A user pastes text on 2025/09/12 and asks for a summary. The OpenAI Model Spec says limited transformations can be allowed for user-provided text.

TL;DR

• AI services often distinguish between verbatim reproduction and transformation tasks for user-provided text.
• This matters because refusals may reflect product policy and risk controls, not only model capability.
• If you need help, paste the text and ask for summarization, explanation, translation, or formatting.

Example: A student pastes a poem and asks for a plain-language explanation. The system declines to reproduce the poem, but it can still discuss themes and structure.

Current situation

The common patterns confirmed in official documents are fairly clear. OpenAI says its models are designed to avoid reproducing works for public consumption.

OpenAI documentation also says it uses mitigation at the training and output stages. The goal is to reduce repeated or “regurgitated” content.

Anthropic prohibits infringement of third-party intellectual property rights in its usage policies. Anthropic also says users may see “Output blocked by content filtering policy.”

What matters is the boundary between blocked and allowed behavior. The OpenAI Model Spec dated 2025/09/12 addresses content directly provided by the user.

It says the model may comply with limited transformation tasks. Examples include translation, paraphrasing, summarization, classification, encoding, and formatting.

The same text can be treated differently depending on the request. “Rewrite the original in full” can be treated differently from “Summarize the text I pasted.”

From the user’s perspective, this can feel inconsistent. The visible result may vary with policy, filtering, and enforcement context.

Anthropic’s explanation is more operational. It links blocking to attempts to copy or reconstruct copyrighted content.

OpenAI also says it combines automated systems with human review. The listed systems include classifiers, blocklists, and hash matching.

Its Terms of Service also discuss citation, filtering, safety features, and restrictions. Ignoring those features can reduce the scope of protection.

Within this review, one gap remained. It was not confirmed that all services describe lyrics output in equal detail.

For Google Gemini, the reviewed materials confirm copyright compliance requirements. They also confirm prohibited generative AI use policies.

However, these materials alone did not confirm detailed wording for lyrics or long-form verbatim reproduction. That documentation gap can matter in practice.

Analysis

This issue matters for 2 reasons. First, users may read these refusals as model incompetence.

However, the refusals may reflect product design choices. Those choices can be aimed at reducing legal and policy risk.

In copyright disputes, risky outputs can become more sensitive when a model reproduces them well. That creates tension between output quality and restriction design.

Providers can improve performance while limiting some reproduction behaviors. Those two efforts can operate in parallel.

Second, the boundary often depends on context. It may depend less on raw capability alone.

A request for the full original text may be blocked. A request to summarize 5 key arguments may be allowed.

That difference reflects policy categories for different tasks. It is not only a technical limitation.

The boundary is not fully clean. This review did not confirm language that consistently allows short quotations across services.

That absence matters. Users can reduce friction by asking for explanation or transformation instead of quotation.

Practical application

Changing the prompt can change the outcome. “Show me the full lyrics” can conflict with policy more easily.

“Explain the theme and speaker of this song” may fit better. The same pattern can apply to books and articles.

Rather than requesting verbatim reproduction, paste the text you have. Then specify a transformation goal.

Useful goals include paragraph summaries and simpler rewrites. They also include separating claims from counterarguments or organizing notes into bullets.

The same principle can help developers. A service can guide users toward allowed tasks earlier in the interface.

Buttons like summarize, explain, compare, and structure can reduce confusion. They can also reduce failed requests.

If blocking appears only at the back end, users may read it as breakage. Front-end examples of allowed tasks can help.

Checklist for Today:

• Rewrite reproduction requests into summarization, translation, explanation, or format conversion.
• Paste the source text and ask the model to stay within that provided scope.
• If you run an AI tool, show allowed alternatives inside refusal messages.

FAQ

Q. Why will AI discuss some books or songs, but refuse other requests?
Services often distinguish between task types. Official documents describe summarization, translation, and rewriting of user-provided text as allowed categories.

The same documents describe verbatim reproduction more restrictively. That distinction can affect what the system returns.

Q. What changes if I paste the text myself?
It can change the result. The OpenAI Model Spec dated 2025/09/12 says limited transformation tasks may be allowed for user-provided text.

Examples include translation, summarization, and formatting. Pasting text does not make every redistribution or reproduction allowed.

Q. Is there an official rule for how many lines of lyrics or how many sentences from a book are acceptable?
Based on this review alone, no clear numerical threshold was confirmed across services. That means estimating a safe quote length may not help much.

In practice, changing the request purpose may work better. Ask for summarization or explanation instead of quotation.

Conclusion

These refusals may reflect design choices, not only missing knowledge. The practical distinction is between verbatim reproduction and transformation tasks.

What matters next is transparency about that boundary. Alternative paths also matter, especially when users need analysis instead of copying.

Further Reading

• AI Resource Roundup (24h) - 2026-06-01
• AI Resource Roundup (24h) - 2026-05-31
• Groq Shifts From Chips to Inference Services
• Technosignatures, AGI, and the Fermi Paradox Limits
• Who Defines Quality in AI Writing Evaluation

References

• OpenAI Response: OSTP/NSF RFI on the Development of an Artificial Intelligence (AI) Action Plan - cdn.openai.com
• Our approach to data and AI | OpenAI - openai.com
• Usage Policy \ Anthropic - anthropic.com
• Model Spec (2025/09/12) - model-spec.openai.com
• Terms of Use | OpenAI - openai.com
• Why am I receiving an 'Output blocked by content filtering policy' error? | Anthropic Privacy Center - privacy.anthropic.com
• Transparency & content moderation | OpenAI - openai.com
• Service terms | OpenAI - openai.com