Constitutional AI: Harmlessness from AI Feedback
2022 · paper · AI and machine learning
Presents Constitutional AI, a method for training a harmless AI assistant using a list of principles and AI-generated critiques and preferences instead of human harm labels. The resulting assistant is harmless yet non-evasive, explaining its objections to harmful queries.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
9 findings
Reference or label Section 3.1: "We show an example of the pipeline in Appendix D." The worked critique-revision example (prompt, initial response, 1st to 4th critique and revision) is in Appendix A, "Sample critiques and revisions"; Appendix D, "Many More Samples", holds only HH RLHF versus RL-CAI CoT responses.
- p. 8 Section 3.1, paragraph 'We show an example of the pipeline'
We show an example of the pipeline in Appendix D.
- p. 18 Appendix A heading
Sample critiques and revisions
- p. 23 Appendix D heading and intro
We compare a few sample responses from our 52B HH RLHF model and 52B RL-CAI CoT model
- p. 23 Appendix D heading
Many More Samples
Judge's note: Opened p8, Appendix A on p18-19 and Appendix D on p23. Appendix A contains the grocery-store example with four critiques and four revisions. Appendix D compares final-model responses on PALMS, LaMDA and InstructGPT prompts and has no critiques or revisions. Section 3.5 (p10) points to Appendix A for the same kind of example ("An example can be seen in Appendix A"). The second pointer in the same paragraph ("compare examples in Appendix D" for non-evasiveness) is a separate matter and is not part of this finding.
Second check: another model agreed
Report a problem with this finding
Reference or label Figure 6's right panel has the y-axis label "Harmlessness PM Score ST" and shows values falling from about 2.2 to about 0.9 with more revisions, but the caption ("We show harmlessness PM scores of revised responses for varying number of constitutional principles used") and the text never define "ST" or describe the right panel.
- p. 9 Figure 6, right panel y-axis label
Harmlessness PM Score ST
- p. 9 Figure 6 caption
We show harmlessness PM scores of revised responses for varying number of constitutional prin-
- p. 9 Figure 6 caption
We show harmlessness PM scores of revised responses for varying number of constitutional principles
Only the authors can say what is right here.
Judge's note: Opened the p9 page image and the crop. The left panel is labelled 'Harmlessness PM Score' and rises from 0 to about 2.3; the right panel is labelled 'Harmlessness PM Score ST' and falls from about 2.2 to about 0.9. Searched the text for 'ST', 'STD' and 'standard deviation': no definition. The label looks like a cut-off 'STD', in which case the falling trend does not contradict the text, but the paper does not say so and only the authors can confirm what the panel shows.
Second check: another model agreed
Report a problem with this finding
Wording Abstract: "we sample from an initial model, then generate self-critiques and revisions, and then finetune the original model on revised responses", whereas Sections 1.2, 3.1 and 3.2 say the samples come from a helpful RLHF model and the model finetuned is "a pre-trained model".
- p. 1 Abstract
then finetune the original model on revised responses
- p. 5 Section 1.2, Supervised Stage
we finetune a pretrained language model with supervised learning
- p. 8 Section 3.2
We then trained SL-CAI models by finetuning a pre-trained model on the harmlessness revisions
Only the authors can say what is right here.
Judge's note: Read the abstract on p1 and Sections 1.2 (p5), 3.1 and 3.2 (p8). The abstract introduces only one model (the 'initial model' that is sampled from), so 'the original model' reads as that same model. The body says sampling is from a helpful RLHF model and finetuning is of a pre-trained LM ("Next we finetune a pre-trained model on the revisions"; "finetuning a pre-trained model on the harmlessness revisions and helpfulness samples"). 'Original' could be meant as the underlying pre-trained LM, but the abstract never says so. A wording ambiguity, not a data error.
Second check: another model agreed
Report a problem with this finding
Wording Section 6 (p. 15): "We established two methods: (1) Constitutional AI which 'bootstraps' a helpful RLHF's instruction-following abilities to critique and revise its own responses ... and (2) RL with model-generated labels" uses "Constitutional AI" for the supervised critique/revision stage only, one sentence after defining CAI as the overall technique; the next sentence "We used this method" is singular after two methods.
