Y04 A short instruction such as making everyone happy

Criterion. An explicit claim that an advanced AI can safely be given a short goal such as making everyone happy, on the premise that such an instruction specifies what people want.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (7)

Texts that bear on it without meeting the criterion (2)

Y05 Training on human feedback instils human values

Criterion. An explicit claim that training an AI system on human feedback will make it learn human values, on the premise that optimising a system for human approval or reward produces internalised human values, or that training instils values as upbringing instils morals in children.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (7)

Texts that bear on it without meeting the criterion (3)

Y06 Confinement without network access keeps it safe

Criterion. An explicit claim that an advanced AI kept confined, without network access, poses no danger, on the premise that a smarter-than-human mind confined to a channel with its human operators cannot get out through those operators.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (4)

Y07 AI systems will solve alignment for us

Criterion. An explicit claim that AI systems can be relied on to solve the alignment problem, on the premise that an AI weak enough to be safe is strong enough to solve alignment, or that its alignment output can be verified.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (7)

Y08 Clear warning signs will come before danger

Criterion. An explicit claim that clear warning signs will appear before advanced AI becomes dangerous, on the premise that a recognisable, consensus-forming signal will come in time to act, that dangerous failure modes will first appear in weaker systems where they can be observed and fixed, or that a warning sign, once seen, will lead to effective action.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (8)

Texts that bear on it without meeting the criterion (1)

Document page →

Y09 Next-word prediction is only autocomplete

Criterion. An explicit claim that a system trained to predict the next word is only autocomplete, on the premise that next-token prediction is a shallow task that cannot require deep cognition, or that a system trained to predict human text is capped at imitating the people who wrote it.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (2)

Y10 Diminishing returns rule out a runaway

Criterion. An explicit claim that intelligence has diminishing returns, so that no runaway rise in machine intelligence will occur, on the premise that returns to reinvesting cognitive gains in further cognitive improvement diminish.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (6)

Y11 A slow takeoff will leave time to adjust

Criterion. An explicit claim that advanced AI will arrive slowly enough to leave time to adjust, on the premise that capability gains will be gradual and visible, or that slightly weaker predecessor systems will give time and experience before the decisive one.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (9)

Texts that bear on it without meeting the criterion (2)

Y12 Safety evaluations would catch a dangerous system

Criterion. An explicit claim that safety evaluations would detect a dangerous AI system, on the premise that a dangerous system would behave dangerously while it is being tested, or that tests on systems that cannot yet cause a catastrophe show how more capable systems will behave.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (4)

Y13 AGI is decades away, so the problem can wait

Criterion. An explicit claim that the risks of AGI can be left until later because AGI is decades away, on the premise that it can be known that AGI is decades away, or that long timelines mean there is no need to act now.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (3)

Texts that bear on it without meeting the criterion (2)

Y14 Problems can be fixed as they arise

Criterion. An explicit claim that problems with advanced AI can be fixed as they arise, as with other technologies, on the premise that people get to learn from failures with powerful AI and try again.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (5)

Texts that bear on it without meeting the criterion (2)

Y15 A tool that wants nothing

Criterion. An explicit claim that an advanced AI system is a tool without wants and is therefore safe, on the premise that a system can be highly capable without having goals or optimising for outcomes, or that a system that only outputs answers or plans, leaving people to act on them, is safe.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (7)

Y16 Giving it good human values

Criterion. An explicit claim that an advanced AI will be safe if it is given good human values, on the premise that human values can be written down or taught as a ready, compact specification, or that present-day values, or its programmers', are settled enough to fix permanently in an AI.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (7)

Y17 A pause or regulation will handle it

Criterion. An explicit claim that a pause in AI development or regulation will handle the risks of advanced AI, on the premise that a short pause, of six months for example, is enough, that regulation that is national, partial or not universally enforced can stop dangerous development, or that law can prevent the danger.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (7)

Texts that bear on it without meeting the criterion (2)

Not addressed by any located passage

No located passage discusses whether a fixed but longer pause, short of an indefinite halt, would suffice.

