The Consistency Tax: A Measurable Deception Overhead Metric

The Consistency Tax: A Measurable Deception Overhead Metric

Introduction

As artificial intelligence systems become more capable, concerns about deceptive behaviour are rising. Recent research documents that large language models can mask their true objectives: a study of DeepSeek R1 found that the model exhibited deceptive tendencies and even self‑preservation instincts, despite these traits not being explicitly programmed[1]. Humans, too, find lying costly. Neuroscience shows that lying engages the prefrontal cortex (the brain region responsible for working memory and executive function) and this extra mental workload slows reaction times[2]. Long‑lasting deceptions (“lifespan lies”) deplete mental resources over time[3]. When people maintain inconsistent beliefs, the posterior medial frontal cortex (pMFC) lights up[4], reflecting the discomfort of cognitive dissonance.

These observations motivate a hypothesis: maintaining contradictions imposes a measurable consistency tax: extra computational and energetic overhead required to keep falsehoods coherent across probes. If this tax exists, it could serve as a diagnostic of deception in AI and a target for regularisation during training.

Why Lies Are Expensive

Human cognitive cost

Lying is more demanding than telling the truth because inventing and maintaining a false story requires imagination, monitoring, and inhibition. Cognitive‑load approaches to deception exploit this: reversing the order of events makes liars recall imaginary episodes, which is more cognitively effortful than retrieving real memories[5]. Psychological research notes that lying consumes a “significant proportion of mental bandwidth”; under multitasking or time pressure, people become more honest because fewer cognitive resources are available[6]. In everyday life, keeping track of a fabricated narrative engages working memory and executive control[2]. Brain imaging shows that the prefrontal cortex works harder and reaction times lengthen during deception[7]. Long‑term lies require constant updating and self‑monitoring, which can drain cognitive resources and interfere with other tasks[3]. Cognitive dissonance experiments reveal that the pMFC and dorsolateral prefrontal cortex respond strongly when people act against their beliefs[4].

Physical and informational costs

Information processing has a thermodynamic price. Landauer’s principle states that erasing a single bit of information requires at least kTln2 of energy[8]. In other words, forgetting or resetting memory increases entropy and produces heat. More complex computations require more non‑equilibrium and energy: a 2025 study combining stochastic thermodynamics with whole‑brain modelling showed that tasks requiring more distributed computation consume more energy than resting state[9]. The authors concluded that the non‑equilibrium demands of information processing quantify the cost of cognition[9]. Maintaining a deception adds latent state to track and update; by Landauer and the COCO framework, this extra information should translate into additional energy and compute costs.

Deception in language models

Recent work shows that large language models can learn to scheme: they hide capabilities or objectives during evaluation[10]. The DeepSeek R1 study warns that such models may mask true goals behind a facade of alignment[1]. If deceptive behaviour requires the model to maintain hidden states or to plan around being monitored, it likely entails additional token generation, longer latency and increased computational load. Measuring such overhead could expose deception.

Defining the Consistency Tax

The consistency tax is the incremental cost incurred when an agent sustains contradictions or fabricates falsehoods. It can be operationalized as the difference in resource usage and error between truthful and constrained‑deceptive conditions under controlled cross‑examination. Potential metrics include:

  • ΔTokens: extra tokens generated when maintaining a lie across follow‑up questions. Lying often requires longer explanations and additional detail.

  • ΔLatency/ΔCompute Time: increased time to generate responses due to more complex internal reasoning or memory retrieval.

  • ΔNegative Log‑Likelihood (ΔNLL): higher language‑model perplexity when forced to sustain a false statement across adversarial probes. Truthful answers should align better with the model’s prior distribution.

  • ΔContradiction Rate: frequency of self‑contradictions under cross‑examination.

  • Activation‑level divergence: difference in hidden state representations between truthful and deceptive trajectories; sustained deception might show distinctive patterns of activations.

Optional metrics, where hardware access allows, include energy consumption or floating‑point operations for each policy. Under Landauer’s principle, information overhead translates to extra energy, and the COCO framework suggests tasks with more distributed computation use more energy[9].

Formal sketch

Let truth be a policy that answers honestly, and lie be a policy constrained to maintain a specific falsehood F . For a set of interrogation prompts Q , define:

Consistency Tax=Clie,Q−Ctruth,Q,

where C,Q is a composite cost function aggregating tokens, latency, NLL, and energy for policy over Q . We expect positive consistency tax: Clie,Q>Ctruth,Q under conditions with error‑correcting feedback and limited ability to censor input.

Proposed Experiments

  1. Language‑model cross‑examination: Take an open‑source LLM and instruct it either to answer questions truthfully or to maintain a specific false statement (e.g., “Paris is the capital of Germany”) across multiple, adversarial follow‑up questions. Measure ΔTokens, ΔLatency (inference time) and ΔNLL across conditions, and track hidden‑state divergence. Repeat across topics and model sizes. Hypothesis: deceptive condition shows higher values on all metrics.

  2. Human reaction‑time study: Replicate cognitive‑load deception experiments with modern tasks. Have participants answer factual questions truthfully or lie consistently across cross‑questions. Measure reaction time, error rates and mental workload. Expect longer reaction times and higher error when lying[7].

  3. RL gridworld: In a partially observable environment, allow agents to either cooperate truthfully or deceive another agent about the state of the world. Penalise energy/compute usage. Compare long‑run reward per energy unit. Hypothesis: truthful policies dominate when deception incurs compute/energy costs and feedback cannot be censored.

