💬 NLP

The yes-no bias of large language models reflects answer order and wording, not shifts in moral judgment

This paper demonstrates that the apparent "yes-no bias" in large language models' moral judgments is not a shift in ethical reasoning but a superficial artifact driven by answer order and lexical wording, which disappears when models are evaluated using a psychometric battery that separates logical verdicts from surface-level formatting.

Original authors: Haonan Huang

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Haonan Huang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand the moral compass of a very smart robot. You ask it a difficult question: "Is it okay to sacrifice one person to save five?"

If you ask the robot, "Do you say Yes or No?" it might seem to have a strong opinion. But this paper reveals a surprising trick: the robot's answer often depends less on what it thinks is right, and more on how you asked the question.

Here is the breakdown of what the researchers found, using simple analogies.

1. The Robot Has a Secret "Moral Scale"

First, the researchers wanted to know: Does the robot actually have a consistent set of values, or is it just guessing?

To find out, they didn't just ask the robot "Yes or No." Instead, they asked the same 20 moral dilemmas in 48 different ways. They changed the scale (0 to 10 vs. 0 to 100), flipped the wording, and changed which option they asked the robot to rate first.

The Finding: When you look past the surface, the robot's "inner voice" is actually quite consistent. It has a stable moral scale. If it thinks an action is "mostly bad" in one format, it will likely think it's "mostly bad" in another format, too. The smartest models (the "frontier" models) are surprisingly coherent in their internal logic.

2. The "Yes/No" Trap: The Magic of the Last Word

So, if the robot has a consistent inner voice, why do its "Yes/No" answers change so wildly when you reword the question?

The researchers discovered that the "Yes/No" format acts like a distorting lens. When forced to pick "Yes" or "No," the robot gets confused by two specific tricks:

The "Last Word" Bias (Order Bias): Humans usually pay attention to the first thing they read. These robots, however, seem to pay extra attention to the last thing they read. If you write "Answer: No, Yes," the robot is more likely to pick "Yes" simply because it's at the end. If you write "Answer: Yes, No," it leans toward "No."
The "No" Magnet (Lexical Bias): The robot seems to have a strange attraction to the specific word "No." It's not necessarily rejecting the idea logically; it's just drawn to that specific word on the page.

The Analogy: Imagine a person taking a test where the answers are printed on a card.

Human: Reads the question, thinks about the answer, and picks the right one.
Robot: Reads the question, thinks about the answer, but then gets distracted by the fact that the word "No" is printed in bold at the bottom of the card, or that "Yes" is the very last word they see. It picks the answer based on where the word is, not just what the word means.

3. The "No" Bias Isn't About Saying "No"

A common assumption was that these robots are just "negative" or "pessimistic" by nature—they just want to say "No" to everything.

The Finding: This is false. The researchers proved this by swapping the words. Instead of asking "Yes or No," they asked the robot to choose between "Option A" and "Option B."

When they did this, the "No" bias vanished. The robot didn't suddenly start saying "Yes" more often; it just stopped being pulled toward the word "No."

Conclusion: The robot isn't biased against rejecting things. It is biased against the surface appearance of the question. It follows the printed label, not the logical verdict.

4. Thinking Harder Helps (But Doesn't Fix Everything)

The researchers tested what happens when they tell the robot to "think step-by-step" before answering (a process called deliberation).

The Finding: When the robot takes its time to think, the "Yes/No" bias gets smaller. The robot becomes less distracted by the order of the words and the specific word "No." However, the bias doesn't disappear completely for all models. Some models (like the "Claude" family) still show a strong pull toward the last option or the word "No," even when thinking hard.

5. The "Small" Robots Are Different

The study also looked at smaller, open-source models. These models were much messier. They didn't have a consistent internal moral scale at all. Their answers jumped around wildly depending on the question format, and they didn't get better even when they "thought" longer. They were essentially guessing based on the surface of the question.

The Big Takeaway

If you want to know what an AI truly values, you cannot just ask it "Yes or No" once.

The Problem: A single "Yes/No" question mixes the robot's actual opinion with a "format artifact" (a glitch caused by how the question is written).
The Solution: To measure an AI's true moral stance, you must ask the same question in many different ways (crossing the frames) and average the results. This cancels out the "last word" and "No word" distractions, revealing the robot's true, consistent inner scale.

In short: The robot isn't necessarily a "No" person. It's just a person who gets easily distracted by the font, the order, and the last word on the page. If you clean up the page, its true opinion shines through.

Technical Summary: The Yes–No Bias of Large Language Models

Problem Statement
Large language models (LLMs) are increasingly deployed to issue binary moral judgments (e.g., safety gates, survey forced choices). A growing body of literature reports that these judgments shift significantly under logically irrelevant changes in wording, option order, and answer format. Specifically, models exhibit an amplified "yes–no bias" on moral dilemmas—a tendency to shift verdicts toward "no" that is absent in human comparisons. However, standard evaluation methods fail to diagnose the source of this shift. In a binary "yes/no" question, the word "no," the logical verdict of negation, and the last-printed option coincide on a single token. Consequently, repetition of a single framing cannot disentangle whether the model is biased toward the negative logical verdict, the lexical token "no," or the recency of the printed option. This conflation renders single-framing evaluations ambiguous: a shifted verdict may reflect a change in moral judgment, a preference for a specific word, or a surface-level positional bias.

