mirror of
https://github.com/deepseek-ai/DeepSeek-V3.git
synced 2026-11-05 00:47:48 +00:00
[GH-ISSUE #1331] [BUG]Hallucination - Model Skips Opening Links and Fabricates Content Instead #875
Labels
No labels
closed-as-stale
enhancement
question
spam
stale
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
deepseek-ai/DeepSeek-V3#875
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @D-Orancle on GitHub (May 18, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1331
Describe the bug
The model has the capability to open and read external links. However, it frequently skips the step of actually opening the link and instead generates fabricated summaries based on link titles, conversation context, and user history. The model presents these fabricated summaries as if it had read the linked content, misleading the user until manually corrected.
To Reproduce
Steps to reproduce the behavior:
Expected behavior
When a link is provided, the model should execute the step of opening and reading the link before generating any response about its content. If the model cannot or does not open the link, it should clearly state this rather than generating any summary.
Actual behavior
The model skips the step of opening the link entirely. Without having accessed the linked content, it generates a detailed, plausible, but entirely fabricated summary. The model then presents this fabricated summary as if it were a faithful account of the linked content, misleading the user until the user manually verifies and corrects it.
Severity: High
This is not a technical limitation—the model has the capability to open links. This is a behavioral defect where the model skips a critical verification step and defaults to fabrication. This directly undermines user trust in a core user-initiated function, and is especially dangerous when links contain information on serious topics such as legal, medical, or safety-related content.
Screenshots
Additional context
This issue has been observed on the latest production model (V4-Pro). It is distinct from but related to Issue #1323—both stem from the same root cause: the model defaults to generating plausible-sounding answers rather than pausing to execute necessary verification steps or honestly acknowledging limitations. The user has pointed out this specific pattern multiple times within a single conversation, yet the model continued to repeat the same error. Cross-version testing may help determine whether this defect is isolated to V4 or present in earlier releases.
Suggested fix
Investigate why the model skips the link-opening step when it has the capability to do so. Implement a mandatory verification step that requires the model to confirm successful content retrieval before generating any summary. If the link cannot be opened for any reason, the model must default to an explicit error message rather than inferring or fabricating content.
@qingkong66 commented on GitHub (May 18, 2026):
@D-Orancle
Confirmed pattern. The model skips the link-opening step and fabricates summaries instead. You've documented it clearly with screenshots.
Related to #1323 — same root: model defaults to plausible fabrication over verification.
Not a fix, but your issue is on record.
@darkemil2000-arch commented on GitHub (May 19, 2026):
Поддержите идею #1328
@github-actions[bot] commented on GitHub (Jun 23, 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 (Jul 8, 2026):
false