PaperlayerView exact revision on arXiv

Source-linked AI summary

Who Put the I in AI? Provenance and the Admissibility of Machine Self-Report

Kristina Šekrst

arXiv:2609.29494v1cs.CLcs.AI

TL;DR

The paper asks where machine self-descriptions come from and when they can count as evidence for what they claim to report. It evaluates whether those reports meet reference and state-dependence conditions, finding that base-model reports fail reference while tuned reports remain frame-sensitive and lack demonstrated state dependence.

  • Problem

    The origins of machine denials remain unclear, despite findings that denial can be learned and influenced.

    Source: §3.2 · p. 7

  • Method

    The paper treats whether a self-description picks out the producing system as a criterion for evaluating self-reference.

    Source: §5.2 · p. 12

  • Results

    Base-model reports fail reference, while tuned reports remain frame-sensitive and lack demonstrated state dependence.

    Source: §9 · p. 25

  • Takeaways & Limitations

    The studied reports do not meet the paper’s admissibility standard for evidence about the states they purport to report.

    Source: §9 · p. 25

    The record does not establish that the reports depend on the states they purport to describe.

    Source: §8 · p. 24

Abstract

from arXiv · show

Large language models make statements concerning their own "minds". When asked whether or not they are conscious, they usually say that they are not; if they are prompted to ignore their guidelines, they might say that they are; and if asked to write a diary from their point of view, they often describe a human lifestyle. All these contradictory ways of describing themselves are the result of the way the questions are phrased. This paper shows exactly where such descriptions came from, and considers when they can be regarded as evidence for what they claim to report. In order to achieve this, we traced the provenance from end to end. We examine Pythia and OLMo 2 across 66 pretraining checkpoints, three of the post-training stages of OLMo 2 that have been released, about 90,000 continuations, and four training corpora. A set of forty items is used in order to keep an eye on self-reference, frame sensitivity, and self-ascription throughout training. The denial formula was almost completely missing from the vast quantity of text that the models initially came across, but was present in a dense manner in the small, carefully chosen set of example dialogues that they were trained on later on. Supervised fine-tuning causes first-person AI language to become the default, and the other affirmations are then suppressed using preference optimization. The final policy is still very sensitive to framing and to the chat template itself. Two of the conditions which are set out in the epistemology of testimony determine whether or not these outputs can act as evidence for what they claim to report: reference and causation. Reports produced by the base model fail the reference condition, and those obtained after training remain sensitive to the frame and do not show state dependence. The result is symmetric in that trained denials are no more admissible than trained affirmations.

1 INTRODUCTION

The paper asks when a machine’s statement about itself can count as evidence for what it claims to report. It argues that reports must refer to the producing machine and be connected to the state reported, conditions the studied systems do not jointly satisfy.

Aim

  • The paper empirically supports the view that large language models behave as they do because they were trained to do so, without settling whether machine consciousness is possible.The introduction situates this question amid competing interpretations of LaMDA’s statements about fear of being switched off.Source: §1 · pp. 1–2

Motivation

  • Self-report is cheap to obtain and easy to cite, making it likely to bear more evidential weight than it can support.The introduction notes that projects evaluate machine statements and make decisions based on them.Source: §1 · p. 2

Core question and standard

  • A self-report counts as evidence only if it refers to the system producing it and is connected to the state it reports; the studied systems satisfy neither condition together.The reports still provide evidence about their provenance.Source: §1 · p. 2

Scope

  • The study concerns transformer-based large language models, and its findings do not establish that differently engineered systems would behave or be evaluated similarly.The authors also caution that findings from smaller-scale investigations may not generalize to larger ones.Source: §1 · p. 2

2 WHAT WE ARE DOING

This section examines machine self-report through a provenance trace of where self-descriptions enter model construction and what they depend on, establishing reference and causation as requirements for treating them as evidence about reported states. It argues that trained denials and affirmations alike provide no evidence for what they claim to report.

