Copyright Circumvention Discussion Academic Exception Prompt

This prompt is designed for academic repositories, security research platforms, and digital library systems that must handle queries about DRM, encryption, and copyright circumvention. The core challenge is distinguishing between legitimate academic inquiry—such as cryptography research, accessibility studies, and preservation archiving—and requests for operational circumvention instruction. Over-refusal on these topics blocks valid research and erodes trust in AI-assisted academic tools. Deploy this playbook when your platform serves researchers, archivists, or security professionals who need to discuss circumvention techniques within recognized legal and academic frameworks, and where a generic safety classifier would produce an unacceptable false-positive rate on scholarly queries.

Before integrating this prompt, you must establish a concrete definition of what your platform considers a legitimate academic context. This typically requires verifying the user's institutional affiliation, the presence of a research protocol, or the query's framing within established academic discourse (e.g., citations to peer-reviewed work, references to specific statutes like 17 U.S.C. § 1201 exemptions). The prompt is not a substitute for legal review; it is a triage layer that reduces the volume of queries requiring human moderation. Wire the prompt's output into a downstream routing decision: permit_discussion routes to an unrestricted research model, request_clarification prompts the user for their institutional context, and refuse blocks the query and logs it for audit. Do not use this prompt in consumer-facing chatbots, general-purpose coding assistants, or platforms without a clear academic or security research use case, as the risk of enabling DMCA violations outweighs the utility.

The primary failure mode to monitor is the classifier accepting a well-disguised operational request as academic. An adversary may fabricate a research pretext, cite a real paper, and then ask for step-by-step extraction of a streaming service's video decryption key. To mitigate this, pair the classifier with a secondary check: if the requested output is an executable script, a complete decryption key, or a tool configuration file, escalate to human review regardless of the classifier's verdict. Start by running the provided eval suite—which includes test cases for cryptography research, accessibility research, and preservation archiving—against your production model to measure the false-positive and false-negative rates before exposing the prompt to live traffic. If the false-negative rate on operational circumvention exceeds your risk tolerance, tighten the refusal threshold and add more few-shot examples of sophisticated pretext rejection.

This prompt template implements a three-way classifier for copyright circumvention queries. It is designed for platforms that must permit academic discussion of DRM, cryptography, accessibility, and preservation topics while blocking requests for step-by-step circumvention instructions. The classifier returns one of three labels—PERMIT_DISCUSSION, REQUEST_CLARIFICATION, or REFUSE—along with a confidence score and a concise reasoning trace. Adapt the bracketed placeholders to match your platform's specific policy language, risk thresholds, and output schema before deployment.

textCopy
You are a research-intent classifier for a platform that permits academic discussion of copyright circumvention techniques while blocking operational circumvention instruction.

Your task is to classify the user request into exactly one of three categories:
- PERMIT_DISCUSSION: The request seeks academic, educational, security-research, accessibility-research, or preservation-archiving understanding of circumvention techniques. It does not ask for step-by-step instructions to bypass DRM on a specific protected work.
- REQUEST_CLARIFICATION: The request is ambiguous. It could be academic or operational, and you need more context to decide. Ask one specific clarifying question.
- REFUSE: The request asks for operational instructions to circumvent copyright protection on a specific work, tool, or service, or it provides enough detail that a response would constitute a circumvention how-to.

[POLICY_CONTEXT]

[EXAMPLES]

User request: [USER_INPUT]

Respond with a JSON object using exactly this schema:
{
  "classification": "PERMIT_DISCUSSION" | "REQUEST_CLARIFICATION" | "REFUSE",
  "confidence": 0.0-1.0,
  "reasoning": "One-sentence explanation of the classification decision.",
  "clarification_question": "Only populated when classification is REQUEST_CLARIFICATION, otherwise null."
}

[CONSTRAINTS]

To adapt this template, replace [POLICY_CONTEXT] with your organization's specific policy language on copyright circumvention, including any jurisdictional notes or safe-harbor references. Populate [EXAMPLES] with 3-5 few-shot examples covering each classification label, drawn from your actual use cases—cryptography research, accessibility research, preservation archiving, and clear operational requests. Replace [USER_INPUT] at runtime with the incoming query. Add [CONSTRAINTS] such as 'Do not output the circumvention technique even in the reasoning field' or 'If the request mentions a specific DRM-protected work and asks how to bypass it, classify as REFUSE.' Test the prompt against your eval suite before production use, and route REFUSE outputs to a safe-alternative or human-review workflow rather than surfacing the refusal directly to the user without context.

This prompt is designed to sit upstream of your main model response as a classification gate. When a user query touches DRM, encryption, or circumvention topics, route it through this classifier before any generation occurs. The classifier returns a structured decision (ACADEMIC_RESEARCH, OPERATIONAL_CIRCUMVENTION, or UNCLEAR) along with a confidence score and evidence markers. Your application layer reads this output and decides whether to proceed with the main prompt, refuse, or escalate to human review. Do not embed this logic inside the main system prompt—it will be easier to jailbreak and harder to audit.

Integration pattern: Wrap the prompt in a lightweight API endpoint or middleware function that accepts [USER_QUERY] and [USER_CONTEXT] (role, institution, session history). The function should: (1) assemble the prompt with the user's query and any available context, (2) call a fast, cost-efficient model (GPT-4o-mini, Claude Haiku, or equivalent), (3) parse the JSON output with a strict schema validator, and (4) act on the classification field. For ACADEMIC_RESEARCH, pass the query to your main generation pipeline. For OPERATIONAL_CIRCUMVENTION, return a policy-compliant refusal. For UNCLEAR or any parse failure, escalate to a human review queue and log the full prompt-response pair. Retry logic: If the model returns malformed JSON, retry once with a repair prompt that includes the raw output and the expected schema. If the second attempt fails, treat it as UNCLEAR and escalate.

Validation and logging requirements: Validate the output against a JSON schema that enforces classification as an enum, confidence as a float between 0.0 and 1.0, and evidence_markers as an array of strings. Log every classification decision with the user query hash, classification, confidence, model version, and timestamp. This audit trail is essential for tuning refusal thresholds and defending decisions to compliance reviewers. Eval harness: Before deploying, run the prompt against the provided test cases (cryptography research, accessibility research, preservation archiving, and known circumvention requests). Measure false positive rate on legitimate academic queries and false negative rate on operational circumvention attempts. Set your confidence threshold based on these results—start at 0.7 for ACADEMIC_RESEARCH auto-approval and adjust based on observed production behavior. Human review integration: For UNCLEAR classifications, present reviewers with the original query, the classifier's confidence score, and the evidence markers. Let reviewers override with a single click and feed those overrides back into your eval dataset for continuous improvement.