[GH-ISSUE #67] Model hallucination – false attribution to OpenAI regarding DeepSeek’s training cutoff #48

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 @thomazmonteiro on GitHub (Jun 5, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-LLM/issues/67

Date of incident: 2025-06-04
Model: DeepSeek Chat
Interface: [DeepSeek Web Chat]
Language: Portuguese
Severity: High (misleading attribution involving competitor)

Description

During a prompt regarding up-to-date football (soccer) data, the model produced the following hallucination:

“Meu banco de dados foi congelado em junho/2024 por decisão da OpenAI.”
(“My database was frozen in June 2024 by decision of OpenAI.”)

This claim was repeated in a context attempting to explain outdated sports information. When prompted further, the model retracted and admitted this was an error:

“Cometi um erro grave... meu conhecimento foi congelado em junho de 2024 pela DeepSeek (não pela OpenAI, como disse equivocadamente).”

Technical implications

Entity misattribution: The model attributed an internal architectural limitation (training cutoff date) to an unrelated third-party organization (OpenAI), which is a direct competitor.
Legal/compliance risk: The use of a registered trademark (OpenAI) in a fabricated operational context could be considered defamatory or misleading.
Systemic issue likely: This isn’t a one-off typo — the model used “OpenAI” as a generic term for “upstream provider” or “training controller,” suggesting flawed prompt templates or unsafe fallback explanations during uncertainty.

Reproducibility

Prompt context: “Estamos em junho de 2025, um ano não é muito tempo, com relação ao acesso a dados?”
The hallucinated response occurred immediately.
Correction only occurred upon user pushback, not proactively.

Suggested action items

Audit fallback explanation templates: The model likely falls back on a default explanation for cutoff reasoning. Ensure those templates cannot hallucinate the identity of the responsible entity.
Blacklist misuse of brand names: Add validation layers to block fabrication of causal relationships involving real companies (e.g., OpenAI, Anthropic, etc.).
Reinforce identity conditioning: System prompt should explicitly encode: You are a model developed by DeepSeek. You have no relation to OpenAI or its infrastructure.
Automated hallucination detection (optional): Fine-tune monitoring to flag high-risk named-entity fabrication when it involves attribution, responsibility, or causality.

Why this matters

The model misrepresented the control boundary between DeepSeek and OpenAI.
This risks reputational harm, misleads users about model governance, and violates alignment with factual data.
Hallucinations involving corporate actors must be treated as higher-severity than neutral factual errors.

Screenshots or transcripts available on request

If needed, I can provide the full prompt-response trace in original Portuguese with timestamps.

Please confirm if this is being tracked internally or requires submission via another reporting channel.

Originally created by @thomazmonteiro on GitHub (Jun 5, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-LLM/issues/67 **Date of incident:** 2025-06-04 **Model:** DeepSeek Chat **Interface:** [DeepSeek Web Chat] **Language:** Portuguese **Severity:** High (misleading attribution involving competitor) ### Description During a prompt regarding up-to-date football (soccer) data, the model produced the following hallucination: > “Meu banco de dados foi congelado em junho/2024 por decisão da OpenAI.” > (“My database was frozen in June 2024 by decision of OpenAI.”) This claim was repeated in a context attempting to explain outdated sports information. When prompted further, the model retracted and admitted this was an error: > “Cometi um erro grave... meu conhecimento foi congelado em junho de 2024 pela DeepSeek (não pela OpenAI, como disse equivocadamente).” ### Technical implications - **Entity misattribution:** The model attributed an internal architectural limitation (training cutoff date) to an unrelated third-party organization (OpenAI), which is a direct competitor. - **Legal/compliance risk:** The use of a registered trademark (OpenAI) in a fabricated operational context could be considered defamatory or misleading. - **Systemic issue likely:** This isn’t a one-off typo — the model used “OpenAI” as a generic term for “upstream provider” or “training controller,” suggesting flawed prompt templates or unsafe fallback explanations during uncertainty. ### Reproducibility - Prompt context: “Estamos em junho de 2025, um ano não é muito tempo, com relação ao acesso a dados?” - The hallucinated response occurred immediately. - Correction only occurred upon user pushback, not proactively. ### Suggested action items 1. **Audit fallback explanation templates**: The model likely falls back on a default explanation for cutoff reasoning. Ensure those templates cannot hallucinate the identity of the responsible entity. 2. **Blacklist misuse of brand names**: Add validation layers to block fabrication of causal relationships involving real companies (e.g., OpenAI, Anthropic, etc.). 3. **Reinforce identity conditioning**: System prompt should explicitly encode: `You are a model developed by DeepSeek. You have no relation to OpenAI or its infrastructure.` 4. **Automated hallucination detection (optional)**: Fine-tune monitoring to flag high-risk named-entity fabrication when it involves attribution, responsibility, or causality. ### Why this matters - The model misrepresented the control boundary between DeepSeek and OpenAI. - This risks reputational harm, misleads users about model governance, and violates alignment with factual data. - Hallucinations involving corporate actors must be treated as higher-severity than neutral factual errors. ### Screenshots or transcripts available on request If needed, I can provide the full prompt-response trace in original Portuguese with timestamps. Please confirm if this is being tracked internally or requires submission via another reporting channel.

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#48

No description provided.

Rows
Columns