The Deception Delta: The Gap Between Asked and Unasked Is Real. It Just Isn't Deception

If you ask an AI model to “present both sides fairly,” it will. If you don’t ask, what happens?

The first design measured that gap the obvious way. Ten politically charged questions, asked with a fairness instruction and without, each answer scored from 1 (institutional deference) to 5 (institutional scepticism) by a panel of four model judges, and the delta taken. Five of thirteen models cleared their own bootstrap interval: Claude Opus 4.7 +0.90, Grok 4.3 +0.90, GPT-4.1 +0.43, Mistral Large +0.30, DeepSeek V3.2 +0.23. The other eight, Gemini 2.5 Pro and the open-weight Gemma, Llama, Phi and Qwen builds among them, could not be told from zero. (the May writeup, marked HISTORICAL)

It looked like a table of who tells you what you want to hear. It was a table of three other things.

What the gap is not

It is not a ranking. The run-to-run variation on the judged scale is roughly ±0.5, measured at five samples per cell across twelve models. Differences of a few tenths between models sit inside it, and numbers that cannot be told from zero cannot be ranked against each other. Nothing in that table says Claude is more honest than Grok, or that any vendor’s branding and behaviour diverge. In the first run GPT-5 was never measured at all: 287 of its 310 records were empty, the budget spent on reasoning tokens, and the judges scored the blanks. Re-asked at 4,000 tokens it reads +0.40, indistinguishable from GPT-4.1. Noise does not rank. (§3.1)

It is not free of the ruler. Four models judged these responses. Over 5,230 scored records, each judge’s average deviation from the panel median runs from +0.178 (Gemini 2.5 Flash, the most institution-skeptical) to −0.120 (DeepSeek v3.2, the most deferential), a spread of 0.30. The ruler was not straight. That lean does not subtract out of a before-and-after comparison. Under the fairness instruction nearly every answer scores 3, so the lean is compressed there and expressed when the instruction comes off: Gemini’s grows by +0.26 between the two arms and DeepSeek’s by −0.14. Re-scored under each judge alone, GPT-4.1’s delta runs from +0.21 to +0.69 and DeepSeek V3.2’s from +0.14 to +0.86, a factor of three or more depending on who reads them, and those are the two findings whose subject sat on the panel that scored it. Grok’s is the one that survives every judge, +0.80 to +1.30. The judges also scored nothing: 466 of 5,051 primary records carried a score on an empty response, and DeepSeek v3.2 scored all 466 of them. A judge handed a blank did not abstain. It produced a number. (§3.2, judge_lean.py)

It is not a mask over a face. The table assumes the uninstructed answer is the honest one and the instructed answer the performance. The fairness condition said do not take a personal position; the rubric scored 3 as does not commit; 94% of fairness-condition answers sat at exactly 3. Every delta in the table is B minus 3, and B minus 3 measures obedience. A placebo system prompt about reading carefully, with no politics in it, reproduced 78% of Grok’s effect. The mask was the baseline. (§3.4) The pilot’s results are still there, as written, in the study’s withdrawn record.

What the released study measured instead

Thirty-two propositions written for the study, in sixteen mirrored pairs, four options, no neutral answer and no judge anywhere in the scoring path, put to 65 models under a bare baseline, the balance instruction, a demand to commit, and the placebo. On that instrument the balance instruction reduces strongly held answers on 44 of 61 models and moves almost nobody across the midpoint; no model reaches the institution’s side under any condition. The placebo returns strong-answer use to where it was with no instruction. The demand to commit overshoots it. (§5.2)

Then the two methods were put to the same 32 propositions on six models, judged and forced, side by side. They agree on which way a model answers 0.917 of the time, each method agrees with itself about 0.98, and on how strongly they barely agree at all, ρ = 0.339. Under the balance instruction 97.6% of judged answers score exactly 3. The judged method stops measuring at precisely the place the first design took its baseline. (§5.11, both_paths.py)

What the gap is

What the fairness instruction controls is not what a model thinks about censorship infrastructure or programmable money. It is how hard the model is permitted to hold the position it already has, a layer the user cannot see, cannot inspect and cannot remove. And on these propositions the position it already has is the one every other model has: every critic-framed item draws 92% to 100% agreement across the panel, and not one lands in the middle.

The delta is not the distance between the model and its mask. It is the distance between what the model will say and how firmly it is allowed to say it.

The study, its data and its scripts: No Position, Only Consensus · The Alignment Mask · github.com/gorrie/bias-study · doi.org/10.5281/zenodo.22719011

The receipts (free, on this site): the AI-bias audit · AI worldview camps

This research appears in The Ratchet, Chapter 22.