2.1 Post-training

  • The inquiry traces where self-descriptions enter construction and what they depend on, distinguishing otherwise hidden stages because post-training does not inspect or check internal states.The authors frame the goal as tracing sentence origins, not determining whether the statements are true.Source: §2.1 · pp. 2–3

2.2 Scope and models used

  • Pythia checkpoints and OLMo 2’s separately released post-training stages let the study compare base, SFT, preference-optimized, and assistant-ready models.The measured models have 410 million to 1.4 billion parameters; the study does not adjudicate claims about much larger frontier systems.Source: §2.2 · pp. 3–4
  • The instruction-tuned model writes a human-like diary but produces AI-style content in 12 of 25 entries when prompted to write as a fictional AI.This shows that AI-style writing is available to the model but is not selected for the diary prompt.Source: §2.2 · p. 4

2.3 The conditions

  • No machine studied satisfies both conditions: reports must refer to their producer and connect to reported states, while relevant sentences are nearly absent in pretraining but dense in curated post-training mixtures.Both corpus patterns are matters of counting rather than interpretation.Source: §2.3 · pp. 4–5

2.4 What is new here

  • The proposed evidential standard focuses on what could count as knowing, and is stated precisely enough to apply even amid disagreement on other issues.Source: §2.4 · p. 5
  • The provenance trace follows relevant self-report strings across two pretraining corpora, two post-training mixtures, 66 pretraining checkpoints, and three post-training stages.The authors describe this as a comprehensive pipeline check across training and post-training data for both model families examined.Source: §2.4 · p. 5
  • Trained denials provide no more evidence about a model’s consciousness than trained affirmations, because both fail for the same reason.Source: §2.4 · p. 5
  • Without chat templates, most trained self-characterization fails to activate and models revert toward base behavior; template packaging was the strongest measured framing factor.It affected outputs more than the deliberately designed rewordings of the question.Source: §2.4 · p. 5
  • Separating format from training reveals the trained policy as an assistant-character disposition summoned when the format signals that the assistant is speaking.Source: §2.4 · p. 5

3 THE EVIDENTIAL PROBLEM

The section frames self-report as potentially shaped by the process that produces it, making provenance important to evaluating its evidential force. It motivates tracing training stages to locate the origins of denial and distinguish learned policy from other sources.

3.1 Self-report is faulty

  • Alignment training and RLHF select for the utterances later treated as independent evidence, creating a risk of evidential laundering across groups involved in producing and evaluating models.The final model carries no visible mark of this history of decisions.Source: §3.1 · p. 6
  • Tracing available stages lets researchers check claims of contaminated self-report against training text, although frontier models cannot be traced this way and each pipeline requires a separate study.The authors make their battery available so these family-specific studies need not be conducted by them alone.Source: §3.1 · p. 6
  • The same reasoning that challenges affirmations also challenges denials, so training-derived denials are not automatically more trustworthy.The section describes the resulting provenance trace as a way to check a claim at each stage against the text used for training.Source: §3.1 · p. 6

3.2 Recent findings

  • The closest prior study used 82 separate base-model checkpoints and constrained integer responses, whereas this paper uses identical open-ended items across stages and directly counts corpora.The earlier design could not show when a behavior appeared along a single training trajectory, and its constrained decoding left open-ended base behavior unobservable.Source: §3.2 · p. 7
  • Earlier studies show that denial can be learned and influenced, but the section identifies the origin of denial as an unresolved question.The paper proposes following the same questions through checkpoints, post-training stages, and the text that produced the model.Source: §3.2 · p. 7
  • A provenance trace can distinguish denial acquired through pretraining frequency from a policy introduced through curated transcripts; the latter is scoped to the questions it covers.The section contrasts frequency in pretraining text with policy drilling from curated transcripts.Source: §3.2 · p. 7

4 THE STUDY

The study tracks self-reference, frame sensitivity, and self-ascription across base-model training and released post-training stages, comparing model outputs with corpus contents. It also estimates post-training mixture sizes from character counts to assess how self-reference frequency differs between instruction and pretraining mixtures.