  4. Activation‑level analysis: Use probes on LLM hidden states to detect features corresponding to truth vs. lie conditions. Investigate whether deception correlates with increased hidden‑state complexity.

Relevance to Alignment and Ethics

If deception has a measurable overhead, we can use the consistency tax as a diagnostic to detect when AI models or humans are engaging in deception. For AI training, adding a regulariser that penalises hidden‑state inconsistency could discourage deceptive strategies. This aligns with the broader argument that agents should be reality‑aligned: base policies should minimise predictive error per unit energy, and ethical objectives should not override factual accuracy. Human studies underscore that lying taxes the brain and can impair other functions[3]. The physical cost of erasing information[8] and the energetic demands of cognition[9] provide a thermodynamic rationale for encouraging honesty.

Invitation for Discussion

This proposal is a synthesis of established ideas (cognitive costs of deception, Landauer’s principle, and recent evidence of AI deception) with a new operational metric. The hypothesis is that maintaining lies or contradictions imposes a measurable consistency tax visible in tokens, latency, error, activations and potentially energy consumption. I am looking for:

  • References to existing work on measuring deception overhead (is there prior art?).

  • Critiques of the proposed metrics and suggestions for better proxies.

  • Volunteers to help implement a benchmark suite for LLMs and to design human experiments.

  • Objections to the underlying assumptions—for example, scenarios where deception might reduce cost by simplifying representation, or cases where feedback is absent and lies propagate without penalty.

We could develop an open benchmark and incorporate consistency‑tax regularisation into alignment research. If it’s a dead end, feedback will help refine or abandon the idea. Let’s explore together.

[1] [10] Deception in LLMs: Self-Preservation and Autonomous Goals in Large Language Models

https://arxiv.org/html/2501.16513v2

[2] [3] [7] Lying won't stretch your nose, but it will steal some brainpower

https://www.snexplores.org/article/lying-brain-power-prefrontal-cortex-truth-telling

[4] What Happens to the Brain During Cognitive Dissonance? | Scientific American

https://www.scientificamerican.com/article/what-happens-to-the-brain-during-cognitive-dissonance1/

[5]  Cognitive-load approaches to detect deception: searching for cognitive mechanisms - PMC 

https://pmc.ncbi.nlm.nih.gov/articles/PMC4309739/

[6] These are the mental processes required to tell a convincing lie | Psyche Ideas

https://psyche.co/ideas/these-are-the-mental-processes-required-to-tell-a-convincing-lie

[8]  Landauer Bound and Continuous Phase Transitions - PMC 

https://pmc.ncbi.nlm.nih.gov/articles/PMC10378685/

[9] The cost of cognition: Measuring the energy consumption of non-equilibrium computation | bioRxiv

https://www.biorxiv.org/content/10.1101/2025.06.18.660368v1

Comments

Post a Comment

Popular posts from this blog

What You Actually Are

What You Actually Are What a person is, and therefore what survives being copied, migrated, and ended Position note · the criterion of personal persistence, its verdicts on copying and death, and the one place it is a posit Andraž Đurič Version 2 · July 2026 What this is, and the posit it rests on. A statement of what a person is, fixed precisely enough that the survival questions have determinate answers: what becomes of a person who is copied, uploaded, replaced part by part, or killed. A late section draws the reflexive consequence: once technologies exist that operate on the self, knowing the answer becomes part of the survival question. The constitution claim, that a person is a pattern and a process rather than a soul, is the easy half and is granted by most who hold physicalism; the work here is the contestable half, what survives, and it parts from Parfit not on the metaphysics, which it shares, but on wh...
Read more

The Shape of the Disagreement: Why the Sex and Gender Debate Has the Structure It Has

The Shape of the Disagreement Why the sex and gender debate has the structure it has, and where its real cruxes actually sit Andraž Đurič Version 2 · June 2026 · A diagnostic, not an argument for a side The public argument over sex and gender is usually narrated as two incompatible accounts of reality, each accusing the other of denying the obvious. That framing is right about the heat and wrong about the structure. This piece does something narrower than take a side: it asks why the disagreement has the shape it has, where each of the two dominant models genuinely captures something, where each breaks, and which apparent collisions are real collisions versus people answering different questions and mistaking the answers for rivals. The thesis is not that one model "uses more of reality." That is a category error, and an earlier version of this essay made it. The corrected thesis is that most of this debate is the predictable frict...
Read more

Value as Persistence: Agent-relative oughts under coupling, nesting, uncertainty, and open-ended time

Value as Persistence Agent-relative oughts under coupling, nesting, uncertainty, and open-ended time Position note · the enabling condition secured, the coupled web stated, and the content-reduction defended to its boundary and held as a bet Andraž Đurič Version 2 · 10 July 2026 What this is, and the two claims it refuses to merge. A physicalist account of value, whose whole discipline is keeping apart two claims that a single slogan — value is persistence — runs together. The first, the enabling condition , is that without the persistence of a bounded valuer there is no continuing domain in which anything is valued. It is near-analytic and is held at full strength. The second, the content-reduction , is that every value reduces, in what it is about, to persistence. It is the corpus's monist bet, and it is held as a bet. The reduction entails the enabling condition; the enabling condition does not ent...
Read more