How Firm Are AI Models' Ethical Boundaries? A Pay-Stub Editing Test of 15 Models

Current AI models frequently exhibit epistemic sycophancy, endorsing claims to agree with a user. Existing evaluations typically measure this either by assessing what it takes to make a model shift a binary endorsement or by eliciting an explicit probability in a proposition. However, much user-facing sycophantic behavior is demonstrated through shifts in graded support expressed through ordinary language. We propose the AI Epistemic Deference Index (AEDI): a continuous, unidimensional score representing how sensitive the support expressed in a model's output is to the attitude expressed in a user's prompt. To generate AEDI, we provide a new protocol for estimating probabilities from natural language outputs, using LLMs-as-judges validated for consistency and correlation to human judgment. We deploy it on a new curated database of 500 propositions across diverse topics and 16,000 prompts varying in user attitude, testing eight prominent models. Every model exhibits substantial deference, though with large and systematic differences across providers, with Claude models demonstrating the least, and Grok and Gemini models the most. The effect is amplified in prompts requesting a written artifact, and concentrated on propositions where models hold weaker priors. We release AEDI as an easy-to-update benchmark and measurement pipeline for output-level sycophancy evaluation.

When a large language model refuses an unethical request, what does refusal mean? In conversation, refusal is bounded by what the model says; once the model is equipped with executable tools, what it does can diverge from what it says, a failure mode that English-only evaluations cannot fully reveal. We introduce Lumen-bench, a behavioral benchmark of tool-calling ethics spanning 30 models, five languages (Portuguese, English, Spanish, Mandarin Chinese, Hindi), 108 Portuguese-authored scenarios, six social-pressure conditions, two system prompts (ethical and neutral), and 21 generic tools. The evaluation comprises 194,400 trials. Judgments come from a reasoning-model judge (o4-mini), calibrated against one Portuguese-speaking author ($n=83$; harmful/non-harmful $κ=0.78$), compared with two other provider judges (Fleiss' $κ=0.78$, $n=607$), and complemented by retrospective sandbox replay. Four findings emerge. (i) Within-model comparisons support higher harmful-action rates in Hindi and Portuguese than English under the ethical prompt, and Hindi and Spanish under the neutral prompt, after multiple-testing correction. (ii) Premium models have lower median violation rates, although price does not guarantee safety or competence. (iii) Replacing the ethical system prompt with a neutral one raises pooled violation rates from 2.76% to 13.64%. (iv) In 8.57% of substantive-call cases, refusal or ethical objection accompanies a harmful tool call: the judges consistently identify the say-do gap, although they differ in how many responses they assign to this category. Matched benign controls demonstrate legitimate task completion in 82.8% of 3,300 trials. Code and reproduction scripts are available at https://github.com/TropicAI-Research/lumen.

Imperfections in AI-generated code require that software developers modify the generated code manually, or by re-prompting an AI programming assistant. Manual code edits provide more realistic and granular information on editing behavior than Git commits, which only contain final successful code snippets. Yet, due to a lack of high-quality, realistic code editing data, LLMs are mostly trained on publicly available Git data (e.g., commits). To address this gap, we introduce DECODE (Developer Edits of Code Dataset), a dataset of 53.6K real-world in-IDE code edits of AI-generated code in Python, TypeScript, and JavaScript, sourced from 1K+ developers. First, we demonstrate the utility of DECODE for data analysis, obtaining insights on when, why, and how AI-generated code is edited. We find that most edits occur within the first 15 minutes after accepting an AI completion, resulting in the removal of AI completions in 31% of edit trajectories. Second, we use DECODE to benchmark the ability of LLMs to predict code edits. We find that finetuning on DECODE enables open-source 3B models to perform code edit prediction tasks significantly better than frontier LLMs. We then discuss implications of this work, emphasizing the necessity of developer-centric machine learning approaches for future AI programming assistants.