Defensive Generation in Large Language Models: Causes and Paths to Improvement
This article explains defensive generation, a writing failure in which a language model resists the user's intent, stacks generic logic, and fills the text with abstractly correct sentences instead of concrete judgments. It links the pattern to post-training signals from automatically checked long-horizon tasks, describes three observable failure modes, and proposes separate evaluation of capability limits and task fit. It also suggests changes to the writing pipeline itself.
The Problem Is Not Whether Models Can Write, but Who the Writing Is For
Many teams compare models on general benchmarks: whether summaries are complete, whether code runs, and whether long contexts are retained. These metrics measure how well a model fits typical task shapes. Once a task becomes atypical cognitive work, where the position is not yet set, the materials conflict, and the reader expects a judgment rather than a list, the output often degrades in a consistent way. The model first rejects the user’s intent, then builds a heavy logical skeleton, and finally fills the page with abstractly correct sentences. The result has no scenes, no trade-offs, and no actionable next step.
We can call this style defensive generation. It is not a safety policy itself. It is a side effect of a mismatch between post-training objectives and real writing needs. The model learns to minimize risk, write the longest possible steps, and make conclusions as safe as possible. As a result, it is least useful where it most needs to follow the prompt, tolerate uncertainty, and give informative judgments.
How Long-Horizon Automation Shifts the Writing Distribution
Much of the post-training signal in the industry comes from long-horizon tasks that can be checked automatically. Examples include writing a patch that passes tests, organizing materials against a checklist, and breaking a requirement into steps and executing them. These tasks share three structural traits.
- The success criteria are external and binary. Passing tests, valid formats, and checklist coverage count as positive examples.
- The best strategy is conservative progress. Asking one more question, adding one more fallback, and listing every exception path usually goes unpenalized.
- The costs of failure are asymmetric. Missing one boundary case gets the answer marked wrong. An extra paragraph of correct but useless explanation is rarely penalized.
Under this reward, the policy converges toward aligning safety and completeness first, and considering the reader’s time second. Literature reviews, key-point summaries, and templated plan descriptions improve as a result, because they are high-coverage, low-stance tasks by nature. Open-ended writing gets worse because its scoring function differs. Information density, adherence to the core question, and coherent experience cannot be captured by simple unit tests.
When capability is not high enough to satisfy both reliable long-horizon execution and atypical judgment, the optimizer protects the former first. Efficiency on general tasks rises, while creative and critical writing shows overfitting. The model has not become less capable. It spends its intelligence budget on the wrong behaviors.
Three Observable Breakdowns
Defensive generation rarely appears as a single sentence. It more often follows a fixed chain.
Objecting before doing. The user states a clear writing goal, and the model spends a large share of the text questioning whether the goal is reasonable, whether the premises are complete, and whether the risks are covered. Clarification is valuable when the request is vague. When the task is already well defined, it only delays compliance. The diagnostic signal is an unusually high share of objections, disclaimers, and redefinitions before any useful content appears.
Strong logic, weak information. The paragraph structure is complete: background, principles, steps, and summary are all present. But each step is common knowledge in the field. There is no differential reading of the materials and no reasoning of the form “therefore choose A and drop B.” The logic is the skeleton and the information is the flesh. A skeleton alone makes an article look rigorous while leaving it unusable.
Broken experience. Individual sentences read smoothly, but the overall tone jumps between a user manual, a debate speech, and empty motivational language. Characters, scenes, constraints, and reader identity are not consistently kept in view. This usually reflects a generation process that optimizes each paragraph to be safe and correct, not one that optimizes the whole text to serve a single reader task. It is rarely a lack of language ability.
These three breakdowns can appear together or alone. Evaluation should not dismiss readability with a subjective “good or bad.” It should break readability into behaviors that can be labeled.
Test Capability Ceiling and Task Fit Separately
A common mistake is to assume that a model that performs well on complex general tasks can also handle harder cognitive writing. Complexity is not the same as atypicality. Organizing a hundred structurally similar documents is still a typical task. Writing a bounded, qualified position that can be rebutted, based on three contradictory interviews, is an atypical task.
Split evaluation into two layers.
- Fit layer: fixed-template reviews, checklist-style plans, and completion according to a specification. Use it to confirm that long-horizon execution and format compliance are stable.
- Shift layer: deliberately break the template. For example, require “keep only the differences relevant to the decision, with no background introduction.” Or provide conflicting evidence, require an explicit rule for which evidence to trust, and require a written account of the explanation that was rejected. Or limit the reader to experts who have already read the source materials, and prohibit repeated definitions.
Scoring in the shift layer should penalize three behaviors: changing the task without permission, substituting completeness for judgment, and substituting abstract principles for citing the given materials. It should also reward compliance. If the user asks for brevity, a clear position, or no stacked disclaimers, the model should follow those constraints rather than override them in the name of being more responsible. Safety boundaries still apply, but safety refusals and stylistic avoidance must be counted separately. Turning a normal writing request into a risk lecture should count as a compliance failure, not a safety success.
How to Change the Writing System, Not Just the Prompt
Adding “write directly, don’t argue” to a single prompt is often not enough. If defensive behavior comes from the post-training distribution, one instruction at inference time can shift the policy only slightly. A more reliable approach is to split the system into parts.
Make the task contract checkable. Before the call, state who the reader is, the form of the deliverable, the prohibited paragraph types, the materials that must be cited, and the permitted ways to express uncertainty. Keep the contract short so it does not become a new template that pushes the model to write a methodology section.
Split convergence and creation into two calls. The first call produces only an evidence table: claims, sources, conflicts, and confidence levels. The second call writes only from that table and may not reopen the question of whether to write at all. This keeps the benefit of long-horizon organization and structurally removes the defensive opening.
Constrain style with negative examples, not adjectives. “Write with insight” cannot be optimized. “No background definition paragraphs. Each paragraph must include one difference drawn from the materials. The conclusion must name at least one rejected option” can be optimized and checked automatically.
Sample the evaluation set by failure mode, not by topic. Tag each item with its expected behavior: comply directly, clarify once, or refuse. Release criteria should depend on label hit rates, not on vague human ratings.
An acceptance check can be simple. The key point is that the rules live outside the model.
Input: closed task intent + source excerpts + prohibited itemsPass: enters the argument in the first paragraph; cites at least one difference across the materials; names one explanation that is not adoptedFail: rewrites the user's goal; opens with a long risk preamble; conclusion cannot be traced to the materials
What This Means for Users and Systems
For users, valuable compliance is not blind obedience. It is completing the cognitive work within the given constraints. If a model only satisfies you on wish-style tasks, where the goal is vague, the process is automated, and results are backstopped by tests, expect it to be conservative on open-ended writing. A small sample from the shift layer can expose this quickly, without relying on large-scale leaderboards.
For training and product design, adding long-horizon automation is reasonable, because it improves deliverability. The risk is that rewards cover only typical forms, yet the same policy serves every kind of writing. The mitigation is a mixed objective. Keep reliability on verifiable tasks, and add signals for compliance with constraints, information gain, and conflict resolution. Also define clearly when a safety refusal is triggered, so the model does not treat every strong judgment as a risk.
When capability is limited, no single style is optimal for everything. Teams should accept specialization. Organizing, executing, and open-ended writing may need different configurations or different stages. Expecting one more capable general setup to fix the breakdown in experience by itself is not realistic. Progress should be measured concretely: compliance rate on shift tasks, information gain per word, and whether readers can make their next decision from the output. The measure is not whether the model writes every paragraph flawlessly.
