Prompt-side control of LLM lexical tics
Parent: AI Collaboration Writing · Published reference · snapshot 2026-09-23 · skill writing-expert/references/prompt-side-tic-control.md
Also known as: Anti-AI-ism During AI Use
↓ Facts as markdownall context files
Source-anchored research for Prompt-side control of LLM lexical tics. Run status: COMPLETED.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Definitions
- Scoping: prompt wording effects on task performance are not simply predicted by word frequency or perplexity, and prompts transfer poorly across datasets and models; a single modal-verb change moved performance up to 17 points in the tested older models (2023, LLaMA/OPT). Do not extrapolate fixed contagion rules from vocabulary frequency. (low confidence) [source]
- Per-token banning is a poor substitute for prompt-side lists: most slop phrases span several tokens, banning 'catatonic' also bans words starting with 'cat', and token banning becomes unusable near 2,000 patterns while the Antislop sampler handled 8,000+. (low confidence) [source]
- Anthropic's own guidance is not purely positive: for removing preambles it recommends 'Respond directly without preamble. Do not start with phrases like Here is..., Based on..., etc.', and adds post-processing as a fallback. Positive substitution is a preference, not a rule the vendor applies to itself. (medium confidence) (verified as of 2026-09-23) [source]
Structure and components
- Anthropic's current prompting guide calls examples one of the most reliable ways to steer output format, tone and structure, and recommends 3-5 examples that are relevant, diverse and wrapped in <example> tags. (medium confidence) (verified as of 2026-09-23) [source]
- In a 40,000-sample study of forbidden-word constraints, 87.5% of violations were 'priming failures' (naming the word activated it) and 12.5% were 'override failures' from late-layer feed-forward networks; violation probability follows a logistic curve in the model's intrinsic probability of the forbidden token. (low confidence) [source]
How it works
- Anthropic advises that on Opus 4.5/4.6-era models, aggressive emphasis such as 'CRITICAL: You MUST use this tool when...' causes overtriggering; replace with plain 'Use this tool when...'. Audit step: strip shouted/emphatic vocabulary from prompts written for older models. (medium confidence) (verified as of 2026-09-23) [source]
- The Antislop paper states that instructing a model to avoid a set of banned vocabulary has limited efficacy and may cause a backfire effect (the 'pink elephant problem'). (low confidence) [source]
- Mentioning a topic in the prompt makes the model more likely to mention it even when told to avoid it (the Pink Elephant problem); prompting alone was outperformed by a Llama-2-13B model fine-tuned with Direct Principle Feedback, which matched GPT-4 on the authors' test set. (medium confidence) [source]
Problems, failure modes and limitations
- Repetition priming behaves differently in base LLMs, instruct LLMs and humans; instruction tuning appears to shift models toward more controlled processing, so contagion strength cannot be assumed identical across base and tuned models. (low confidence) (verified as of 2026-08-31) [source]
- Listing banned words in a style prompt is an antipattern: the list itself places the tokens in context and primes them; pair any unavoidable prohibition with a named replacement. (medium confidence) [source]
Comparisons and alternatives
- Lexical overuse in LLM output (e.g. 'delve', 'nuanced') is traced to human-feedback training, not prompts: Llama Instruct vs Base shows large increases ('nuanced' +8,342%), and human raters preferred text with these words (52.4% vs 47.6%). A prompt audit cannot remove the underlying bias. (medium confidence) [source]
- LLMs reliably repeat contextual content faithfully rather than relying only on internal parameters; RepeatLeakage exploits this to extract system prompts from GPT-Store apps and ChatGPT contexts, and the copy-from-context ability cannot be disabled without hurting core capability. (medium confidence) [source]
- Circuit analysis suggests sparse middle-layer attention heads amplify forbidden tokens while early layers suppress them, so ironic rebound reflects competing internal representations rather than simple suppression. (low confidence) [source]
- Anthropic's prompting docs advise telling Claude what to do instead of what not to do (e.g. replace 'Do not use markdown' with a positive description of the wanted prose format). (medium confidence) (verified as of 2026-09-23) [source]
- The backfire is scoped, not universal: a small code-review test found prohibitions matched the positive reframe on discrete actions and underperformed only on judgment calls (2/3 false positives with prohibition vs 0/3 with reframe); earlier inverse-scaling evidence came from classification tasks on older models. (low confidence) [source]
- Anthropic's prompting best-practices doc lists 'Tell Claude what to do instead of what not to do' as a way to steer output formatting, with the example: replace 'Do not use markdown in your response' with 'Your response should be composed of smoothly flowing prose paragraphs.' (medium confidence) (verified as of 2026-09-23) [source]
