DeepSeek回应用户“对话泄露”疑虑：特殊字符引发模型幻觉_推荐_大众网. 

Detected Manipulations & Suggested Changes

Unsubstantiated claims

Presenting a claim as fact without providing evidence, data, or independent verification.

“输入‘##_FORMAT_LT_## think ##_FORMAT_GT_##’等字符触发返回异常内容，属于特殊字符引发的模型幻觉，不涉及安全问题或隐私泄露。” The article relays DeepSeek’s conclusion that the issue is purely a hallucination caused by special characters and that it does not involve security or privacy leaks. No technical details, methodology of the ‘全面排查’, or third‑party confirmation are provided to substantiate this assurance.

3 suggestions

Add information on how the conclusion was reached, e.g.: “DeepSeek表示，其技术团队通过日志审计、代码审查和模拟复现等方式进行了全面排查，未发现任何对话内容被外泄或被第三方访问的证据。”

Include any available independent or external assessment, e.g.: “目前尚无第三方安全机构报告发现相关隐私泄露问题。”

Qualify the certainty of the claim instead of stating it as absolute fact, e.g.: “目前排查结果显示，该问题应为特殊字符引发的模型幻觉，暂未发现涉及安全问题或隐私泄露的证据。”

Omission of key information

Leaving out relevant context or details that would help readers fully assess the situation.

The article mentions that users had ‘对话泄露’疑虑 and that DeepSeek conducted a ‘全面排查’, but it does not explain: - What specific ‘不可预期的内容’ looked like (e.g., did it resemble other users’ conversations?). - What ‘全面排查’ entailed in technical terms. - Whether any limitations or remaining uncertainties exist. This omission makes it harder for readers to independently judge how convincing the company’s explanation is.

3 suggestions

Briefly describe the nature of the abnormal outputs, e.g.: “部分用户反馈，模型在输入上述字符后，会生成与当前对话无关、看似来自其他场景的文本片段。”

Outline the scope of the investigation, e.g.: “DeepSeek称，其排查包括模型训练数据来源、推理日志、接口调用记录等多个环节。”

Mention any remaining uncertainties or monitoring plans, e.g.: “DeepSeek表示，将继续监控相关触发模式，如发现新的风险线索将及时更新说明。”

Unbalanced reporting

Presenting mainly one side’s perspective while giving minimal or no space to other relevant viewpoints.

The article primarily reproduces DeepSeek’s official statement. Users’ concerns are only summarized in one sentence: “使部分用户产生了‘对话泄露’的疑虑。” No direct user quotes, examples of their experiences, or external expert opinions are included. This makes the piece function more as a company clarification than a balanced report on the controversy.

3 suggestions

Include at least one or two anonymized user descriptions of what they observed, e.g.: “有用户表示，模型生成的内容与其本人从未输入过的对话高度相似，因此担心存在泄露风险。”

Add an independent expert or security researcher comment, even if brief, e.g.: “一位安全领域专家指出，从目前公开信息看，更可能是模型在训练数据中学习到的通用文本模式，但仍需更多技术细节来完全排除泄露可能。”

Clarify that the article is mainly relaying the company’s position, e.g.: “以下为DeepSeek方面的说明：” so readers can better contextualize the one‑sided nature of the information.

Appeal to authority

Relying on the authority or expertise of a party as the main basis for accepting a claim, without providing supporting evidence.

“针对这一问题，我们的技术团队经全面排查后发现：……不涉及安全问题或隐私泄露。” The conclusion is justified primarily by invoking ‘技术团队’ and ‘全面排查’ rather than by presenting verifiable evidence or allowing external scrutiny. Readers are implicitly asked to trust the company’s internal authority.

3 suggestions

Supplement the appeal to internal expertise with concrete, checkable information, e.g.: “DeepSeek公布了部分技术细节，说明异常内容来自模型训练语料中的公开文本片段，而非实时用户对话。”

Clarify the limits of internal authority, e.g.: “DeepSeek方面表示，基于目前内部排查结果，尚未发现安全或隐私问题。”

If possible, reference external standards or audits, e.g.: “公司称，其数据安全流程已通过某某认证，相关机制在此次排查中也得到验证。”