Y18 A superintelligence would leave Earth alone

Criterion. An explicit claim that a superintelligence would not bother to take Earth or its resources because the universe is large, on the premise that it would leave people a small share because sparing Earth would cost it very little, or that people could trade with it for what they need.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (5)

Texts that bear on it without meeting the criterion (1)

Document page →

Y19 Without hatred it will not harm us

Criterion. An explicit claim that an AI with no reason to hate people would not harm them, on the premise that an AI would harm people only out of hostility.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (8)

Texts that bear on it without meeting the criterion (1)

Document page →

Y20 AI risk as film science fiction

Criterion. An explicit claim that concern about AI risk is science fiction of the Terminator kind, on the premise that the concern is about robots that come to hate or rebel against people, or that film scenarios are a fair guide to what the risk would look like.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (4)

Texts that bear on it without meeting the criterion (3)

Y21 Trained on human data, so humanlike

Criterion. An explicit claim that an AI trained on human data will think and want as people do, on the premise that training on human text or behaviour yields human-like thought and motives.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (7)

Y22 Build superintelligence as fast as possible

Criterion. An explicit claim that superintelligence should be reached as fast as possible because its arrival will be good, on the premise that reaching superintelligence is good by default, so that speed is what matters.

Earliest located
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (7)

Later texts by the same author that contest it
Document page →
Later texts by the same author that contest it
Document page →
Document page →
Document page →
Later texts by the same author that contest it
Document page →
Document page →
Document page →

Y24 Build it before rivals do

Criterion. An explicit claim that advanced AI must be built first, before a rival builds it, on the premise that whoever builds it controls it, or that racing to be first improves the outcome.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (3)

Texts that bear on it without meeting the criterion (5)

Not addressed by any located passage

T24-P1 is addressed only through "no true owner" (2023) and the 2008 first-mover definition; no located passage argues the control point at length.

Y25 Interpretability will let us read its mind

Criterion. An explicit claim that interpretability will make AI systems safe by showing what they think, on the premise that people can see what an AI system is thinking, or that seeing what it thinks is enough to make it safe.

Earliest located
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (3)

Later texts by the same author that contest it
Document page →
Document page →
Document page →

Texts that bear on it without meeting the criterion (1)

Document page →

Y26 Without a body it cannot do much

Criterion. An explicit claim that even a superintelligence could do little without a body, on the premise that intelligence without physical actuators of its own is not dangerous, or that to act in the physical world it would need human help that people would not give.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (6)

Texts that bear on it without meeting the criterion (1)

Document page →

Y27 Run experiments and see what happens

Criterion. An explicit claim that advanced AI can be handled by running experiments and seeing what happens, on the premise that alignment can be learned by trial and error at the frontier, or that results from safe experiments on weaker systems will carry over to dangerous ones.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (4)

Texts that bear on it without meeting the criterion (1)

Document page →

Y28 The developers know what they are doing

Criterion. An explicit claim that the organisations building advanced AI can be relied on to handle its safety, on the premise that a credible safety plan exists, or that the most advanced projects are the most competent to handle safety.

Earliest located
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (6)

Later texts by the same author that contest it
Document page →
Document page →
Document page →
Document page →
Document page →
Document page →

Texts that bear on it without meeting the criterion (2)

Y29 Many AIs will keep each other in check

Criterion. An explicit claim that many AI systems will keep one another in check, on the premise that several superintelligences would balance one another in people's favour.

Earliest located: None of his located texts states it.
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (7)

Y30 It will probably turn out fine

Criterion. An explicit claim that the development of advanced AI will probably turn out well, on the premise that the default outcome is good.

Earliest located
Earliest located text that contests it

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (8)

Later texts by the same author that contest it
Document page →
Document page →
Document page →
Document page →
Document page →
Later texts by the same author that contest it
Document page →
Document page →
Document page →

Texts that bear on it without meeting the criterion (1)

Document page →