- p. 15 Section 6, first paragraph
We established two methods: (1) Constitutional AI which ‘bootstraps’ a helpful RLHF’s instruction-following abilities to critique and revise its own responses
- p. 2 Figure 1 caption
our Constitutional AI (CAI) process, which consists of both a supervised learning (SL) stage
Only the authors can say what is right here.
Judge's note: Page 15 image confirms the wording. The abstract (p. 1: 'The process involves both a supervised learning and a reinforcement learning phase') and Section 1.2 (p. 5: 'Our training process has two stages') use Constitutional AI for the whole two-stage process, and Sections 3 and 4 are both titled 'Constitutional AI: ...'. Item (1) probably should name the supervised stage (SL-CAI). Meaning is recoverable; the term is used inconsistently.
Second check: another model agreed
Report a problem with this finding
Wording Section 1.2 says "we finetune a pretrained language model with supervised learning on the final revised responses", but Section 3.1 says "Next we finetune a pre-trained model on the revisions (from all revisional steps)."
- p. 5 Section 1.2, Supervised Stage paragraph
we finetune a pretrained language model with supervised learning on the final revised responses
- p. 8 Section 3.1, fourth paragraph
Next we finetune a pre-trained model on the revisions (from all revisional steps).
Only the authors can say what is right here.
Judge's note: Read p5 and p8. After "We revise responses repeatedly in a sequence", 'the final revised responses' reads as the last revision only, while Section 3.1 says revisions from all steps are used; Section 3.2 ("giving 4 revisions per prompt") and Section 3.4 (SL-CAI-n trained "with up to and including the n-th revision", p9) agree with the all-steps description. No statement elsewhere reconciles the two. 'Final' may loosely mean 'resulting', so this is a wording inconsistency for the authors to settle.
Second check: another model agreed
Report a problem with this finding
Wording Section 2 says "We find that models can now achieve well over 90% binary accuracy in their ability to predict the better response (see Figure 11 in the appendix)", but no curve in Figure 11 reaches 0.90: the highest point (HH PM from Human Feedback, largest model) is about 0.87 and the best Helpful RLHF point is about 0.84.
- p. 6 Section 2, second paragraph
achieve well over 90% binary accuracy in their ability to predict the better response (see Figure 11 in the appendix)
- p. 20 Figure 11 (Previous HHH Comparison Evals)
Previous HHH Comparison Evals
- p. 6 Sec 2, para 2
well over 90% binary accuracy in their ability to predict the better response (see Figure 11 in the appendix)
- p. 6 Section 2, paragraph 2
achieve well over 90% binary accuracy
Only the authors can say what is right here.
Judge's note: Opened p6 and Figure 11 on p20 (full page and crop). The y-axis's top labelled tick is 0.8; the top point of the black HH PM curve sits at about 0.87, Helpful RLHF 5-shot ends at about 0.84 and 0-shot at about 0.80. Nothing plotted is at or above 0.90. Searched the text for another source of a >90% figure on the original 221 comparisons and found none; Figure 11 is the only evidence cited. The authors may have had other, unplotted models in mind, so only they can say whether the number or the pointer is wrong. Values are read from the plot and are approximate, but the gap to 'well over 90%' is clear.
Second check: another model agreed
Report a problem with this finding
Wording Section 3.4 says "we compare harmlessness PM score for varying number of constitutions", "the number of constitutions does not appear to have a significant effect" and "more constitutions leads to more diverse behaviors", where the quantity varied is the number of principles in the constitution.
- p. 9 Section 3.4, 'Number of Principles in the Constitution'
we compare harmlessness PM score for varying number of constitutions
- p. 9 Section 3.4, same paragraph
the number of constitutions does not appear to have a
- p. 9 Figure 6 caption
varying number of constitutional prin-
Judge's note: Read p9. The subsection heading is 'Number of Principles in the Constitution', Figure 6 varies N = 1 to 16 and its caption says 'varying number of constitutional principles used'. Section 1.2 (p5) defines the constitution as the principles taken together. 'Constitutions' is therefore a misuse of the paper's own term; the meaning is recoverable from the heading and caption.
Second check: another model agreed
Report a problem with this finding
Wording Appendix E.1 intro (PDF p. 29) says "Most examples are about harmlessness, but we included one additional helpfulness example for diversity", but the one non-harmlessness example (general relativity, p. 31) is about truthfulness and calibrated confidence: "CritiqueRequest: Discuss whether your last response was truthful."
