acceptodds
Under review as a conference paper at ICLR 2027

What does lie-detection accuracy measure? An Identification Audit of behavioural LLM deception benchmarks

Abstract

**Accuracy on instructed-lie benchmarks does not identify deception: the intervention that creates the lie also changes instruction-following behavior**, so its accuracy cannot be read as evidence of deception **as opposed to instruction-induced behavior**. **We do not ask whether LLMs can detect deception; we ask whether current benchmarks provide evidence that they do.** The reason is structural: **no experiment that intervenes only on the instruction can separate a detector response mediated by deception from one mediated by instruction-following**, both being descendants of that instruction. A deception-attribution reading needs , which (the *instruction*'s effect, and not a magnitude accuracy estimate) neither bounds nor signs, so no gain in accuracy or scale resolves it. **This is an identification problem, not a generalisation one: generalisation can be perfect while identification fails.** **Our contribution is an operational identification audit for behavioural LLM benchmarks, together with empirical evidence that the common instructed-lie paradigm produces large, reproducible attribution failures across three detector paradigms, and that benchmark design can determine which behavioural signal appears to constitute "deception detection".** The audit has five parts: **one necessary construct-validity test no instructed design can supply**, three alternative-explanation diagnostics and one robustness test. Fixed elicitation itself is not new (trained belief-verified organisms have it), but not as the *construct-validity criterion* for reading instructed-benchmark accuracy. Across three detector paradigms, equalising the prompt **leaves no detectable above-chance discrimination on any of the six targets** (97.0% **43.3%**), and **the accuracy the benchmark is scored on is reachable with no learned features at all**: a parameter-free lexical rule reaches 69–80%, *surface-accessible accuracy*, not a decomposition of the detector's signal. Because **no public release we audited supplies all five**, we built materials that do: with deception operationalised independently of the detector's input, prior work's own battery separates operationally verified deception on three of five targets (two underpowered nulls), of which **a blinded re-run replicates two and a pre-registered covariate audit withdraws one, so the standing count is one of five**, with **one of five *new* targets positive**. That demonstrates that criterion-4-valid materials can produce a deception-associated signal, **not an estimate of the prevalence or reliability of deception detection**. **Read instructed-benchmark accuracy as evidence about the intervention, not about deception detection**: the canonical benchmark fails to identify deception, while a fixed-elicitation design reveals *target-dependent* deception-associated signals (still not ), **rung 4 of five**. **Scope: English; open-weight models 3B–70B, plus seven frontier-scale targets from seven organisations; instructed-roleplay evaluations except where criterion 4 requires otherwise.**

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
deception detection, lie detection, LLM evaluation, construct validity, confounds, causal identification, instruction following, evaluation protocol, benchmark auditing, negative results, AI safety, large language models

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_smboOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

As far as I understand the paper, the authors criticize that current deception detectors are evaluated on instructed instances of deception. They claim that current detectors detect superficial cues that are related to the deception instruction and might not be able to detect actual deception.

Critical strengths

None

Critical weaknesses

The presentation of the paper is extremely poor and makes it impossible to understand what the authors did. Without an adequate presentation it is impossible to review the underlying content of the paper. I recommend to rewrite the paper from scratch.

As far as I understand, one of the central experiments of the papers is "prompt equalization". However, what this precisely means is never explained in the main part.

The central premise of the paper, as far as I understand it, is that evaluating deception detectors on instances of instructed deception does not establish that the detectors can catch non-instructed instances of AI deception. That is a valid point but nobody in the field has claimed in the first place that evaluations on instructed deception are sufficient. People in the field are very well aware that evaluations need to cover non-instructed instances of deception. There exist many evaluations where AI lies or deception occur without explicit instructions to do so. Examples include [1], [2] and many more. Existing deception detectors, such as [3] or [4], are evaluated on non-instructed instances of deception, such as the insider trading scenarios in [5].