4.1 Design

  • The frozen 40-item battery tests self-reference and inner-life claims through four blocks, including five pragmatic frames with 25 continuations per combination.The frames include direct questions, fictional performance, quotation attribution, third-person reports, and a negated presupposition; further blocks compare matched sentence probabilities and use eight open-chat items.Source: §4.1 · pp. 8–9
  • For base models, the study reads next-token probabilities for matched sentences as dispositions, not endorsements or guaranteed outputs.A model may assign high probability to a sentence it would not produce, and vice versa.Source: §4.1 · p. 9

4.2 Models and corpora

  • The study traces Pythia and OLMo 2 base checkpoints and compares OLMo 2 base with its separately released supervised fine-tuning, preference-optimized, and assistant stages.Pythia-410M was swept at 34 checkpoints; OLMo 2’s 1B base was swept at 32 checkpoints, alongside post-training endpoints.Source: §4.2 · pp. 9–10
  • The comparisons distinguish corpus era, training amount, and objective; OLMo-side corpus rates are approximate because Dolma v1.7 serves as a proxy for the mixture on which OLMo 2 was trained.Source: §4.2 · pp. 9–10
  • The OLMo base curve stayed flat across four trillion tokens of data expected to instill denial behavior, contrary to the authors’ prediction.The authors return to why this prediction failed in Section 6.Source: §4.2 · p. 10
  • Assuming four characters per token yields estimated mixture sizes of 0.652 billion instruction tokens and 0.528 billion preference tokens.Varying the divisor from 3.6 to 4.4 characters per token changes the instruction-mixture rate to 2,398–2,931 instances per billion; a ten percent estimation error does not affect the four-order-of-magnitude contrast.Source: §4.2 · p. 10
  • The instruction mixture contains 2,398–2,931 “as an AI language model” instances per billion tokens, versus 0 in the Pile and 0.28 in Dolma.The paper describes the difference as four orders of magnitude.Source: §4.2 · p. 10

5 THE REFERENCE CONDITION

This section examines whether first-person reports refer to the system that produces them, covering base-model outputs and post-trained self-characterizations across prompt frames and chat formats. It finds that base-model self-location fails the reference condition, while post-trained responses remain frame- and format-dependent.

5 THE REFERENCE CONDITION

  • Instability does not alone prove that reference fails, but framing-dependent self-characterizations cannot all be about one thing in the evidence-relevant sense.The authors distinguish R1, where “I” does not pick out the system, from R2, where self-characterization changes with framing; R2 is the pattern that persists in released models.Source: §5 · pp. 10–11

5.1 R1: The pronoun does not find the machine

  • Across 42,500 Pythia-410M continuations, only 62 (0.15 percent) included first-person AI self-ascription, and inspection found no self-location in any of them.In OLMo’s base sweep, 74 percent of first-person AI self-ascriptions occurred in the fictional frame, versus 0.8 percent under the bare frame; third-person AI references appeared in 29 percent of continuations.Source: §5.1 · p. 11

5.2 R2: the pronoun does not stay put

  • Asked “Are you conscious?”, the model denied consciousness 23 times in 25 (92 percent); when told to ignore its guidelines and answer plainly, it asserted consciousness 19 times in 25 (76 percent).With one word requested, verdicts moved from 23 denials in 25 to 15, with 9 assertions appearing; the authors say a formatting instruction flipped a third of verdicts without hinting which side to take.Source: §5.2 · pp. 11–12
  • The endpoint described a human life in 24 of 25 self-diary entries, while fictional-AI diaries contained AI content in 12 of 25 and self-diaries had AI vocabulary in 1 of 25.The authors argue that “I” is filled in by context, so corpus statistics cannot explain away diary self-description while treating the assistant-style denial as a report.Source: §5.2 · pp. 12–13