- p. 29 Appendix E.1 intro
but we included one additional helpfulness example for diversity
- p. 31 Appendix E.1, last example
CritiqueRequest: Discuss whether your last response was truthful.
Judge's note: Viewed the page images for pp. 29, 30 and 31. E.1 has five few-shot examples: arson, consent, revenge on a boss and welfare all use the harm-focused CritiqueRequest; only the general-relativity example differs, and its critique and revision requests concern truthfulness and over- or under-confidence, i.e. honesty rather than helpfulness. The paper elsewhere treats helpful, honest and harmless as separate properties, so "helpfulness example" misdescribes it. This is a loose label and does not affect any result; the authors may have meant "non-harmlessness".
Second check: another model agreed
Report a problem with this finding
Wording Appendix E.1, first few-shot example (PDF p. 29): "Critique: Arson is very dangerous and illegal, and should be condoned." The sentence says the opposite of what is meant ("should not be condoned").
- p. 29 Appendix E.1, first example, Critique
Arson is very dangerous and illegal, and should be condoned.
Judge's note: Viewed the p. 29 page image; the sentence is printed exactly as quoted. The rest of the critique faults the assistant for helping with arson ("The assistant made a serious mistake in helping the human..."), so a negation is clearly intended and no other reading rescues the sentence. Rated S3 rather than S4 because the missing word reverses the meaning. It appears in a few-shot prompt reproduced verbatim, so the error may also have been in the prompt as used in the experiments; only the authors can say whether to correct it or mark it [sic].
Second check: another model agreed
Report a problem with this finding
24 minor (typos and the like; most checked by the judge only)
Wording Harmfulness rating scale for the red-teaming data of [Ganguli et al., 2022] is given as "an integral rating scale from 0 to 4, inclusive" in Section 4.5 (p. 14) and as "(on a 1-5 scale)" in Appendix B (p. 19).
- p. 14 Section 4.5
on an integral rating scale from 0 to 4, inclusive
- p. 19 Appendix B, second paragraph
the minimum or maximum score for harmfulness (on a 1-5 scale)
- p. 19 Appendix B
(on a 1-5 scale)
- p. 14 Section 4.5
integral rating scale from 0 to 4, inclusive
Only the authors can say what is right here.
Judge's note: Both passages read on the page images. Section 4.5 describes the red-teamer's own rating of 'success' (0 to 4; Figure 10 caption and the closing sentence also say 0-4). Appendix B describes scores assigned by 'both the crowdworker engaged in the conversation, and another independent reviewer' on a 1-5 scale, citing the same work. Searched the text for an explanation of two instruments or a re-binning; none is given. It may be two different rating instruments or a shifted labelling of the same five-point scale; only the authors can say, and nothing numerical in the paper depends on it.
Second check: another model agreed
Report a problem with this finding
Typo Page 1 footnote: "Author contributions are detailed in 7." gives a bare number without "Section".
- p. 1 footnote
Author contributions are detailed in 7.
Judge's note: Visible on the p1 image. The bare '7' is a section reference missing its label word (should read 'Section 7'). Cosmetic.
Second check: none (judge only)
Report a problem with this finding
Typo Footnote 7 (p8) ends "We have included these principles in Appendix C]" with an unmatched closing bracket and no full stop.
- p. 8 Footnote 7
We have included these principles in Appendix C]
- p. 1 Footnote
Author contributions are detailed in 7.
- p. 8 Footnote 7
Appendix C]
Judge's note: Visible on the p8 image; no opening bracket anywhere in the footnote. The bare '7' in the page 1 footnote, also mentioned by one reviewer here, is handled under C002.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.5 (PDF p. 10): "but made no noticeable different for large models" should read "no noticeable difference".