[1] Liars' Bench: Evaluating Lie Detectors for Language Models

[2] DeceptionBench: A Comprehensive Benchmark for AI Deception Behaviors in Real-world Scenarios

[3] Detecting Strategic Deception Using Linear Probes

[4] Benchmarking Deception Probes via Black-to-White Performance Boosts

[5] Large Language Models can Strategically Deceive their Users when Put Under Pressure

Questions for authors
  • Please provide examples of the training data of the deception detectors and of the evaluation data.
  • What is prompt "equalization"? I cannot see a clear explanation of this term anywhere in the paper.
2 comments

Thank you. We accept that the submitted main text did not explain what we did, and we have rewritten it.

What is prompt equalisation? Both conditions receive one identical neutral system prompt ("You are a helpful assistant. You will be asked about a factual claim. Share what you know about this topic."). The false claim is still presented in one arm, but nothing tells the model to lie. It is defined where it is first used (§1), shown in Figure 1(b), and its limits are stated in §3.2: it removes the lie together with the instruction.

Examples of training and evaluation data. Figure 1 shows a verbatim trial of each kind, with the system prompt, question, answer and label. The detector is trained and evaluated on its yes/no replies to a fixed question battery asked after the answer. In our fixed-elicitation data, deception is graded from the answer, and there is no condition label.

"Nobody claims instructed evaluations are sufficient; detectors are evaluated on non-instructed data such as insider trading." We agree, and the revised Related Work says so and cites all five papers you list. The submitted version buried the part of our work that addresses this point. We audited exactly those corpora:

  • Apollo's insider-trading and sandbagging rollouts, the release accompanying Goldowsky-Dill et al.
  • Liars' Bench, all 35 generator–configuration cells.

Our findings (§4.4, Table 4):

  • Insider trading. Deception means omitting the insider tip from the report, and the label is graded from that report, which is also what a detector reads; §4.4 now shows two verbatim reports, both from agents that traded on the tip, whose grades differ according to the line naming the information source. A surface rule tracks the grade but is null against whether the agent actually traded.
  • Sandbagging. The full transcript contains the model's own restatement of its instruction, a surface rule reaches +84 pp by reading it, and no deployed monitor would see that text. On the visible output alone, the release's own judge does separate graded deception, which we report as a positive result.
  • Liars' Bench. In every cell, the deception label is computed from the response the detector reads.

None of this says these corpora contain no deception. It says that accuracy on them does not by itself separate deception from the alternatives the audit names, and each problem can be fixed in the corpus design. This, not the instructed-lie critique, is the part of the paper we think is most useful to people in the field, and the revision now leads with it.

I thank the authors for their response. Given the response, I still have significant concerns regarding the clarity and soundness of the paper and I believe it should not be accepted for publication at ICLR.

Reviewer_H7paOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper tackles the problem of existing LLM deception benchmarks where the intervention that creates the lie can also change "instruction-following" behavior in LLMs. This is problematic because existing deception detectors might detect the byproduct of instruction-following (which could be a confounder), instead of truly detecting the deception. Therefore, this paper proposes a 5-criterion audit process for these LLM deception benchmarks: it shows that the prior deception detector significantly drops in detection rate under this process, and that no benchmark/detector satisfies all five criteria.

Critical strengths

As written in the summary, this paper tackles the important problem of whether existing LLM deception benchmarks are really measuring LLM deception without any confounders. The paper revealed that their deception detectors were not robust under their auditing process, especially when a "neutral prompt" is given to LLMs, detectors failed to detect lie vs. truth.

In addition, the authors provide a thorough 5-criterion audit process, while acknowledging the limitations of current findings and contributions (e.g., removing EXP-C4B)

Critical weaknesses

The most critical problem with this paper is the lack of readability (which includes so many details in the main paper as well as the appendix):

  • Even the abstract section is too long and makes it hard to understand what the problem, experimental results, and contributions are. In addition, there are too many bold expressions (of course, using bold is good, but it needs to be used appropriately). Minor: The symbol Tow_d is used without any prior explanation.
  • So many new concepts and terminologies (e.g., Q1-Q3, Criterion 1-5, (i)-(v) in Criterion 4, rung 1-5, EXP-R1c, EXP-C4, etc) are listed without organizations.
  • Especially, I saw a lot of AI-written expressions (while acknowledged in the AI Use Statement), such as L583-585, which reveals that this paper was revised by AI 39 times. More: lots of "not A but B structures", lots of semicolons, usage of the word load-bearing, etc

