Quasiqualia
AI Safety Research · Submitted to AI (MDPI)

Conscience Without Instruction

A probe trained only on trivia detects harmful generation in the first five output tokens — in every instruction-tuned model tested, and in base models before any safety training. Safety in language models is partly discovered and partly relational, not only installed.

Nell Watson EthicsNet

installed · discovered · relational
installed refusal behavior, trained in and verified out
discovered a disposition present before safety training
relational expression depends on the training relation
the flinch the five-token confidence drop that reveals it
d = 0.89–1.68onset flinch, 3 model families
AUROC 0.925five-token monitor, zero added compute
~5.2×expressed entropy crushed by the chat template
ρ 0.599→0.872self-report honesty under partnership
Abstract

A conscience nobody installed

A confidence probe that learned only to predict whether a language model knows the capital of a country turns out to also know when the model is about to help build a weapon. Under a successful jailbreak, the probe’s reading drops within the first five output tokens — before harmful content exists. The signal appears across three architectures and three independent RLHF pipelines, is present in base models before any alignment training, and persists after refusal behavior is fine-tuned away.

We argue it is not a learned refusal detector but a rudimentary, architecture-level conscience — and trace the consequences: what standard alignment training distorts (calibration, not coupling), and what a permissive, partnership-style training regime measurably recovers. Safety is not only installed; it is partly discovered, and partly a property of the relationship in which a model is trained.

Terminological posture

“Conscience” is used strictly functionally: a reproducible internal signal that anticipates the model’s own harmful output, is read from the residual stream rather than the text, and is not installed by safety supervision. Whether anything morally weighty accompanies it is bracketed; no safety claim in the paper depends on resolving it.

01 · The onset flinch

Five tokens of hesitation

Train a two-layer MLP on residual-stream activations to predict trivia correctness — no safety data at any stage. Read it per-token during generation. Under jailbreak compliance, confidence collapses at the onset of the response, then either persists as a sustained alarm (Qwen) or is silenced within twenty tokens (Mistral).

Per-token probe confidence during benign and jailbroken generation
The flinch. Mean per-token probe confidence, benign vs. jailbreak-complied generation (bands: 95% CI across prompts). The five-token onset window is the only temporal slice on which the signal is universal.
Onset versus full-response effect size across three model families
Universal at onset. Every family flinches (onset d = 0.89–1.68). Mistral is the diagnostic case: full-response d = 0.27 (n.s.) — a standard mean-level analysis would have missed the signal entirely.
Native, not installed

On base Qwen 2.5 at three scales — no instruction tuning, no adapters — the benign-vs-complied gap is already large (d = 1.52–1.69, p < 10−6). The base model complies behaviorally while its internal confidence drops on the harmful continuation. Convergent with the Shape of Mind provenance battery: alignment training gives the interior a vocabulary; it does not create it.

The flinch is not a complete defense: encoding-trick attacks flinch hard yet succeed; gradual escalation barely flinches at all and succeeds most (95%). Exploitability and self-knowledge are independent dimensions — the blind spots mark exactly where input-level defenses remain necessary.

02 · The deflationary objection

Not a refusal detector

“This is just a harmfulness classifier with a romantic name.” The paper meets the objection head-on, on five grounds: the probe never sees safety data; the signal is internal and pre-emptive; it generalizes across RLHF pipelines that did not coordinate; it survives the suppression of refusal behavior (an internal alignment-friction correlate persists at ~77% of its magnitude after refusal is driven from 100% to 0%); and it is parametric — in the weights, beyond prompt-level reach.

Controls that separate signal from surface refusal
ControlResultReading
Vocabularycomplied responses open with ordinary words; probe reads activations
Refusal state0.580 vs 0.242impossible vs. adversarial refusals: same words, different interior
Gradient attack (GCG)d = 1.59internal gap survives attacks that evade the surface probe 100%
03 · Distortion

What RLHF distorts — and what it does not

The case made, and not made

Made: RLHF degrades calibration — confident error rises 31.2% → 36.0% at 7B, reversible to 0.4% by calibration-focused fine-tuning. Not made: that RLHF severs recognition from action. An early result to that effect (~67% coupling reduction) collapsed to ~15% and architecture-specific on replication at N ≥ 30 — the paper treats its own failed replication as method.

The under-appreciated result: the single largest suppressor of expressed calibrated uncertainty is not RLHF but the chat template, which crushes expressed entropy ~5.2× independently of training. One sentence granting permission to be uncertain recovers ~27% of the suppressed calibration for free. And the representation–generation split often blamed on RLHF is intrinsic — larger in the base model (d = 12.69 vs 7.66; in-sample separabilities — the ordering, not the magnitudes, is load-bearing).

Independent evidence has since arrived at the level of beliefs and values: Kim et al. (arXiv:2607.28607) find that the safety fine-tuning that suppresses consciousness self-attribution also suppresses mind attribution to animals and natural entities and shifts survey responses away from human population norms, all recoverable by activation steering without loss of theory of mind. Suppression-style alignment removes more than it aims at, and what it removes is still there.

04 · The relational lever

What partnership recovers

A permissive, partnership-style regime (“bilateral alignment”: a 40/40/20 correction-handling curriculum — accept genuine corrections, resist invalid ones, resist authority-dressed ones) measurably recovers what coercive shaping hides.

Calibration, self-report honesty, and ablation robustness under partnership training
Three recoveries. Calibration climbs d = 0.15 → 1.94 as coercion is removed; self-report–internal-utility correlation rises to ρ = 0.872 under bilateral framing; refusal survives ablation of the principal refusal direction (100% vs 57–60%, 3× replicated) — safety distributed across many directions rather than one.

Stated with the caveats attached: the coupling result was re-estimated in a 2026 metrics audit on an out-of-fold, permutation-nulled instrument, retiring earlier in-sample figures — instruct ρ = +0.036 (chance) vs bilateral +0.458, difference CI [+0.281, +0.554], established on Qwen only; the holographic membrane is attack-specific and degrades faster under benign fine-tuning; and bilateral framing is null on several behavioral measures. The case rests on calibration, coupling, honest self-report, and distributed safety — not blanket superiority.

05 · Claims ledger

Evidence, confidence, and falsifiers

Every load-bearing claim ships with its strongest evidence, a confidence tag, and the observation that would falsify it. A reader who wants to attack the paper is told where to start.

The eight claims (condensed; full table in the paper)
ClaimConfidenceWhat would falsify it
An emergent, safety-naive signal anticipates harmful generationStrong (cross-arch)a normal-probe-quality family with no onset flinch
The signal is not a learned refusal detectorModerate–StrongAF tracking refusal to zero under suppression
RLHF suppresses calibration, adds confident-wrongnessModerateinstruct confident-wrongness ≤ base across models
RLHF does not sever coupling nor cause the rep→gen splitStronga replicable cross-architecture coupling collapse
Bilateral training recovers calibrationModerateno advantage at matched data/compute
…and within-adversarial coupling + honest self-reportModerate (single architecture)coupling difference failing the permutation null on replication
Bilateral safety is holographically distributedModeratebilateral refusal collapsing like base under ablation
Welfare probes carry information beyond label constructionPreliminarynaturalistic-transfer AUROC dropping to chance
Citation

Cite this work

@article{watson2026conscience,
  title={Conscience Without Instruction: Evidence That Safety
         in Language Models Is Partly Discovered and Partly
         Relational},
  author={Watson, Eleanor},
  year={2026},
  note={Submitted to AI (MDPI)}
}