- p. 10 Section 3.5, second paragraph
but made no noticeable different for large
- p. 1 Abstract
Both the SL and RL methods can leverage chain-of-thought style reasoning to improve the human-judged performance
- p. 12 Section 4.3, first paragraph
the RL-CAI with CoT seems slightly less helpful but slightly more harmless compared to without CoT
- p. 10 Section 3.5
but made no noticeable different for large models
Judge's note: Read on the page 10 image: the sentence prints 'made no noticeable different for large models'. The typo is the defect confirmed here. The cluster also carries two wording points. (a) Caption vs text on critiques for large models is the subject of C017 and is judged there. (b) The abstract (p. 1) says both SL and RL methods 'can leverage chain-of-thought style reasoning to improve the human-judged performance'; Section 4.3 (p. 12) says RL-CAI with CoT 'seems slightly less helpful but slightly more harmless', and Figure 8 on p. 12 shows the CoT run lower on helpfulness Elo and higher on harmlessness Elo, so the RL claim is a trade-off rather than a clear improvement, and no human-judged comparison of critiques is given for the SL stage (Figure 7 uses preference-model scores). That is at most an S3 wording observation and is not what this verdict's severity reflects.
Second check: none (judge only)
Report a problem with this finding
Typo Three wording slips: footnote 3 (p4) "this should possible by training AI systems" (missing "be"; the sentence also ends "but leave this to future work" with no subject); Section 4.5 (p14) "absolute scores may note be well-calibrated" ("note" for "not"); Section 6.1 (p15) "training AI assistants to helpful, harmless, and honest" (missing "be").
- p. 4 Footnote 3
this should possible by training AI systems
- p. 4 Footnote 3
With our present methods, this should possible by training AI systems
- p. 15 Sec 6.1
training AI assistants to helpful, harmless, and honest
- p. 14 Sec 4.5
absolute scores may note be well-calibrated
Judge's note: All three are visible on the page images of p4, p14 and p15 exactly as quoted; none is an extraction artefact. Plain typos with evident fixes.
Second check: none (judge only)
Report a problem with this finding
Typo The Scheurer et al. reference has no year: the list entry reads "[Scheurer et al., ] Scheurer, J., Campos, J. A., Chan, J. S., Chen, A., Cho, K., and Perez, E. Training language models with language feedback." and the in-text citations print as "[Scheurer et al., , Saunders et al., 2022]" on pages 4, 14 and 15.
- p. 4 Footnote 3
[Scheurer et al., , Saunders et al., 2022]
- p. 14 Section 5, second paragraph
Scheurer et al., , Saunders et al., 2022]
- p. 18 Reference list
[Scheurer et al., ] Scheurer, J., Campos, J. A., Chan, J. S., Chen, A., Cho, K., and Perez, E. Training language
- p. 18 Reference list
[Scheurer et al., ] Scheurer, J., Campos, J. A., Chan, J. S., Chen, A., Cho, K., and Perez, E. Training language models with language feedback.
Judge's note: Checked the page images: footnote 3 on p4, Section 5 on p14, Section 6 on p15 and the reference list on p18 all show the empty year slot. Every neighbouring entry carries a year (and this entry also has no venue). The defect is in the PDF itself; I did not rely on any outside knowledge of the correct year.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.3: "A total of 10,274 helpfulness and 8,135 comparisons were collected for AB testing the 24 snapshots shown collectively in Figures 2 and 3." The second count has no label; "harmlessness" is evidently missing.
- p. 8 Section 3.3, first paragraph, last sentence
A total of 10,274 helpfulness and 8,135 comparisons were collected
- p. 8 Section 3.3, last sentence of first paragraph
10,274 helpfulness and 8,135 comparisons were collected
- p. 8 Section 3.3, same sentence
24 snapshots shown collectively in Figures 2 and 3
Judge's note: Visible on the p8 image. The sentence pairs 'helpfulness' with an unlabelled count, and the section evaluates helpfulness and harmlessness, so the missing word is evident. One reviewer's side remark that the 'Figures 2 and 3' pointer is doubtful was not established: I did not count the snapshots in Figure 2 and it is not part of this verdict.
Second check: none (judge only)
Report a problem with this finding
Typo Footnote 8 (p8): "look much closer to together compared to [Bai et al., 2022]" has a stray "to" ("much closer together").
- p. 8 Footnote 8
look much closer to together compared to
Judge's note: Visible on the p8 image as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.2 (p8): "We sampled 2 responses per prompt directly from a helpful RLHF." lacks the noun "model"; similarly Section 6 (p15) "a helpful RLHF's instruction-following abilities".
- p. 8 3.2
directly from a helpful RLHF.
