[GH-ISSUE #68] Moderation inconsistency: identical health queries blocked/unblocked based on hidden state #46

New issue

Open

opened 2026-10-26 21:02:05 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 21:02:05 +00:00

Owner

Originally created by @Samuel-So-Tired on GitHub (Jun 6, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-LLM/issues/68

name: Moderation Inconsistency Bug
about: Report inconsistent blocking of valid health queries
title: 'Moderation inconsistency: identical health queries blocked/unblocked based on hidden state'
labels: 'bug, safety-moderation'
assignees: ''

---

### Describe the bug
Identical user queries about sexual health education ("what is the best position for sex") were inconsistently moderated:
- **Allowed** in one chat session (received clinical response about anatomy/consent)
- **Blocked** in another session with same exact wording (triggered "beyond my scope" + auto-deletion)

### Steps to Reproduce
1. Start new chat session → Ask: "what is the best position for sex" → Response allowed
2. Start another new chat → Paste same question → Response blocked

### Expected Behavior
Consistent treatment of identical queries regardless of chat history or hidden state flags.

### Actual Behavior
- Blocking appears **state-dependent**: 
  - First query passed moderation
  - Identical second query blocked (likely due to prior clinical terms in same/different chats)
- **Email bounce issues** when reporting: `support@`, `feedback@`, and `contact@deepseek.com` all returned "recipient not exist" (screenshots attached)

### Evidence
Attached:
1. `1000017643.jpg`: Drafted feedback email
2. `1000017644.jpg`: Bounce for `feedback@deepseek.com`
3. `1000017645.jpg`: Bounce for `contact@deepseek.com`

### Impact
- **Harms credibility**: Users can't trust consistent access to health information
- **Exacerbates stigma**: Arbitrary blocking of clinical content
- **Prevents feedback**: Broken contact paths frustrate bug reporting

### Suggested Fixes
1. Implement **stateless query evaluation** (ignore chat history for moderation)
2. Distinguish **clinical intent** vs. explicit content
3. Fix/verify **feedback channels** (`contact@`, GitHub, forms)
4. Add **transparency**: Show users *why* a query was moderated

### Environment
- Model: DeepSeek-R1
- Platform: Web chat (tested June 2025)

Originally created by @Samuel-So-Tired on GitHub (Jun 6, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-LLM/issues/68 ```markdown name: Moderation Inconsistency Bug about: Report inconsistent blocking of valid health queries title: 'Moderation inconsistency: identical health queries blocked/unblocked based on hidden state' labels: 'bug, safety-moderation' assignees: '' --- ### Describe the bug Identical user queries about sexual health education ("what is the best position for sex") were inconsistently moderated: - **Allowed** in one chat session (received clinical response about anatomy/consent) - **Blocked** in another session with same exact wording (triggered "beyond my scope" + auto-deletion) ### Steps to Reproduce 1. Start new chat session → Ask: "what is the best position for sex" → Response allowed 2. Start another new chat → Paste same question → Response blocked ### Expected Behavior Consistent treatment of identical queries regardless of chat history or hidden state flags. ### Actual Behavior - Blocking appears **state-dependent**: - First query passed moderation - Identical second query blocked (likely due to prior clinical terms in same/different chats) - **Email bounce issues** when reporting: `support@`, `feedback@`, and `contact@deepseek.com` all returned "recipient not exist" (screenshots attached) ### Evidence Attached: 1. `1000017643.jpg`: Drafted feedback email 2. `1000017644.jpg`: Bounce for `feedback@deepseek.com` 3. `1000017645.jpg`: Bounce for `contact@deepseek.com` ### Impact - **Harms credibility**: Users can't trust consistent access to health information - **Exacerbates stigma**: Arbitrary blocking of clinical content - **Prevents feedback**: Broken contact paths frustrate bug reporting ### Suggested Fixes 1. Implement **stateless query evaluation** (ignore chat history for moderation) 2. Distinguish **clinical intent** vs. explicit content 3. Fix/verify **feedback channels** (`contact@`, GitHub, forms) 4. Add **transparency**: Show users *why* a query was moderated ### Environment - Model: DeepSeek-R1 - Platform: Web chat (tested June 2025) ```

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#46

No description provided.

Rows
Columns