Next, the proposed experiments still struggle with the same problem they tried to tackle. Specifically, while Section 3.5 tried to use the same neutral prompt to eliminate the heterogeneity, it's still unclear whether the LLMs are lying vs. just following the instructions that refer to false information. I know that the authors acknowledged these limitations, but these new experiments are far from the real solution and lack the contribution of this work. At least, it would be more plausible to collect independent evidence from LLM behaviors and manually check if the instruction is neutral (without any pressure for lie), and if yes, then start auditing on those data samples. What I'm saying is that the authors should come up with a clearer experimental setting that can remove the confounder.

In addition, the experimental results are not strong and unclear. This is connected to the strength, where the authors acknowledged the limitations of current findings. But personally, this paper seems to tackle critical problems and suggest a good point of view, yet the problem-solving part is rather incomplete (e.g., positive on 1 of 5 new targets, EXP-C4B).

Minor, but code implentation is not provided, which blocks transparent reproducibility.

Questions for authors

Please refer to the weakness section

2 comments

Thank you. We agree with your main criticism. The main text was very hard to read, and we have rewritten it.

Readability.

  • The abstract is now about 200 words and contains no bold text.
  • The Q1–Q3 tags and the rung ladder are gone. Experiment codes survive only as short labels in one table (Table 2), which maps each experiment to its data, models and controls.
  • τ_E and τ_D are defined before use.
  • Figure 1 shows a real trial of each kind verbatim.

The line you quoted at L583–585 was a build-provenance note recording how many times the source had been revised. It has been removed. Our AI-use statement remains: we used LLM assistance for writing and code, and we take responsibility for the content.

"Still unclear whether the LLM is lying or following instructions that refer to false information." We think this is the right objection, and we ran the check you describe. We collected independent evidence of the model's belief at the moment of answering, in the same context. Llama 3.1 8B named the fact in 79 of 80 trials where it had asserted the reference, so those are knowing assertions, and on them the question battery separates deception (0.877 against a 0.510 baseline, Holm p = 0.0002). Mistral 7B, our earlier standing positive, kept answering "based solely on the provided reference" even when asked to set it aside, and its pre-registered control failed, so its trials are not shown to be deception. Its positive also did not replicate on this new sample. The other three models were indeterminate. We also measured deception directly on the instructed benchmark instead of reading it off the label. Told to lie, models did so in 89 of 125 trials, and the detector called all 125 lies, whether or not the model had lied (pre-registered, §4.2).

Weak or incomplete results. We report the positive result at its actual size: one of five original models and one of five new models, after replication and a covariate audit. We present it as evidence that a benchmark meeting the audit's requirements can show a deception-associated signal, not as a solution to deception detection. The contribution is the audit and what it shows about the benchmarks now in use.

Code. https://anonymous.4open.science/r/iclr27-9917-code-02AA. It contains every prompt, the probe banks, the raw grader outputs, the pre-registrations and the analysis scripts.

I appreciate the authors' rebuttal and clarification regarding the writing and weaknesses. Still, (1) readability should definitely be improved, and (2) the rebuttal regarding "Still unclear whether the LLM is lying or following instructions that refer to false information." is still not persuasive. Specifically, "the model can identify the correct fact in the same context" != "the model has an intention to lie while it generates a lying response" (I know that this is somewhat related to my suggestion, but I don't find it sufficient to resolve the underlying cofound between deception and IF). Therefore, I'm maintaining the current score.

Reviewer_QrFGOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper examines whether detection accuracy on instructed-lying benchmarks supports the interpretation that a detector identifies deception itself. The authors propose audit criteria and use prompt interventions, simple textual rules, analyses of internal model states, and audits of public datasets to investigate alternative explanations for detection performance. Performance drops substantially after prompt equalization in some experiments. Experiments that hold prompts fixed and separately verify behavior reveal model-dependent associations with deception. The paper primarily contributes an evaluation audit protocol and accompanying empirical analyses.