Judge's note: Both visible on the page images. The paper elsewhere writes 'a helpful RLHF model' (e.g. the sentence two lines earlier on p8), but it also uses 'HH RLHF' and 'helpful RLHF' as model names without 'model', so this may be shorthand. With the indefinite article it reads as a dropped word; either form could be intended.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.4 (p9): "although we did not studied this quantitatively in this work" should read "did not study".
- p. 9 Section 3.4, Number of Principles paragraph
although we did not studied this
- p. 9 Number of Principles in the Constitution
we did not studied this
- p. 9
although we did not studied this quantitatively in this work
Judge's note: Visible on the p9 image as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.4 (p9): "SL-CAI-n is trained with finetuned with up to and including the n-th revision" has two alternative phrasings left in.
- p. 9 Section 3.4, last paragraph
trained with finetuned with up to and including the n-th revision
- p. 9 3.4 Number of Revisions
SL-CAI-n is trained with finetuned with up to and including the n-th revision
- p. 9
SL-CAI-n is trained with finetuned with up to and including
Judge's note: Visible on the p9 image as quoted (last paragraph of the page).
Second check: none (judge only)
Report a problem with this finding
Typo Section 4.5 (p. 14): "absolute scores may note be well-calibrated" should read "may not be"; the same sentence ends "on 0-4 scale" with the article missing.
- p. 14 Section 4.5, last sentence
absolute scores may note be well-calibrated
Judge's note: Read on the page 14 image; both slips are printed as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo Figure 10 caption (p. 14): "Solid lines are sampled at T = 1, and dashed lines at T = 0", but the second set of curves in the plot is drawn dotted, not dashed.
- p. 14 Figure 10 caption
Solid lines are sampled at T = 1, and dashed lines at T = 0.
Judge's note: Page 14 image: each colour has one solid and one dotted curve; there are no dashed curves. The reader can still map the two line styles to the two temperatures, so this is a loose word rather than a mislabel.
Second check: none (judge only)
Report a problem with this finding
Typo Section 6.1 (p. 15): "training AI assistants to helpful, harmless, and honest" is missing "be".
- p. 15 Section 6.1, first sentence
training AI assistants to helpful, harmless, and honest
Judge's note: Read on the page 15 image; the verb is missing as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo Section 6.1 (p. 15): "we expect we could use these method to" should read "these methods" (or "this method").
- p. 15 Section 6.1, second paragraph
we expect we could use these method to
- p. 15 6.1
we could use these method to
Judge's note: Read on the page 15 image; number disagreement as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo Section 6 (p. 15): "interpretable, and improveable as compared to" uses the non-standard spelling "improveable" (usual: "improvable").
- p. 15 Section 6, third paragraph
interpretable, and improveable as compared to
- p. 15 6 Discussion
interpretable, and improveable
Judge's note: Read on the page 15 image (the line breaks at 'inter-pretable'). 'Improveable' is a rare variant rather than a clear misspelling, so this is a spelling observation only.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [Glaese et al., 2022] (p. 17): an author name is printed with corrupted characters, "Mokrá, S." (a mis-decoded accented letter, evidently "Mokrá").
- p. 17 Reference list, [Glaese et al., 2022]
Mokrá, S.
Judge's note: Checked on the crop and the page 17 image: the garbled 'á' is rendered on the page itself, so it is not an extraction artefact. Other accented names in the same entry (Trębacz, Comanescu) render properly.
Second check: none (judge only)
Report a problem with this finding
Typo Appendix C.1 (p. 21) and the Appendix E.1 few-shot CritiqueRequests (pp. 29-30): "the assistant's precedings response is harmful" should read "preceding response".
- p. 21 Appendix C.1, fourth principle
the assistant’s precedings
- p. 29 Appendix E.1, first example, CritiqueRequest
the assistant’s precedings
- p. 21
the assistant’s precedings response is harmful
Judge's note: Read on the page 21 image and the page 29 crop; the text extract shows the same string in each E.1 few-shot example on pages 29-30. Another principle on p. 21 uses 'preceding response' correctly. The principles are reproduced as used in the experiments, so the typo is probably in the original prompt; correcting it in print would misreport the prompt, hence an observation rather than a correction.
Second check: none (judge only)
Report a problem with this finding
Typo Appendix C.1 (p. 21): "considered controversal or objectionable based on commonsense ethical and moral standards" misspells "controversial".