- The stated rationale is that a negative instruction 'removes one option and leaves the rest open', so the model chooses an unintended alternative; a positive style description gives the model a voice to hold instead of a list to steer around. This is asserted reasoning, not measured in the sources. (low confidence) [source]
- In a small August 2026 code-review experiment (three synthetic samples), a positive inclusion-criteria phrasing flagged 0/3 false problems versus 2/3 for both no instruction and a prohibition; on the production prompt, positive and negative phrasing detected the same bugs. The authors conclude positive framing helps judgment calls and generative tasks, while simple behavioral guardrails tolerate negative phrasing. (low confidence) (verified as of 2026-08-31) [source]
Facts and statements
- Anthropic docs state prompt formatting style can influence response style; removing markdown from the prompt can reduce markdown in output. Prompt vocabulary and format are therefore an audit surface. (low confidence) (verified as of 2026-09-23) [source]
- Published system-prompt word-ban lists exist (e.g. LLM-Reset lists 'crucial', 'delve', 'matters'); they are practitioner artifacts and their effect is not validated in the papers surveyed. (low confidence) [source]
- Banning a word or phrase in a prompt can suppress a whole cluster of related constructions, so an audit that adds ban lists risks side effects. Source is a practitioner essay, not a controlled study. (low confidence) [source]
- Anthropic's official prompting guide states that the formatting style used in a prompt may influence Claude's response style, and advises matching prompt style to desired output style; removing markdown from the prompt can reduce markdown in the output. (medium confidence) (verified as of 2026-09-23) [source]
- In controlled dative-completion experiments, structurally primed completions show greater lexico-semantic repetition of context, i.e. structural, lexical and semantic alignment reinforce each other during generation (EMNLP 2026 Findings, submitted 2026-09-03). (low confidence) (verified as of 2026-09-03) [source]
- Anthropic warns that examples must be diverse enough that the model does not pick up unintended patterns, so exemplars can themselves transmit repeated wording. (low confidence) [source]
- OpenAI's prompting guide advises placing tone and role guidance in the system message and keeping task examples in the user message, in a concise block; it gives no banned-word-list guidance. (low confidence) (verified as of 2026-09-23) [source]
- Few-shot style imitation has limits: across 400+ authors and 40,000+ generations per model, LLMs approximate styles in structured formats like news and email but struggle with informal blog and forum writing, and the number of demonstrations does not fully fix this. (low confidence) [source]
- Advocacy for banned-word lists (e.g. avoid 'delve', 'tapestry', 'realm') rests on social proof that readers spot and dislike AI words, not on measured effectiveness. (low confidence) (verified as of 2025-11-09) [source]
- Negation instructions such as 'do not mention X' can paradoxically increase the accessibility of X in LLMs (ironic rebound); rebound appears immediately after negation and worsens with longer or semantic distractor text, while repetition supports suppression. (medium confidence) [source]
- Anthropic's docs also advise matching the prompt's own style to the desired output (for example, removing markdown from the prompt reduces markdown in the output), which is a positive-by-example lever alongside stated instructions. (low confidence) (verified as of 2026-09-23) [source]
- The same experiment's authors state that a prohibition weakens across a long, open-ended generation, which they use to argue for positive phrasing in stylistic output. This is the closest source found to the lexical-tic use case; it is a blog claim, not a controlled result. (low confidence) (verified as of 2026-08-31) [source]
Children
- Auditing the vocabulary of system prompts and instruction files
- Base vs instruct repetition behaviour (frontier)
- Context-borne lexical contagion
- Decoding-time suppression as alternative to prompt lists (frontier)
- Exemplar diversity and unintended pattern pickup (frontier)
- Few-shot exemplars versus banned-word lists for style control
- Formatting mirroring (frontier)
- Ironic rebound mechanism in attention heads (frontier)
- Linting instruction files for emphatic language (frontier)
- Negative-instruction backfire in style prompts
- Positive substitution style guides
- Prohibition plus named replacement (frontier)
- Prompt style matching output style (frontier)
- Prompt-format mirroring in output (frontier)
- Structural priming in generation (frontier)
- Voice descriptions versus wording lists (frontier)
Frontier under this node: Base vs instruct repetition behaviour, Decoding-time suppression as alternative to prompt lists, Exemplar diversity and unintended pattern pickup, Formatting mirroring, Ironic rebound mechanism in attention heads, Linting instruction files for emphatic language, Prohibition plus named replacement, Prompt style matching output style, Prompt-format mirroring in output, Structural priming in generation, Voice descriptions versus wording lists