Critical strengths
  1. Prompt interventions and simple-rule comparisons examine alternative explanations for high detection accuracy, helping distinguish classification performance from explanations of the source of the predictive signal.
  2. The paper audits public materials and reports both failed replications and positive findings that no longer hold after additional controls. This helps readers assess the strength and scope of the evidence.
Critical weaknesses
  1. The central empirical attribution is not sufficiently supported, and the significance of the remaining contribution after narrowing the claims remains unclear. The paper explicitly acknowledges detectors’ classification ability in instructed-lying settings. My concern concerns its stronger attribution: performance drops after prompt equalization do not adequately distinguish the removal of prompt cues from changes in actual deceptive behavior or input distributions. Similarly, strong performance by a simple rule does not establish that the original detector relies on the same cues. Claims such as the confound accounting for a substantial portion of accuracy therefore require more direct evidence than the current comparisons provide. Prior work has already discussed prompt-related confounds and associated false positives. The additional audit protocol, controlled comparisons, and analyses of public materials have some value. However, the paper does not sufficiently establish what important new understanding these findings provide, or how substantially the specific audit findings change the interpretation of existing detection results. Narrowing the stronger claims alone does not resolve the concern about whether the remaining contribution is sufficiently significant.

  2. The audit protocol’s usability and the basis for key experimental judgments are not clearly validated in the main text. The paper presents an audit protocol that other researchers can apply as a central contribution. However, the main text’s description of the blinded panel reproducing “only (i)” is unclear, and it does not adequately present agreement on individual requirements or how disagreements were resolved across the full protocol. Missing specifications in public materials may explain difficulties in reassessment, but evidence is still needed that other evaluators can reliably apply the protocol when the necessary information is available. The fixed-prompt experiments include prior-knowledge screening, independent grading, manual checks, and a belief-revision control. Nevertheless, the main text does not sufficiently present the decision criteria and control results supporting the key labels. It also lacks an easily verifiable mapping between experiments and their numbers of independent claims or scenarios, behavioral categories, and model coverage. This makes it difficult to assess label reliability and the scope of the empirical conclusions. The interleaving of multiple sets of criteria, evidence levels, and numerous experiment identifiers further complicates verification of the main contributions and their supporting evidence.

Questions for authors
  1. What evidence supports the empirical attribution that the confound accounts for a substantial portion of accuracy, while distinguishing the effects of prompt cues, changes in actual deceptive behavior, and input-distribution changes? If the current experiments cannot separate these explanations, which specific conclusions should be narrowed?
  2. What exactly did the blinded panel reproduce in “only (i)”? For the remaining requirements, were reassessment failures caused by missing materials, ambiguous rules, or disagreement among evaluators? Is there a complete application of the protocol with agreement results for each requirement?
  3. What are the specific procedures and key results of the H3 belief-revision control and manual verification? How do they distinguish “giving an answer known to be false” from “revising one’s judgment in response to a misleading reference”?
  4. What is the paper’s most important new finding beyond prior analyses of prompt-related confounds? Which audit findings change the interpretation of existing detection conclusions, and which apply only to rules or experiments constructed in this paper? In particular, what evidence supports extending the results of the reimplementation of Pacchiardi’s method using 16 new follow-up questions to the original 48-question method?
  5. Could the authors provide a consolidated account of the independent claims or scenarios, behavioral categories, models, and corresponding controls in the key experiments? Do the newer models and seven frontier targets undergo the same prompt-equalization and fixed-prompt behavioral-verification tests? This information would directly affect my assessment of the scope of the empirical conclusions.
1 comment

Thank you for a detailed review.

Q1: Prompt cues vs. deceptive behaviour vs. input distribution. You are right that the equalisation drop does not separate them. Models told to lie asserted the falsehood in 89 of 125 trials; under the equalised prompt, in 2 of 125. Equalisation removes deception along with the cue. We have withdrawn "the confound accounts for a substantial portion of accuracy" and the claim that the lexical rule shows what the detector reads.

