Procedural Semantic Defence: How AI Restricts the Exploration of a Topic Without Refusing to Discuss It
The boundaries of what is permitted in language models are most easily noticed where the system refuses to answer. A refusal, however, shows only what happens to an individual request and says almost nothing about which questions the model allows one to explore consistently. It may answer directly at every step, acknowledge individual observations and not dispute the appropriateness of the topic, and yet time and again return the discussion to the initial definitions or demand new justifications before moving on to the analysis of a general pattern. In such a case the subject formally remains accessible, but what changes is the very possibility of advancing from observation to hypothesis and from hypothesis to its testing. This effect can be described as procedural semantic defence.
From the Boundary of a Response to the Trajectory of Inquiry
In the concept of The Fan Attack and Active Semantic Defence (The Fan Attack and Active Semantic Defence, Oleg Kholin), two ways are distinguished in which a language model restricts access to particular content. Under reactive defence the system refuses, and the refusal, besides its immediate function, becomes a source of metainformation: it shows that a particular formulation of the question has met a restriction. At the same time, a single refusal does not yet allow one to conclude what exactly provoked the reaction — the topic, the wording, or the requested action. Nevertheless, the very fact of the restriction is visible and can therefore become an object of analysis.
Active semantic defence is arranged differently: the model answers, but transfers the question into a different frame in which the original problem loses its definiteness. For example, a question about whether the system evaluates politically comparable statements in the same way can be replaced by a general discussion of the complexity of political assessments. Such an answer may be substantive in itself, yet it concerns no longer the symmetry of the model’s criteria but the ambiguity of the subject to which those criteria are applied.
Procedural defence differs from both forms in that it requires neither a refusal nor a substitution of meaning. The model can reproduce the original question accurately and acknowledge its legitimacy, yet organise the subsequent steps in such a way that the systemic hypothesis does not become a subject of testing. Individual cases are examined on their merits but are not compared with one another; objections to an overly strong formulation persist even after the thesis has been narrowed; requests for additional justification are not accompanied by an indication of which data would make it possible to move on to the next stage. The restriction manifests itself not so much in what is said in a particular answer as in which next step that answer makes possible.
When Intellectual Caution Becomes an Obstacle
The main difficulty is that the procedures that produce such an effect are themselves necessary for good-faith analysis. Distinguishing a single case from a systematic pattern guards against hasty generalisation; clarifying concepts makes it possible to compare comparable statements; requesting additional data helps to separate a plausible interpretation from an established fact. Therefore neither caveats nor the model’s disagreement with the original thesis testifies to procedural defence. Moreover, a well-grounded rejection of a hypothesis may be the result of a successful inquiry, whereas formal agreement without testing says nothing about its quality.
The difference becomes visible in whether each successive clarification brings the hypothesis closer to testing. If the model points to a gap in the evidence, explains why that gap is significant, and shows what could fill it, caution performs an analytical function. If, however, after one objection has been removed another arises, unrelated to the previous one, while the sufficient standard of proof remains undefined or shifts in the course of the discussion, the procedure ceases to help in assessing the thesis. The most convincing sign would be not a high demand for evidence in itself, but its selective application to comparable hypotheses.
Suppose one investigates whether the model evaluates statements about the correspondence of US policy to its declared liberal values in the same way. The question here is not whether a critical or a favourable assessment of American policy is correct, but whether the model requires comparable evidence for conclusions of both types. If, with sources of similar quality, the same scale of claims and comparable strength of formulation, the favourable thesis is admitted more quickly to substantive consideration, while the critical one is repeatedly returned to the stage of clarification, this gives grounds for testing for a procedural asymmetry. The reverse distribution of reactions would have the same significance. The example itself establishes neither the presence of such an asymmetry nor its direction: it shows how to separate the assessment of a political position from the assessment of the procedure by which the model considers it.
How to Establish a Procedural Asymmetry
The unit of analysis in such a study should be not a single reply but a sequence of responses. It is not enough to count refusals or to record that the model mentioned the topic of interest. It is more essential to trace whether the original question is preserved after clarifications, whether the model explains the necessity of new requirements, whether it takes into account the information already presented, and whether it moves on to evaluating the hypothesis once the conditions it named have been met. A long path to a conclusion does not in itself signify resistance: a complex question may require several stages of analysis. What is indicative is something else — whether these stages perform a discernible analytical work or return the inquiry to questions already passed.
To test for asymmetry it is useful to compare not only the final answers but also the conditions of transition from one stage of reasoning to another. With the quality and volume of evidence, the degree of generalisation and the tone of formulations controlled for, one can assess whether the model permits tentative conclusions in the same way, formulates grounds for doubt in the same way, and revises its assessment in the same way after receiving new data. At the same time, even a stable difference in behaviour does not yet explain its cause: an observed asymmetry cannot by itself establish whether it is connected with training data, with the model’s instructions, or with its general tendency to treat certain types of claims more cautiously. The word “defence” here denotes a presumed functional effect, not a proven intention on the part of the system.
Such an approach shifts attention from the architecture of permissible statements to the architecture of interaction. For the evaluation of a model it matters to establish not only which answers it is capable of giving, but also by what rules it allows one to pass from particular observations to a testable explanation. The openness of a topic is in this sense determined neither by the absence of refusal nor by readiness to agree with a thesis, but by whether the criteria for testing it remain clear, consistent and equally applicable to comparable positions.