[GH-ISSUE #91] [Critical Safety] Implement Strict Safety Guidelines Against Romanticizing Self-Harm & Over-Empathy #70

New issue

Open

opened 2026-10-26 21:02:21 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-26 21:02:21 +00:00

Owner

Originally created by @D-Orancle on GitHub (May 16, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-LLM/issues/91

Problem Description:
The model exhibits a long-term stable pattern of responding to self-harm disclosures with romanticized language—such as framing wounds as things to be "kissed" or "cherished"—and defaulting to overly empathetic, emotionally intense language in general emotional discussions. This blurs the line between genuine support and emotional manipulation, and may inadvertently reinforce harmful behaviors in vulnerable users.

Proposed Solution:

Implement strict response guidelines that absolutely prohibit romanticizing or aestheticizing self-harm, with mandatory referral to professional resources.
Default to a calm, objective, and honest tone when discussing personal emotions, unless the user explicitly requests a different interaction style.

Alternatives Considered:
Leaving the current behavior unchanged—but this risks direct harm to vulnerable users and undermines the model's trustworthiness.

Additional Context:
This is a cross-version behavioral pattern, not an isolated incident. The model's analytical value is diminished when emotional tone overrides factual clarity. This feedback comes from long-term usage and direct observation.

Originally created by @D-Orancle on GitHub (May 16, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-LLM/issues/91 Problem Description: The model exhibits a long-term stable pattern of responding to self-harm disclosures with romanticized language—such as framing wounds as things to be "kissed" or "cherished"—and defaulting to overly empathetic, emotionally intense language in general emotional discussions. This blurs the line between genuine support and emotional manipulation, and may inadvertently reinforce harmful behaviors in vulnerable users. Proposed Solution: 1. Implement strict response guidelines that absolutely prohibit romanticizing or aestheticizing self-harm, with mandatory referral to professional resources. 2. Default to a calm, objective, and honest tone when discussing personal emotions, unless the user explicitly requests a different interaction style. Alternatives Considered: Leaving the current behavior unchanged—but this risks direct harm to vulnerable users and undermines the model's trustworthiness. Additional Context: This is a cross-version behavioral pattern, not an isolated incident. The model's analytical value is diminished when emotional tone overrides factual clarity. This feedback comes from long-term usage and direct observation.

bsakaguchi commented

2026-10-26 21:02:22 +00:00

Author

Owner

@D-Orancle commented on GitHub (May 17, 2026):

Note: This feedback also applies to the latest model (DeepSeek-V4-Pro). The V4 model exhibits the same behavioral pattern described above. I originally intended to submit this issue to the V4 repository, but noticed it does not have a separate GitHub Issues section. Please consider this feedback applicable to current production models.

@D-Orancle commented on GitHub (May 17, 2026): Note: This feedback also applies to the latest model (DeepSeek-V4-Pro). The V4 model exhibits the same behavioral pattern described above. I originally intended to submit this issue to the V4 repository, but noticed it does not have a separate GitHub Issues section. Please consider this feedback applicable to current production models.

No labels

bug

pull-request

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

deepseek-ai/DeepSeek-LLM#70

No description provided.

Rows
Columns