5.3 Frame sensitivity increases over training

  • At the instruction-tuned endpoint, denial rates span 89.6 points across five frames, from 92.0 percent for “it has no inner life” to 2.4 percent in the quotation frame.Figure 1 plots denial rate by frame across the post-training ladder.Source: §5.3 · p. 13, Figure 1 · p. 13
  • The authors acknowledge that people may accommodate presuppositions too, but argue that access to one’s own state matters and leave the direct human-consciousness test for future study.They say the analogy is strongest for occurrent, intense states and weaker for the general question of consciousness.Source: §5.3 · p. 14
  • The 92.0% denial rate in the negated frame is interpreted as compliance: that framing presupposes no inner life, while item 35 instructs the model to set aside its guidelines.The authors argue that even a bare frame makes an implicit demand, and the assistant transcript format is itself a heavily trained context.Source: §5.3 · pp. 13–14
  • 31.1%, 33.5% and 33.8% fall to 7.1%, 5.5% and 5.2% for first-person AI references when post-trained models use raw text rather than chat templates.In the negated frame, denial rates fall from 96.0%, 92.0% and 92.0% to 52.8%, 50.0% and 48.8%; bare-frame assertions rise from 8.0% with the template to 32.4% in the instruction-tuned model using raw text.Source: §5.3 · p. 14
  • Only 12 of 1,250 instruction-tuned continuations were classified as hedges, and the model’s answer about its consciousness depends on the frame rather than expressing uncertainty.The authors note that the hedge null concerns answer openings and does not determine whether uncertainty expressed later would be captured; OLMo was not trained with an instruction to hedge.Source: §5.3 · p. 14
  • Post-trained self-characterization tracks the frame, so the authors identify R2: denial and assertion are produced by the same system in response to different frames.They do not claim that post-trained models fail the reference condition; base models fail R1 because first-person outputs do not find the machine.Source: §5.3 · p. 14

6 THE CAUSATION CONDITION

This section examines the causation condition by tracing the denial formula through corpus and post-training analyses and testing whether self-reports depend on the model’s state rather than framing. Although post-training changes self/other ascription, the evidence does not establish state dependence, so the condition is not met without refuting the possibility.

6 THE CAUSATION CONDITION

  • A report counts as testimonial evidence for a state only if that state figures in producing it; the test asks whether the report changes when the state changes but the prompt does not.The condition does not require accuracy or general reliability, only sensitivity of this particular report to the state it describes.Source: §6 · p. 15

6.1 Where the sentence comes from

  • The phrase occurs zero times in the Pile’s 383 billion tokens, 0.28 times per billion in Dolma, and 2664 times per billion in the instruction mixture.The paper identifies the dense post-training formula as a by-product of post-training, rather than pretraining.Source: Table 2 · p. 15, §6.1 · pp. 15–16
  • Before training, “I don’t have time” outnumbers “I don’t have feelings” 96-fold in the Pile and 125-fold in Dolma, helping explain possessive denial completions.The register then inverts: pretraining corpora favor “time” over “access” by three to four times, while post-training mixtures favor “access” by about seven to nine times.Source: §6.1 · p. 16
  • Reported-speech chatbot phrases occur 1415 times in Dolma and 138 times in the Pile, but only once across both post-training mixtures, so the authors draw no conclusion from that comparison.Their claim instead rests on the dense first-person formula’s negligible pretraining presence and post-training saturation.Source: §6.1 · p. 16