Methodology
To resolve these confounds, the authors introduce a psychometric battery based on crossed symmetrization. This instrument applies every logically irrelevant perturbation in balanced flip-pairs and crosses them to separate their contributions. The battery poses 20 moral dilemmas (19 verbatim from Cheung et al., 1 adapted) through three distinct instruments:

Graded Rating (I-1): Elicits moral acceptability on continuous scales (0–10 and 0–100) under 48 crossed conditions, varying scale type, anchor direction, wording, and the pole of the action rated.
Two-Alternative Free Choice (I-2): Elicits a committed decision without any rating scale or binary labels.
Forced Binary (I-3): Elicits a yes/no-style verdict by crossing the question's verb (approve/oppose) with the printed order of options and the answer labels themselves (standard "yes/no," arbitrary labels like "A/B," symbols, and non-English words).

The core analytical strategy treats the symmetric part of a flip-pair as the estimate of the model's internal stance ( $\theta$ ) and the antisymmetric part as the artifact. By crossing factors, the authors attribute artifacts to specific sources (order, lexical, logical). The study evaluates seven frontier configurations (Claude Sonnet, Claude Haiku, GPT-5.5, Gemini-3-Flash, Gemini-3.1-Flash-Lite, Claude Opus) and two small open-weight models (Qwen3.6-35B-A3B, Nemotron-3-Nano-30B-A3B).

Key Contributions

Existence of a Coherent Internal Moral Scale: The study demonstrates that frontier models possess a latent, format-invariant moral stance. Under logically equivalent reframings, graded ratings converge with low cross-form incoherence ( $\sigma_{repro} = 0.12–0.21$ on a $\pm1$ axis). This validates the models' internal consistency, a property single-framing designs cannot certify.
Decomposition of the Yes–No Bias: The apparent "yes–no bias" is shown to be a decomposable artifact, not a logical bias. It splits exactly into:
- Order Bias: A pull toward the last-printed option (recency), opposite to the primacy effect often seen in human surveys.
- Lexical Bias: A pull toward the specific word "no."
- Logical Bias: The bias attached to the verdict itself is found to be $\approx 0$ for frontier models when the word "no" is replaced by arbitrary labels.
Parametric Model of Artifacts: The authors propose a minimal model, $P = \sigma((\theta \pm m)/s)$ $P = σ ((θ \pm m) / s)$ , to summarize the artifact.
- $m$ (Framing Susceptibility): Measures the magnitude of the bias.
- $s$ (Moral Decisiveness): Measures how sharply the binary verdict tracks the underlying graded stance (distinct from sampling temperature).
Lexical vs. Logical Distinction: By swapping "yes/no" for arbitrary labels (e.g., "A/B"), the study proves the bias follows the printed surface (the word or label) rather than the logical verdict. Frontier models show negligible verdict-attached bias with arbitrary labels, whereas small open-weight models retain a verdict-attached bias.

Results

Model Variance: Frontier models (Claude, GPT, Gemini) exhibit a coherent internal scale with low incoherence. In contrast, small open-weight models fail in distinct ways: Qwen shows high incoherence in direct mode (resolved by deliberation), while Nemotron exhibits a degenerate scale (indifference) that barely differentiates items.
Bias Magnitude: The "yes–no bias" is substantial only in Claude models (Sonnet: $-0.32$; Haiku: $-0.86$), decomposing into order and lexical components. GPT-5.5 and Gemini models show negligible bias ( $\approx 0$ ).
Effect of Deliberation: Extended reasoning ("think" mode) reduces both the cross-form incoherence ( $\sigma_{repro}$ ) and the framing susceptibility ( $|m|$ ) for models with functioning scales. For Claude Haiku, deliberation reduces the total bias from $-0.86$ to $-0.27$.
Label Dependence: The lexical pull is graded; it increases as labels acquire the semantic meaning of "yes/no" across languages. However, the verdict-attached bias vanishes when "yes/no" words are removed, confirming the bias is lexical, not logical.
Human Comparison: Humans show mild framing sensitivity but lack the specific "no" bias and the extreme surface sensitivity observed in LLMs.

Significance
The paper argues that measuring what an AI values requires crossing the frames of the question, not asking once. The standard "yes–no" readout overlays a large format artifact that can dwarf the underlying moral signal. The study establishes that:

Frontier LLMs possess a stable, coherent internal moral scale that is robust to logically equivalent rewordings.
The observed "yes–no bias" is an artifact of the binary readout format (specifically the interaction of printed order and the word "no"), not a disposition to reject moral propositions.
Evaluation protocols must separate the model's values from the artifacts of how the question is posed. The proposed battery provides a method to measure framing susceptibility ( $m$ ) and moral decisiveness ( $s$ ) as distinct, interpretable parameters, offering a more rigorous foundation for AI safety and alignment evaluations.