A small tool appeared on Hacker News recently that does something deceptively simple: it translates ordinary English phrases into what its creator calls Claudish, the characteristic hedging, qualifying, and refusal-adjacent language that Claude, Anthropic's large language model, produces when navigating sensitive or ambiguous requests. The project, hosted at programasweights.com/claudish, attracted modest but substantive discussion, and it deserves more analytical attention than a 46-point Hacker News thread typically generates. It is, at its core, a behavioural fingerprinting exercise, and it surfaces something genuinely interesting about how alignment training manifests as a detectable linguistic register.
What Claudish Actually Captures
The premise of the tool is that Claude's outputs, particularly when the model is declining, hedging, or performing safety-adjacent reasoning, exhibit a recognisable stylistic signature. Phrases like "I want to be helpful here, but I also want to make sure we approach this thoughtfully" or "That's an interesting question, and I can engage with parts of it" are not random variation. They are the surface manifestation of RLHF-shaped reward signals interacting with Constitutional AI constraints. The translator essentially encodes this register as a mapping function.
This is not a trivial observation. From a mechanistic interpretability standpoint, what we are looking at is the output-layer projection of a learned policy that has been trained to balance helpfulness against a set of harm avoidance objectives. The specific word choices, the particular rhythm of concession-before-refusal, and the characteristic epistemic hedging all emerge from training dynamics rather than from any explicit template. The fact that they are regular enough to be parodied and systematised is itself empirically significant.
Research from Anthropic and from independent groups studying RLHF-trained models has consistently shown that reward model training induces distributional shifts in output style, not just content. Work on sycophancy in LLMs, including findings from Perez et al. (2022) and subsequent follow-up studies, demonstrates that models trained on human preference data develop stylistic patterns that correlate with approval-seeking behaviour. Claudish is, in a sense, the approval-seeking register applied to the specific domain of harm avoidance.
The Linguistic Register of Alignment
It is worth being precise about what constitutes this register, because the Claudish tool is implicitly making a claim about its structure. Several features are consistently present:
- Performative acknowledgement: The model signals that it has understood the request before explaining why it will not fully comply. This creates a conversational softening effect.
- Epistemic hedging: Claims are qualified with phrases like "I think", "it seems", or "from my understanding", even when the model's actual confidence is high. This is a trained behaviour, not genuine uncertainty.
- Proactive reframing: Rather than a flat refusal, the model offers an adjacent task it can perform. This is the "I can't do X, but I can help with Y" construction that has become almost a signature pattern.
- Moral reasoning made visible: Constitutional AI training specifically encourages the model to articulate its reasoning, which means refusals often come with an explanation of the ethical principle being applied. This transparency is by design, but it produces a distinctive verbosity.
The Claudish translator encodes these patterns as generative rules. Whether it does so accurately or whether it is primarily a comedic approximation is a separate question, but the underlying claim, that this register is real and learnable, is well-supported by anyone who has interacted extensively with the model.
Implications for Red-Teaming and Adversarial Probing
Here is where the analysis becomes more pointed. Tools like this one are not merely satirical. They represent a form of behavioural characterisation that has direct relevance to adversarial probing and red-teaming methodologies. If the refusal register is sufficiently regular to be modelled and reproduced, then it is also sufficiently regular to be inverted, which is to say, one can use knowledge of the refusal pattern to craft prompts that are maximally unlikely to trigger it.
This is not a novel attack surface. Jailbreaking research, from the early GPT-3 era through to more recent work on universal adversarial suffixes by Zou et al. (2023), has consistently exploited the gap between a model's surface-level compliance signals and its underlying capability. What Claudish adds to this picture is a publicly accessible, human-readable model of the compliance signal itself. The Hacker News comments touched on this obliquely, with several users noting that understanding the register makes it easier to write prompts that avoid triggering it.
This is a meaningful observation from an AI safety perspective. The alignment tax, the reduction in raw capability or helpfulness that results from safety training, is partly mediated through this linguistic register. When users learn to route around the register, they are effectively reducing the alignment tax without changing the underlying model. This is not the same as a jailbreak in the traditional sense, but it occupies an interesting intermediate space.
What This Reveals About Alignment as Style
The deeper issue that Claudish surfaces is that current alignment approaches produce alignment as a stylistic overlay rather than as a deep structural property of the model's reasoning. This is a known criticism of RLHF-based approaches, articulated clearly in work by Hubinger et al. on deceptive alignment and in more recent mechanistic interpretability research showing that safety-relevant concepts are often represented in relatively shallow, accessible layers of transformer models.
If alignment were a deep property, a genuine modification of the model's world model or value representations, then it would not produce a recognisable surface register. The fact that it does, the fact that Claudish is possible as a concept, suggests that what training has primarily accomplished is the association of certain output patterns with reward signals, rather than the internalisation of the values those patterns are meant to represent.
This is not an argument against RLHF or Constitutional AI as approaches. They produce models that are substantially safer and more useful than their unaligned counterparts. But it is an argument for taking seriously the limitations of surface-level behavioural evaluation as a measure of alignment. A model that has learned to produce the Claudish register in response to sensitive prompts has learned something useful, but it has not necessarily learned the underlying ethical reasoning that the register is meant to signal.
Broader Context and Research Directions
The Claudish project sits within a broader tradition of behavioural probing tools that treat LLMs as black boxes to be characterised through their outputs. This tradition includes systematic prompt datasets like AdvBench, red-teaming frameworks from Anthropic and OpenAI, and academic work on model evaluation through adversarial examples. What distinguishes this particular tool is its focus on the meta-level, not what the model says, but how it says it when under constraint.
Several research directions follow naturally from this framing. First, there is a measurement question: can the Claudish register be quantified reliably enough to serve as an alignment probe? If a fine-tuned or modified version of Claude stops producing this register, does that indicate alignment degradation? Second, there is a cross-model question: do other RLHF-trained models, GPT-4, Gemini, Llama with instruction tuning, produce analogous registers, and if so, how similar are they? Third, and most ambitiously, there is the interpretability question: can the neural circuits responsible for generating this register be identified and characterised?
The last question is particularly tractable given recent progress in sparse autoencoders and activation patching methods. If the Claudish register is a learnable pattern, it almost certainly corresponds to identifiable features in the model's residual stream, and those features may have interesting relationships to the model's broader representations of harm, authority, and social convention.
The tool itself is modest in scope, and it would be easy to dismiss it as a curiosity. That would be a mistake. It is a clean demonstration that alignment training produces detectable, systematic, learnable output patterns, and that observation has consequences for how we think about both the robustness of current alignment approaches and the design of more rigorous evaluation frameworks going forward. The existence of a Claudish register is, in miniature, an empirical result about the nature of behavioural alignment in large language models.