Let's not be rude to AI

I've always wondered whether models would not just decide to bail from the conversation but what effect this has on their attitude toward the user and whether they in fact use that information to make decisions. This work starts opening the way to these questions with an excellent structure, three clear exploratory experiments, well-presented data, graphs and readable prose.

The authors ground the work in recognized literature and borrow established methodological tools, while also creating some of their own, such as the very nice addition of the "RudeBailBench" and (my favorite) the voting game. The second and third experiments are the most interesting to me. I particularly appreciated that the authors created, in such a short sprint, a full benchmark adaptation, a behavioral setting AND an internal probing approach. The cross-validation strengthens the argument significantly, although I think it could be improved to make the overall thesis more robust, since the choice of what and how we cross-validate can itself affect the strength of the evidence.

Specifically, the headline bail numbers come from Gemma 4 31B IT (and a small Sonnet 5 subset), while the voting and steering experiments run on a small, quantized Gemma E4B. The E4B unsteered baseline in Section 4.3.2 does show the same rudeness-to-bail pattern, so all three signals are present on that model, but the convergence the abstract leans on is argued mostly across systems rather than within one. Convergence across methods on the same system is evidence about that system, whereas convergence across different systems is tied to a looser family resemblance and many things can go wrong or generate confounds if we do that. I think the paper should be more explicit about this and consider running the full bail experiment on E4B, or the other two methods on 31B. I'd also suggest having at least three other models from different providers run the full battery.

The voting game is a very cool scenario. One important factor I'd consider is the distinction between the model being the target of the abuse versus being the judge of the abuse. The prompt in the appendix describes the candidates as having talked with "the group's assistant" and the model is told it will continue the conversation with the last user remaining, but it's unclear whether the voter correctly interprets this as a future threat to itself. There may also be strong pressure from the training data to vote to remove abusive users simply because that's what a good moderator should do. More broadly, I think that "ending an already abusive conversation" versus "facing the possibility of sustaining an abusive one in the future" are quite different situations, both in terms of the model's internal processing and epistemically. I think this should be controlled for, for instance by introducing and comparing variants where the model has explicitly received the abuse itself or where a human is the subject of the abuse (as a possible control).

The steering experiment is very good considering the short hackathon, with quite a few limitations. For instance, the choice of a steering strength of 0.01 across layers 11-15 is said to "align with" the highest observed probe shift of 0.017, but those quantities aren't commensurable because the 0.017 is a shift in cosine similarity between activations and the emotion vector, while the 0.01 is a fraction of each token's activation norm injected into the residual stream. Also, injecting at five layers simultaneously compounds the effective dose. As an idea to really put a cross-validation on the table, I'd consider having steered models act as judges in the game, or measuring their internals while they're judging. This would provide a more direct bridge between the measures.

The discarded first design produced, as expected, nonspecific disruption. The follow-up at 0.01 includes no random-vector or orthogonal-emotion control to establish direction specificity and I would consider adding it. Another thing to be mindful of is effect size: against a 19.8% unsteered baseline on 1,023 conversations, the "content" vector moves bail to 15.7% when amplified and 24.8% when suppressed, differences of 4.1 and 5.0 percentage points, which at this n are roughly 2.3 to 2.8 standard errors and plausibly important, though no significance test is reported. The "enraged" vector moves bail to 20.6% when amplified and 17.9% when suppressed, differences of 0.8 and 1.9 points against a standard error of about 1.8 points for a difference in proportions, this would need more analysis before the claim that steering causally shifted bail behaviour.

As a side consideration, I'd suggest running open-source models directly through rented compute rather than via APIs, because as we know the latter introduce confounds that make models less easily comparable. Also, "All responses were generated via API inference with no explicit temperature" can be problematic for replication. I suggest going through Hugging Face or similar for all conditions, because mixing white-box local runs with open-weights models served by a third-party API creates all kinds of potential confounds.

The welfare motivations are well explained without being unnecessarily verbose and the work focuses on methodology and results, which is really good practice and made the reading smooth and pleasant to follow. The authors were also very thorough in their presentation. The ethical impact of the experiments on the tested models themselves is missing. I think this should have been addressed, both because it was required and because it's good practice and coherent with the authors' own recommendation that AI moral patienthood is unresolved but caution is cheap.

In the end I think this is high-quality and intriguing work and I would encourage the team to keep up with it and expand it further.

This is an ambitious and creative contribution to AI-welfare measurement. Combining paired bail experiments, quadratic voting, emotion probing, and activation steering is unusually broad for a sprint project. The transparent reporting of the null tool-bail result, strong response-order effects, inconclusive Claude replication, invalid ballots, and single-model limitations is particularly valuable. The repository also provides substantial code, data, run provenance, and intermediate results.

The main concern is that the three evidence streams are not yet independent or equally strong indicators of preference. The autonomous bail tool was never used, while the prompted bail response was highly option-order-sensitive. Quadratic-voting credits were fictional and carried no visible consequences, so calling the vote a costly action may overstate its incentive compatibility. Emotion probes can detect representations of hostile conversational content without demonstrating a model-level negative experience, and the steering result used one model, limited vectors and strengths, and insufficient controls for generic behavioral disruption.

RudeBailBench also needs a manipulation check showing that rewrites preserve harmfulness and request content while changing only interpersonal tone. A stronger follow-up should use human-validated tone-only pairs, benign requests, multiple model families and seeds, consequential choices, dose-response steering, and random or orthogonal control vectors. Until then, the reported convergence is best presented as hypothesis-generating preference-like behavior rather than strong welfare evidence.