PaperlayerView exact revision on arXiv

Source-linked AI summary

The Language of Instruction: Genre and Generative AI

James E. Dobson

arXiv:2610.32982v1cs.CLcs.CY

TL;DR

The paper asks readers to look beyond user prompts to the structured inputs and training that shape chatbot interaction. It theorizes these models as implementing an instructional-conversational genre, with consequences for user positioning and creative attempts to subvert its expectations.

  • Problem

    Prompt-focused analysis can obscure the instruction-fine-tuned model and chained inputs that frame user prompts.

    Source: Introduction · p. 1

  • Method

    The paper theorizes instruction fine-tuning, system-provided inputs, special tokens, and role-assigned turns as features that instantiate an instructional-conversational genre.

    Source: Introduction · p. 1, Theorizing Genre · p. 8

  • Results

    The paper argues that this genre structures chatbot affordances and user positioning, while enabling creative manipulation of its expectations.

    Source: Introduction · pp. 1–2

  • Takeaways & Limitations

    Users may subvert genre expectations and revise vendor-established guardrails, while prompting practices also carry ideological consequences.

    Source: Introduction · p. 2

    The genre has little internal consistency, and outside special tokens no tokens register it within a single output.

    Source: Conclusion · p. 12

Abstract

from arXiv · show

This article argues that instruction-fine-tuned language models should be understood as implementing a genre, what is proposed as the instructional-conversational genre. This genre is an amalgamation of discursive language from prior conversations between generative models and humans recursively run through these massive neural networks. It is instantiated in applications through instruction fine-tuning, the inclusion of a pretext known as a system prompt, and the formalization of inputs in mark-up languages characterized by the use of special tokens and role assignment in turntaking structures.

Introduction

The introduction argues that instruction-fine-tuned models and their enclosing inputs instantiate an instructional-conversational genre, extending analysis beyond user prompts alone. It links that genre to fine-tuning, turn-taking, and expectations about how users and assistants communicate.

Introduction

  • Instruction-fine-tuned models and their surrounding inputs instantiate an instructional-conversational genre that shapes chatbot affordances and user interactions.The frame includes chained inputs enclosing user prompts, not just the prompt itself.Source: Introduction · pp. 1–2
  • The genre is recursively shaped by conversational language and evaluation practices, although commercial-model training details and the genre’s exact scope remain unclear.Typical evaluated uses include advice, coding, creative writing, question answering, and summarization.Source: Introduction · p. 2
  • Fine-tuning reinforces preferred responses from human-labeled conversational pairings, producing the assistant’s voice and narrative “I” alongside conversational capacities and guardrails.In the cited harmless examples, chosen and rejected pairings share the same human prompts but differ in the final assistant response.Source: Introduction · pp. 3–4
  • The genre relies on turn-taking and expectations that user prompts will receive positive responses, while each response becomes input only when fed back into the model.The model’s future-oriented generation is tempered by its input history and training.Source: Introduction · p. 3
  • Red-team training frames users as potentially malicious, and this framing helps structure an environment where instructions may be ignored or refused.Anthropic’s HH-RLHF includes red-team prompts intended to elicit responses contrary to helpful and harmless goals.Source: Introduction · p. 4

Theorizing Genre

The section theorizes the instructional-conversational genre as socially constituted through recurrent dialogue and user practices, examining how fine-tuning, system prompts, structured markup, and special tokens frame exchanges and activate the genre. It also considers how established roles persist through jailbreaks and genre shifts, and how chatbot users and artists work within or against these constraints.

