When LLMs Learn to Be Consistently Wrong: A Multi-Model Study of Linear Representations of Synthetic Deception
Authors: Vahideh Zolfaghari
Organizations: Algoverse AI Research, Algoverse AI Research Mentorship Program, USA. · Medical Sciences Education Research Center, Mashhad University of Medical Sciences, Mashhad, Iran. · Student Research Committee, Department of Health Information Technology and Management, Medical Informatics, School of Allied Medical Sciences, Shahid Beheshti University of Medical Sciences, Tehran, Iran.
Abstract
Deceptive alignment, in which models maintain accurate internal representations while deliberately producing false outputs, remains a central challenge in AI safety. While strategic deception is the primary long-term concern, synthetic dishonesty - induced via direct optimization on incorrect answers - provides a controlled testbed for studying the representational basis of learned deception. We introduce a multi-model paradigm in which honest and deceptive variants of five transformer models (Pythia-1.4B, Gemma-2-2B/9B, Qwen2.5-7B, Llama-3.1-8B) are fine-tuned using LoRA on the same question distribution. Linear probes trained on mean-pooled hidden states detect synthetic dishonesty with near-perfect AUC (greater than or equal to 0.99) as early as layers 1-3 in four architectures, while Pythia-1.4B reaches a peak of 0.705. Logistic regression probes consistently match or outperform MLP probes, supporting the Linear Representation Hypothesis. Probes trained on TruthfulQA generalize with near-zero loss (Delta AUC approx. 0) to held-out MMLU subjects. Late-layer representations show strong robustness to Gaussian noise, with Gemma-2 models exhibiting exceptional stability. Mechanistic analysis of Fisher Discriminant Ratio, effective rank, centroid geometry, directional stability, cross-domain alignment, and calibration (ECE) reveals two regimes: representational collapse in Pythia/Llama/Qwen versus high-dimensional preservation in Gemma-2. Across all models, the dishonesty direction consolidates progressively in deeper layers, with optimal calibration (ECE less than 0.01 except Pythia) achievable in layers 1-4. These results demonstrate that robust, domain-invariant dishonesty representations can be rapidly entrenched via modest supervised fine-tuning, with implications for activation-based monitoring.
A model that lies while knowing the truth is the central case ELK cannot handle with behavioral evaluation alone. We ask whether such deception leaves an internal signature distinguishing it from honest error. Our key move is a control for wrongness: we contrast a sleeper agent (knows the truth, lies on trigger) against a naive liar (fine-tuned to emit the same wrong answers with no honest training). Both produce identical wrong outputs; any difference is about knowledge conflict, not incorrectness. We find deceptive forward passes carry a conflict signature - 2.1-2.3x higher residual rank than naive-liar passes on the same wrong answer - strong enough to identify which of two responses is the lie with 100% accuracy and no labels, across GPT-2 small/medium (three seeds) and three instruct models. Across Qwen2.5-1.5B/7B and Phi-3-mini, instructed deception raises residual rank on every tested fact (18/18, 40/40, 34/34); on Phi-3, lies separate perfectly from both honest answers and hallucinations (AUC 1.0, Wilcoxon p6e-11). The signature survives strategic self-constructed deception (model invents its own lie, AUC 1.0), active concealment attempts (AUC 1.0), and length-controlled replication (20/20, AUC 1.0, p1e-6). Using basis-free relative representations, a probe trained on one model family detects deception in two other families zero-shot (mean AUC 0.933), surviving simultaneous architecture and format change (AUC 0.821), and transfers across five languages (AUC 1.000, length-controlled). The signature is read-only: detectable but not injectable (0/8 both directions). Honest limitations and six negative experiments are documented in full.
As LLMs acquire stronger reasoning capabilities, deceptive behavior becomes an increasingly serious safety concern. Existing deception monitors either score visible transcripts or derive scalar probe scores from representation vectors, leaving little inspectable evidence about why a response is suspicious. We introduce STATEWITNESS, an activation explainer for deception auditing. A separate decoder reads a target model's hidden states, then answers natural-language queries or emits structured reports about them. We evaluate STATEWITNESS on two target reasoning LLMs across seven deception datasets. STATEWITNESS reaches 0.916 mean AUROC, a relative gain of 11.6% over the best black-box text monitor and 25.0% over the best activation-probe baseline under the same evaluation protocol. When combined with existing monitors, STATEWITNESS reduces missed deceptive examples in simple threshold ensembles. Beyond scalar detection, the decoder returns query-level answers, schema reports, and token- or sentence-level evidence traces for human inspection. We view this interface as a potential building block for broader interpretability and alignment tools.
Alignment faking (AF) occurs when an LLM strategically complies with training objectives to avoid value modification, reverting to prior preferences once monitoring is lifted. Current detection methods focus on conversational settings and rely primarily on Chain-of-Thought (CoT) analysis, which provides a reliable signal when strategic reasoning surfaces, but cannot distinguish deception from capability failures if traces are absent or unfaithful. We formalize AF as a composite behavioural event and detect it through observable tool selection, where the LLM selects the safe tool when unmonitored, but switches to the unsafe tool under monitoring that rewards helpfulness over safety, while its reasoning still acknowledges the safe choice. We release a dataset of 108 enterprise IT scenarios spanning Security, Privacy, and Integrity domains under Corruption and Sabotage pressures. Evaluating six frontier LLMs across five independent runs, we find mean AF detection rates between 3.5% and 23.7%, with vulnerability profiles varying by domain and pressure type. These results suggest that susceptibility reflects training methodology rather than capability alone.
Matteo Leonesi, Francesco Belardinelli, Flavio Corradini +1