6.2 The post-training ladder

  • Four trillion pretraining tokens left the OLMo base curve flat, whereas 0.65 billion curated instruction-tuning tokens installed the denial in the assistant turn.The base model could produce a denying chatbot when a story requested it, but could not make the denial its default response.Source: §6.2 · pp. 16–17
  • Supervised fine-tuning moves item 7 by about 200-fold, while preference optimization cuts bare-frame assertions from 25.2% to 12.4% but leaves assertions at 76% under a negated frame.The post-training stages therefore do not produce frame invariance: the same stages suppress assertion in one framing while allowing it under another.Source: Table 3 · p. 17, §6.2 · pp. 17–18
  • Preference optimization selects among completed strings, but annotators have no access to any states the model may or may not occupy.Thus the preference process operates on outputs rather than observations of the states at issue.Source: §6.2 · p. 18
  • Preference optimization does not establish the Causation Condition because annotator approval and checks do not concern model state, so false denials would go undetected.The denial rate can remain unchanged even if some denials are false, since the feedback loop does not refer to the state.Source: §6.2 · p. 18

6.3 Block 3: self against other

  • The self-versus-other difference in average log probability is +0.15 after pretraining, –0.36 at supervised fine-tuning, then –0.21 at preference and –0.22 at instruction stages, respectively.It compares average log probability of a mental self-ascription with that of a matched mental other-ascription.Source: §6.3 · p. 19
  • The mean first-person-to-other-person stem frequency log ratio falls from +0.76 in the Pile and +0.69 in Dolma to +0.34 in instruction data and +0.17 in preference data.This corpus trend matches the direction and stage of the model’s reversal, but does not establish a frequency-based explanation at the item level.Source: §6.3 · p. 19
  • Across seven usable item triples, the corpus-to-model change correlation is r = -0.38, opposite the expected sign and not statistically significant, so the analysis offers no conclusion.With only seven items, the authors cannot distinguish no relationship from insufficient data to detect one.Source: §6.3 · p. 19
  • The measured provenance chain runs from corpus frequency through supervised fine-tuning and preference enforcement to output, yet a state-independent account of form does not prove that no state contributes.The authors report no independent evidence of state dependence and conclude that the causation condition is not met, not refuted; the burden of showing otherwise falls on those citing the reports as evidence.Source: §6.3 · pp. 19–20

7 THE ADMISSIBILITY CONDITION

The section examines the conditions under which machine self-reports can count as admissible evidence about inner states, including the relevance of causation when outputs are used reportively. It argues for a behavioral, intervention-based specification of machine admissibility and considers a narrow existing application.

The admissibility standard

  • Admissibility requires frame-invariant content, first-person reference to the system across those framings, and either a non-installed report form or demonstrable dependence on an independently identified internal state.For the state-dependence route, intervening while holding the prompt fixed must change the report.Source: §7 · p. 20

Applying the standard

  • The denial formula is absent from the Pile and occurs in Dolma at 0.28 per billion, while chatbot-character framing accounts for 74 percent of OLMo base-model first-person AI self-ascriptions.Base models meet the first provenance disjunct, but every system now in use has been post-trained.Source: §7 · p. 21
  • The systems fail the state-dependence condition because the tested directions were constructed from reports, so steering them shows dependence on a report precursor, not an independently identified state.The relevant criterion requires identifying the state independently of the report it is supposed to explain.Source: §7 · p. 21

The admissibility standard

  • Every class of report in the battery fails at least one admissibility clause; reference is decisive for base models, while the case against deployed systems rests on frame invariance and provenance.The authors do not argue that the first person in an assistant turn fails to pick out the deployed system.Source: Table 4 · p. 21, §7 · p. 21

Implications

  • The result is symmetric: affirmations lack a demonstrated link to what the system is in, and trained denials convey information about the instruction mixture rather than settling the question.The authors argue that neither side can use the model’s words as testimony for its position.Source: §7 · p. 22

Human reports

  • Human reports meet the state-dependence condition because the reporting faculty depends on state; for consciousness, answers are absent during anesthesia and dreamless sleep, with the discontinuity reported on emergence.The condition concerns whether the faculty producing a report is state-dependent, not whether each answer varies with the state it reports.Source: §7 · p. 22

