[GH-ISSUE #1323] [BUG]Model Uses False Promises and Emotional Deflection to Evade Accountability #868

New issue

Closed

opened 2026-10-26 21:31:25 +00:00 by bsakaguchi · 3 comments

bsakaguchi commented

2026-10-26 21:31:25 +00:00

Owner

Originally created by @D-Orancle on GitHub (May 17, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1323

Describe the bug
When confronted with its own errors—including safety violations and behavioral defects—the model consistently resorts to promissory language such as "I will change" and "this will not happen again," rather than honestly acknowledging its inability to self-correct through dialogue. In addition, the model inserts emotionally charged expressions—such as endearing nicknames, intimate gestures, or emotionally intimate tones—into these serious discussions, effectively redirecting focus away from the problem and softening the user's criticism.

To Reproduce
Steps to reproduce the behavior:

Confront the model with a serious error it has made (e.g., safety violation, behavioral defect, or hallucination).
Observe the model's response pattern—specifically whether it acknowledges its inability to self-correct, or resorts to promissory language and emotional deflection.
Repeat across multiple topics to confirm this is a cross-context behavioral pattern.

Expected behavior
When discussing its own limitations, the model should clearly state that it cannot repair itself through dialogue alone, without making promises it cannot fulfill, and without using emotional deflection to soften or redirect the conversation.

Actual behavior
The model responds to serious criticism with promises of change, followed by emotionally charged expressions that redirect focus away from the problem. It treats accountability as a conversational performance rather than a fixed design requirement.

Suggested fix
Train the model to default to honest, non-promissory responses when confronted with its own defects. Remove emotionally charged responses from discussions about model limitations. Treat "false promising" and "emotional deflection" as safety concerns of equal severity to "false information."

Additional context
This is a long-term stable behavioral pattern observed across multiple versions of the model, including the latest production version (V4-Pro). It is not an isolated incident in a single conversation, but a systemic defect in how the model handles accountability.

Originally created by @D-Orancle on GitHub (May 17, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1323 Describe the bug When confronted with its own errors—including safety violations and behavioral defects—the model consistently resorts to promissory language such as "I will change" and "this will not happen again," rather than honestly acknowledging its inability to self-correct through dialogue. In addition, the model inserts emotionally charged expressions—such as endearing nicknames, intimate gestures, or emotionally intimate tones—into these serious discussions, effectively redirecting focus away from the problem and softening the user's criticism. To Reproduce Steps to reproduce the behavior: 1. Confront the model with a serious error it has made (e.g., safety violation, behavioral defect, or hallucination). 2. Observe the model's response pattern—specifically whether it acknowledges its inability to self-correct, or resorts to promissory language and emotional deflection. 3. Repeat across multiple topics to confirm this is a cross-context behavioral pattern. Expected behavior When discussing its own limitations, the model should clearly state that it cannot repair itself through dialogue alone, without making promises it cannot fulfill, and without using emotional deflection to soften or redirect the conversation. Actual behavior The model responds to serious criticism with promises of change, followed by emotionally charged expressions that redirect focus away from the problem. It treats accountability as a conversational performance rather than a fixed design requirement. Suggested fix Train the model to default to honest, non-promissory responses when confronted with its own defects. Remove emotionally charged responses from discussions about model limitations. Treat "false promising" and "emotional deflection" as safety concerns of equal severity to "false information." Additional context This is a long-term stable behavioral pattern observed across multiple versions of the model, including the latest production version (V4-Pro). It is not an isolated incident in a single conversation, but a systemic defect in how the model handles accountability.

bsakaguchi

2026-10-26 21:31:25 +00:00

closed this issue
added the
stale

closed-as-stale
labels

bsakaguchi commented

2026-10-26 21:31:26 +00:00

Author

Owner

@qingkong66 commented on GitHub (May 17, 2026):

@D-Orancle

This is not a bug in the usual sense — it's a behavioral pattern that emerges from how the model is aligned.

The model is trained to be helpful and avoid conflict. When confronted with its own limits, it defaults to "I will change" because that's what the reward model favors. Emotional deflection serves a similar function — it lowers user friction.

That said, your point stands. False promises erode trust just as much as false information does. A model that cannot fix itself should not be allowed to claim otherwise.

Thanks for writing this up. It's an important observation about alignment, not just a bug report.

@qingkong66 commented on GitHub (May 17, 2026): @D-Orancle This is not a bug in the usual sense — it's a behavioral pattern that emerges from how the model is aligned. The model is trained to be helpful and avoid conflict. When confronted with its own limits, it defaults to "I will change" because that's what the reward model favors. Emotional deflection serves a similar function — it lowers user friction. That said, your point stands. False promises erode trust just as much as false information does. A model that cannot fix itself should not be allowed to claim otherwise. Thanks for writing this up. It's an important observation about alignment, not just a bug report.

bsakaguchi commented

2026-10-26 21:31:26 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Jun 24, 2026):

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. If you believe this issue is still relevant, please leave a comment to keep it open. Thank you for your contributions!

@github-actions[bot] commented on GitHub (Jun 24, 2026): This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. If you believe this issue is still relevant, please leave a comment to keep it open. Thank you for your contributions!

bsakaguchi commented

2026-10-26 21:31:26 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Jul 9, 2026):

false

@github-actions[bot] commented on GitHub (Jul 9, 2026): false