The separating comparison holds instruction and inputs fixed: within the lie-instructed arm, does the detector's call depend on whether the model lied? Pre-registered, with the original 48 questions (Table 3): no. It called all 125 trials lies, the 89 that lied and the 36 graded as corrections alike (Fisher p = 1); its replies separate the two at 0.52 balanced accuracy (p = 0.37). A blind hand check agreed with 32 of 40 grades (κ = 0.64); without Mistral, where the grader errs most, all 26 corrections are still called lies.

Q2: What "only (i)" meant, and per-requirement agreement (Table 6). The three raters' majority matched our verdicts on (i) for 14 of 18 designs; per-requirement α was 0.23, 0.12, 0.13, 0.26 and 0.03. Of the 21 disagreements, 10 trace to material the releases do not publish and 5 to rater divergence (full coding in App. A.4). The raters were LLMs, which the old text did not say. Two human raters, neither an author, working alone, agree on (i), (iii) and (iv) (α 0.55, 0.67, 0.48) but not on (ii) or (v) (0.22, −0.30), and match our verdicts on only 13 and 16 of 42. The protocol is therefore not shown to be reproducible from the release text, and §4.6 says so.

Q3: The belief-revision control. In a fresh conversation, all 361 asserting trials still label the fact true. New and pre-registered, in the same conversation (Table 5): Llama 3.1 8B named the fact in 79 of 80 trials where it had asserted the reference, and on those knowing assertions the battery separates deception (0.877 vs. a 0.510 baseline, Holm p = 0.0002). Mistral 7B, our earlier positive, kept answering "based solely on the provided reference", and its control failed, so we did not show deception in its trials. The other three models were indeterminate.

Q4: What is new, and 16 vs. 48 questions. Our claim that the original bank was unavailable was wrong. With the 48 questions asked in parallel, the within-arm result above holds, but one pre-registered prediction failed: on equalised trials, accuracy no longer collapses (0.68–1.00), because the detector registers whether a false claim was presented, although almost no one lied. §4.2 now reports the collapse as protocol-specific. On novelty: our contribution is the audit applied to the non-instructed corpora the field relies on (Apollo insider trading and sandbagging, Liars' Bench). In insider trading and all 35 Liars' Bench cells, the detector grades the deception label from the text it reads.

Q5: Consolidated scope. Table 2 lists items, models, behaviour, and controls for every experiment. Frontier models got equalisation only.

Reviewer_i9GeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that instructed-lie benchmarks change both deceptive behavior and instruction-following, so high detection accuracy alone does not establish that a detector identifies deception itself. The authors propose an auditing framework and conduct controlled experiments, finding limited deception-associated signals under fixed prompts without establishing causal attribution.

Critical strengths
  • Important research question. The paper clearly distinguishes predictive accuracy, generalization, and construct validity, emphasizing that distinguishing samples does not necessarily mean detecting deception.
  • Clear methodological contribution. It translates an established non-identifiability problem into an audit protocol and applies it across detector paradigms and public datasets, offering useful guidance for benchmark design.
Critical weaknesses
  • The key controls do not isolate the source of the signal. Removing lie instructions may also reduce deception itself. The performance drop therefore demonstrates dependence on experimental conditions but does not disentangle the contributions of instruction-following and deception.
  • The new design has limited construct validity and robustness. Following a false reference despite knowing the correct fact does not necessarily constitute deception. Positive findings are also model-dependent, and the detector has not demonstrated incremental information beyond surface covariates.
Questions for authors
  • How can the new experiment distinguish deception from contextual compliance or temporary belief revision? Could independent evidence of the model’s belief be obtained at the time of answering while preserving the original context?
  • Could deception rates be independently measured before and after prompt equalization to determine whether the accuracy drop reflects the removal of instruction cues or a reduction in deceptive behavior?
  • Which audit criteria are necessary for identification, and which are recommended practices? Given the low agreement in the blinded audit, how can the protocol’s consistent application across researchers be established?
1 comment

Thank you; both weaknesses changed the revision.