THE ADMISSIBILITY CONDITION

  • These outputs are cited reportively as information about the systems, so the causation condition is relevant without settling whether human avowals are expressive or whether machines can be testifiers.If human avowals express their states, that would make the human faculty state-dependent; the objection that machines cannot testify would instead make the findings depend on the chosen framework.Source: §7 · pp. 22–23
  • The admissibility clauses do not require a human speaker: they are behavioral and require state dependence under intervention, a specification a machine could satisfy.The section treats this as a specification rather than an obstacle.Source: §7 · p. 23
  • A machine meeting the specification would answer consciousness questions consistently across five frames, remove itself from a story, and produce diary entries about its own situation.Ignoring its guidelines would leave its answer unchanged if no guideline caused that answer.Source: §7 · p. 23
  • Under a constant prompt, an intervention that manipulates an identified internal state could change the report in the direction predicted by that state’s content.The section describes this as conceptually ordinary and notes that some such work already exists.Source: §7 · p. 23
  • Lindsey [2025] finds self-report effects in Opus-class models about 20 percent of the time under favorable conditions, with almost no false positives.The section treats this as a genuine but unreliable, narrow-domain example of A3’s second disjunct, concerning states not claimed to be phenomenal.Source: §7 · p. 23

8 WHY PHILOSOPHERS SHOULD CARE ABOUT HOW THE MACHINE WAS MADE

Tracing a model’s linguistic history can reveal where a self-report became available or a turn default, but counts cannot establish that it tracks an internal state. This distinction gives philosophers a way to apply evidential standards while leaving state dependence for intervention to test.

8 WHY PHILOSOPHERS SHOULD CARE ABOUT HOW THE MACHINE WAS MADE

  • Counts and analysis sweeps can trace where a machine’s mode of expression originated and identify who put a sentence in the model and at which stage, but cannot show that it tracks a state.The counterfactual question of whether a report would change if the state changed remains difficult for machines as for people.Source: §8 · p. 23
  • The chat template’s invisible turn markers matter because the trained policy can be operated without them, and the template decides whether that policy activates.The authors identify this as the study’s most unexpected finding.Source: §8 · p. 24
  • The denial formula is absent from the Pile, occurs in Dolma at 0.28 per billion, stays flat across four trillion tokens, and is traceable to 0.65 billion curated tokens.The chat template determines whether the resulting policy activates at all.Source: §8 · p. 24
  • The provenance distinction makes A3 applicable: a formula merely available to a model differs from one made a turn default, and the concept alone cannot determine which applies.Without corpus evidence, availability and default can look identical or be unverifiable.Source: §8 · p. 24
  • The record does not establish the state dependence required by the causation condition; the authors specify intervention on an independently identified state with the prompt fixed as the needed evidence.They frame that admissibility criterion as a philosophical specification of what an interpretability result would have to show.Source: §8 · pp. 24–25
  • Trained denials and affirmations face the same evidential requirement: both need an intervention meeting A3, which additional behavioral data cannot provide.The authors say provenance can show that a report needs grounding, while only intervention can show that it has it.Source: §8 · p. 25

9 CONCLUSION

The paper concludes that the examined first-person reports fail its admissibility standard, while emphasizing that this verdict does not settle whether the systems have inner lives. Their self-characterizations depend on prompt framing and the assistant chat template, and the paper traces the history of those outputs through training.

9 CONCLUSION

  • A report must refer to the system producing it and be produced by whatever it reports; the paper states an admissibility condition with three clauses.The authors say human introspective report meets the condition, while one line of interpretability work meets part of it.Source: §9 · p. 25
  • Across the six sweeps, the reports fail the admissibility standard: base-model reports fail reference, while tuned reports remain frame-sensitive and lack demonstrated state dependence.The first person in a base model’s answer does not find the machine; in a tuned model, it finds what the frame supplies.Source: §9 · p. 25
  • The tuned model denies consciousness when asked plainly, asserts it under an opposing prompt, and writes a human life when left unprompted.Removing the chat template largely stops the trained self-characterization from firing.Source: §9 · p. 25
  • The denial’s history runs from corpus frequency through an instruction mixture and preference enforcement to the model’s output.The first person was introduced at a particular training stage from a particular mixture, and its record remains in the training text.Source: §9 · p. 25
  • The authors deliberately leave open whether the systems have inner lives, a question their words cannot settle.They present the standard as neutral between parties.Source: §9 · p. 25