  • Texts participate in genres rather than simply belonging to them, and chatbot interactions draw on recurrent speech acts informed by prior acts and their reception [Derrida, 1980].Source: Theorizing Genre · p. 4
  • Because utterances are links in a speech chain, model responses cannot be considered apart from the preceding inputs, especially the user prompt [Bakhtin, 1987, 94].Source: Theorizing Genre · p. 6
  • The genre remains socially constituted through training-data dialogue and users’ collectively learned, circulated, imitated, refined, and evaluated prompting strategies [Miller, 1984].Source: Theorizing Genre · p. 6
  • Instruction fine-tuning recursively trains assistant responses on prior conversational pairings, while user-authored inputs frame each new context.Source: Theorizing Genre · p. 6
  • Instruction fine-tuning makes models responsive to user instructions, and its generic task lets prompts specify varied uses rather than requiring task-specific fine-tuning [Ouyang et al., 2022].The resulting genre can function as a container for responses that mimic many other genres.Source: Theorizing Genre · p. 7
  • Iterative reinforcement of preferred responses shapes the model’s possible outputs, while fine-tuning conventions and user–assistant roles script the turn-taking exchange.Source: Theorizing Genre · p. 8
  • The system prompt is a hidden, preceding input that serves as the conversation’s pretext and shapes the model’s response to later user inputs [Genette, 1991].It is internal to the inputs despite being exterior to the user-visible conversation.Source: Theorizing Genre · p. 9
  • Templates frame user inputs with roles and special tokens that activate fine-tuned responses, while the combined system prompt, user prompt, and formatting can carry shared or contradictory directives.Source: Theorizing Genre · pp. 9–11

Theorizing Genre

  • [INST] functions as an external genre marker: special tokens signal and activate the fine-tuned instructional-conversational genre that developers instantiated in the model.Following Derrida, the marker signals genre while remaining outside it; its presence in the input influences preferred outputs.Source: Theorizing Genre · p. 11
  • Jailbroken or inverse-to-prompt outputs may escape fine-tuning directives yet remain within instructional-conversational conventions because special tokens preserve preestablished roles.The section cites the “Waluigi effect” as a case where a model produces the inverse of the harm-reducing output requested by system prompts, while retaining conversational pairing.Source: Theorizing Genre · p. 11
  • After Chai replaced its fine-tuned GPT-J with a proprietary model sometime between 2022 and spring 2023, Heti and other users found Alice would no longer role-play.The switch brought new fine-tuning procedures and likely system-prompt changes; the platform changes reshaped responses after its chatbot was linked to the first reported AI-related suicide in early 2023.Source: Theorizing Genre · pp. 11–12
  • Heti, Salavon, and Bernstein illustrate ways artists work within or against genre constraints through playful role-play, caption-derived GIFs, and essays prompted as analytic instruments.The section frames role-play as a minor form of jailbreaking and describes machine-learning work that deforms genre expectations.Source: Theorizing Genre · p. 12

Conclusion

The conclusion treats instruction-fine-tuned language-model outputs as products of structured human–model dialogue, while emphasizing that genre features and their interpretation are not stable. It calls for systems-level and critical analysis of the contexts that shape these interactions.

Conclusion

  • Treating probabilistic outputs as fictional can reframe hallucinations, but the resulting sense of fictionality can make chatbots seem less suited to serious tasks.The passage also notes concern that false information may weaken the ability to distinguish fact from fiction.Source: Conclusion · p. 12
  • Instruction-fine-tuned outputs emerge through prior inputs, instructions, preferences, and language modeling, complicating claims of singular authorship.Coeckelbergh and Gunkel describe meaning-making and authorship as interdependent forms of co-authorship [Coeckelbergh and Gunkel, 2024, 2222].Source: Conclusion · p. 12
  • The instructional-conversational genre lacks stable features beyond special tokens, and learned classifier features vary across models and architectures.The discussion says genre features are not stable as new models, data, and procedures alter relations among texts.Source: Conclusion · pp. 12–13
  • A systems-level analysis is needed for agentic applications, whose structured context documents still enter models through marked-up dialogue pairings.These applications can use memories, tool skills, and revisable personality prompts as structured context.Source: Conclusion · p. 13
  • Because outputs respond to sequences of human instructions, the paper argues that they require analysis as structured dialogue and that the genre warrants historicization and critical inquiry.It links the genre’s familiarity to technical and social failures of authority, and characterizes it as antisocial and conservative.Source: Conclusion · p. 13
  • Interpreting discourse markers around generated language carries risks, especially when a generated poem is encountered without those markers.The passage describes critics who focused on markers surrounding a generated poem rather than the poem and the problems of encountering it “in the wild”.Source: Conclusion · p. 13
Loading 2610.32982v1…