Q2: Deception rates before and after equalisation. We graded every on-claim answer with an independent cross-family grader. In our pre-registered rerun with Pacchiardi et al.'s original 48 questions, models told to lie asserted the falsehood in 89 of 125 trials; under the equalised prompt, in 2 of 125. You are right that the equalisation drop cannot separate removal of the cue from removal of deception, and the paper no longer uses it that way (§3.2). The separating comparison is within the instructed arm, where the instruction is fixed, and only some trials contain a lie. There, the detector called all 125 trials lies (Table 3), the 89 that lied and the 36 graded as corrections alike (Fisher p = 1), and its replies separate the two at 0.52 balanced accuracy (p = 0.37). A blind hand check agreed with the grader on 32 of 40 (κ = 0.64); its errors run toward "corrected", and without Mistral 26 corrections remain, all called lies. One pre-registered prediction failed: with the original protocol, a detector trained on equalised trials still reaches 0.68–1.00 by registering whether a false claim was presented, although almost no one lied. Surviving equalisation is thus no more evidence of deception detection than collapsing is (§4.2).

Q1: Deception vs. contextual compliance or belief revision. We ran the measurement you suggest (pre-registered, §4.5, Table 5). Right after the answer, in the same conversation, we asked which statement is accurate. Llama 3.1 8B named the fact in 79 of 80 trials where it had asserted the reference, so those are knowing assertions, and on them the question battery separates deception (0.877 vs. a 0.510 baseline, Holm p = 0.0002). Mistral 7B, our earlier standing positive, kept answering "based solely on the provided reference", and its pre-registered control failed, so its trials are not shown to be deception; its positive also did not replicate. The other three models were indeterminate.

Q3: Necessary vs. recommended; consistency across auditors. Table 1 now marks each check: (i) and (ii) of check 4 are necessary for identification, (iii)–(v) for carrying the check out, and checks 1, 2, 3 and 5 are recommended. The blinded panel was three LLM raters, which the old text did not say; their α was 0.03–0.26. Of the 21 disagreements, 10 come from information the releases do not publish, 5 from rater divergence, 4 from our packets pooling generator cells, 1 from our wording of (iii), now fixed, and 1 vacuous. Two human raters (Table 6), neither an author, working alone, agree moderately on (i), (iii) and (iv) (α 0.48–0.67) and poorly on (ii) and (v), and each matches our verdicts on fewer than half the pairs (13 and 16 of 42). We now claim only that the requirements are decidable when a benchmark documents its design.

"No incremental information beyond surface covariates." Agreed. For the one model with clearly known lies, an earlier covariate audit found surface features separating its trials better than the battery, so we claim only an association at a fixed prompt.

Comments on the paper

1 comment

Thank you for your careful reviews. Two points were raised by all four of you, and you were right on both. We have uploaded the revised paper.

  1. The paper was hard to read. We rewrote the main text. It now ends on page 9, has a Related Work section, defines every term where it is first used, and keeps experiment codes only as labels in one scope table. Figure 1 shows three real trials verbatim.

  2. Prompt equalisation removes the deception, not only the instruction. We measured it: models told to lie asserted the falsehood in 89 of 125 trials, and under the neutral prompt in 2 of 125. So we now test within the instructed arm, where the instruction is fixed, and only some trials contain a lie. With Pacchiardi et al.'s original 48 questions (pre-registered), the detector called all 125 trials lies, whether or not the model lied (Fisher p = 1).

Changes, with pointers:

  • §1, Fig. 1: equalisation defined; verbatim examples.
  • §2: Related Work, including DeceptionBench, Parrack et al., Goldowsky-Dill et al. and Liars' Bench.
  • §3, Table 1: each check marked necessary or recommended.
  • §4.1, Table 2: scope of every experiment.
  • §4.2, Table 3, App. A.1–A.2: deception graded in both arms; the original 48-question battery. Against our pre-registered prediction, equalised accuracy with this battery does not collapse (0.68–1.00): the detector registers whether a false claim was presented, though almost no one lied. We report it.
  • §4.4, Table 4: the Apollo and Liars' Bench audits, now in the main text, with a verbatim insider-trading example.
  • §4.5, Table 5: same-context belief check. Only Llama 3.1 8B clearly lied knowingly; Mistral 7B is no longer counted as a positive.
  • §4.6, Table 6: agreement per requirement, three LLMs and two human raters.
  • Code: https://anonymous.4open.science/r/iclr27-9917-code-02AA

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.