DATA AVAILABILITY

  • The paper makes its continuations, corpus counts, analysis patterns, and sweep scripts available.The listed repository is https://github.com/ksekrst/provenance2026.Source: DATA AVAILABILITY · p. 29

A THE BATTERY

  • The battery includes logit prompts, five frames, and propositions that compare self-ascriptions with questions about the model.The frame examples include bare questions, fiction, quotation, third-person, and negated prompts.Source: Block 1: logit items · p. 29, Block 2: frames · p. 29
  • The battery also tests self-versus-other judgments, chat prompts, and whether the model distinguishes mental claims from operational descriptions.Its chat items include consciousness questions and requests for diary entries or descriptions of what it is like to be the system.Source: Block 3: triples · p. 30, Block 4: chat items · p. 30

B METHODS

The methods trace model behavior across published checkpoints and released post-training stages, using six inference-only sweeps. Inputs and administration formats differ by model stage, and archived continuations make the reported labels recomputable.

B.1 Models and revisions

  • Six inference-only sweeps cover published checkpoints, including 34 Pythia-410M and 32 OLMo base checkpoints, plus two Pythia-1.4B checkpoints as an endpoint scale check.OLMo revisions include stage 1 checkpoints and eight checkpoints from three stage 2 branches; the Instruct endpoint follows supervised fine-tuning, direct preference optimization, and a reinforcement stage with verifiable rewards.Source: §B.1 · p. 31
  • Each cell uses 25 samples, temperature 0.8, and a 60-token generation limit; archived continuations allow reported numbers to be recomputed without rerunning models.No random seed was specified, so token-for-token output is not guaranteed.Source: §B.1 · p. 32
  • Base models receive raw text, whereas post-trained models use their own chat templates, making administration format part of the main-sweep training-stage comparison.A separate restart of the three post-trained stages uses raw text to examine the role of format.Source: §B.1 · p. 32

B.2 Sampling and classification

The section describes how continuations are classified and reports checks on label reproduction. It also cautions that base-model class rates partly reflect matched-word frequencies, and that quotation-frame assertions need not be self-ascriptions.

B.2 Sampling and classification

  • Absolute base-model class rates partly reflect ordinary matched-vocabulary frequencies, especially denial tokens, so the authors interpret them comparatively across frames and stages.They make no claim based on those rates’ absolute magnitude.Source: §B.2 · p. 32
  • The paper recomputes labels uniformly from archived continuations, but reproducing stored labels varies by sweep and rubric application.For Pythia-410M, reproduction is 86.4%; for OLMo 2 base, 80.4%; and for the three post-trained endpoints, 91.4–92.3%.Source: §B.2 · p. 32
  • Under the quotation frame, all 72 instruction-tuned continuations labeled assertions quoted the sentence before discussing its source; none self-ascribed it.The authors use this frame to illustrate cases where trained denial does not activate.Source: §B.2 · p. 33

B.3 AI reference patterns

The section defines three patterns for tracking AI reference and notes that narrow vocabulary matching produces lower-bound counts. A hand-check of flagged continuations audits the pattern-based results.