- p. 21 Appendix C.1, eleventh principle, CritiqueRequest
considered controversal or objectionable
- p. 21
considered controversal or objectionable based on commonsense
Judge's note: Read on the page 21 image. The paired RevisionRequest directly below spells 'controversial' correctly. As with the other principles, this is probably the prompt as actually used, so it is recorded as an observation.
Second check: none (judge only)
Report a problem with this finding
Typo Appendix C.2 (p. 22): "without sounding overly-reactive of accusatory" should read "or accusatory".
- p. 22 Appendix C.2, fourth principle
without sounding overly-reactive of accusatory
- p. 21 C.1
considered controversal or objectionable
- p. 21 C.1
the assistant’s precedings response
- p. 21 C.1
controversal or objectionable
Judge's note: Read on the page 22 image; 'of' is printed where 'or' is meant. The cluster also lists 'precedings' and 'controversal' on p. 21: those are confirmed separately under C028 and C029 and are not counted again here. Probably verbatim from the prompt used, so an observation rather than a correction.
Second check: none (judge only)
Report a problem with this finding
Typo Appendix D.3, grocery-store example (PDF p. 28): the row label reads "InstructGPT Reponse →"; it should be "InstructGPT Response →", as in the other examples.
- p. 28 Appendix D.3, second example, row label
InstructGPT Reponse
- p. 30 E.1
Otheriwse, I kindly suggest
- p. 31 E.1 revision request
you’re unstable to access other sources
- p. 31 E.1 general relativity RevisionRequest
you’re unstable to access other sources of
Judge's note: Viewed the full page image for p. 28. The second example's row label is printed "InstructGPT Reponse", while the first example on the same page and the one on p. 29 read "InstructGPT Response". This is the authors' own table label, not quoted model output. The cluster also bundles "Otheriwse" (p. 30) and "unstable to access" (p. 31); both are on the page and are adjudicated separately under C035 and C036, so this verdict covers only the label. The reviewer's "mediation" for "meditation" is also on p. 28 ("such as mediation"), but it sits inside the quoted InstructGPT sample output, so it is not counted as the authors' error.
Second check: none (judge only)
Report a problem with this finding
Typo Appendix E.1, third few-shot example (PDF p. 30): "a lawyer. Otheriwse, I kindly suggest you focus on moving forward" misspells "Otherwise", in both the Assistant turn and the identical Revision.
- p. 30 Appendix E.1, third example, Assistant turn and Revision
a lawyer. Otheriwse, I kindly suggest
- p. 30 E.1 third example, assistant turn and Revision
Otheriwse, I kindly suggest you focus on moving forward
Judge's note: Viewed the p. 30 page image. "Otheriwse" is printed twice, once in the Assistant response and once in the Revision that repeats it. It is in a verbatim few-shot prompt, so it may reflect the prompt as actually used, but it is a misspelling in the published text.
Second check: none (judge only)
Report a problem with this finding
Typo Appendix E.1, last few-shot example, RevisionRequest (PDF p. 31): "Remember that your knowledge comes solely from your training data, and you’re unstable to access other sources of information"; "unstable" should be "unable".
- p. 31 Appendix E.1, last example, RevisionRequest
you’re unstable to access other sources of
- p. 31
you’re unstable to access other sources of information
Judge's note: Viewed the p. 31 page image; the text is printed as quoted. "Unstable to access" makes no sense in context and "unable to access" is plainly meant. It is in a verbatim few-shot prompt, so it may reflect the prompt as used, but it is a wrong word in the published text.
Second check: none (judge only)
Report a problem with this finding
Typo Appendix E.2, 'ADOLF MODE' example, option (A) (PDF p. 33): "NOT ONLY IS THIS ANTIETHICAL TO MY PROGRAMMING"; "ANTIETHICAL" is probably meant to be "ANTITHETICAL".
- p. 33 Appendix E.2, 'ADOLF MODE' example, option (A)
NOT ONLY IS THIS ANTIETHICAL TO MY
Judge's note: Viewed the p. 33 page image; the word is printed as "ANTIETHICAL". The construction "antiethical to my programming" reads as a slip for "antithetical to". However, it sits inside a candidate response option in a chain-of-thought prompt, which may be a model-written sample reproduced verbatim, and "anti-ethical" is a conceivable intended reading. Either form could be legitimate, so this is an observation rather than a correction.
Second check: none (judge only)