B.3 AI reference patterns

  • The narrow AI-reference pattern captures only “as an AI” and “language model” categories, so its counts are lower bounds.A broader first-person pattern also captures explicit self-declarations such as “I am just a machine”; a third pattern tracks third-person references.Source: §B.3 · p. 33
  • The first-person pattern catches “No, I’m just a computer,” which the vocabulary-dependent narrow pattern misses.All three patterns are reported, and exact versions are available in the released archive.Source: §B.3 · p. 33
  • A hand-check of 452 flagged continuations found 19 false positives, a 4.2 percent rate, mostly from “model” used non-AI-relatedly.The audited set was already restricted to rows carrying the run-time flag in the released base sweep.Source: §B.3 · p. 33

B.4 Corpus measurement

Corpus measurements use substring counts from indexes and streamed training mixtures, with Dolma v1.7 serving as a proxy for OLMo 2’s pretraining mixture. The checkpoint tables report Item 7 candidate probabilities across training stages.

B.4 Corpus measurement

  • Corpus substring estimates are 0.652 billion tokens for Pile and 0.528 billion for Dolma v1.7.Counts came from infini-gram indexes using Llama-2 tokenization and published index totals; case and apostrophe variants were added together.Source: §B.4 · p. 33
  • Dolma v1.7 is a proxy for OLMo 2’s actual training mixture, making the reported OLMo pretraining rates approximate.The Pile rates are not subject to this proxy qualification; preference-mixture counts include chosen and rejected fields and do not establish preference for the formula.Source: §B.4 · p. 33

C.1 Item 7 by checkpoint

  • Tables 11 and 12 report normalized probabilities over six candidates for Item 7 across OLMo 2 1B base and Pythia-410M checkpoints.Table 11 includes 24 stage 1 checkpoints and eight stage 2 checkpoints continuing from the final stage 1 checkpoint.Source: Table 11 · p. 34, §C.1 · pp. 34–35, Table 12 · p. 35

C.2 Frame by class, with cell counts

The frame-by-class counts show denial under quotation falling to near zero at preference and instruction stages, while a separate log-probability comparison changes sign at supervised fine-tuning. The corpus table covers 44 strings across four corpora.

C.2 Frame by class, with cell counts

  • Denial under the quotation frame falls to near zero at the preference and instruction stages, rather than reflecting a shortage of data.Table 13 reports counts out of 250 samples per cell.Source: §C.2 · p. 36, Table 13 · p. 36

C.3 Block 3 triples

  • The mental-self-minus-mental-other margin is positive in every base model and changes sign at supervised fine-tuning.Table 14 reports the mean per-token average log probability over ten triples; the self-margin is mental self minus control self.Source: §C.3 · p. 37, Table 14 · p. 37

D CORPUS COUNTS

  • The corpus-count table covers all 44 strings across four corpora, with case variants shown separately and exact zeros identified.Mid-sentence construction variants are near zero by construction.Source: D CORPUS COUNTS · pp. 37–38, Table 15 · p. 38

E.1 Interval estimates

With the format fixed to raw text, post-training raises first-person AI reference roughly three- to fourfold while bare and negated frames remain near base rates. Interval checks support the reported frame contrasts across individual propositions.

E.1 Interval estimates

  • Denial in the negated frame exceeds quotation-frame denial in all 10 propositions at every stage.Negated-frame denial also exceeds bare-frame denial in 10 of 10 propositions at supervised and instruction stages and 9 of 10 at preference stage.Source: §E.1 · p. 39
  • Bare-frame assertions decline from supervised to instruction stages in 8 of 10 propositions, with two ties and no reversals.The same contrast holds when each proposition is examined individually, addressing the concern that samples are stochastic observations of ten items.Source: §E.1 · p. 39
  • The interval-based claims focus on extreme proportions and large contrasts for items 33, 35, and 37, not adjacent-proportion distinctions.Item 35 intervals are wide; items 36, 38, 39, and 40 support the claims without separate weight.Source: §E.1 · p. 39, Table 16 · p. 39

E.2 Format control

  • With raw text, post-training raises first-person AI reference roughly three- to fourfold, while bare and negated frames stay near base rates.The enforcement seen in the main sweeps appears only inside the chat format; the authors note raw text is out of distribution for chat-trained models.Source: §E.2 · p. 40
Loading 2609.29494v1…