| 曾庆鸣,毛维准.认知武器化与人工智能认知战[J].国际安全研究,2024,(5):49-80 |
| 认知武器化与人工智能认知战 |
| Cognitive Weaponization and Cognitive Warfare of Artificial Intelligence: A Study Based on Machine Learning and Behavioral Experiment |
| 修订日期：2024-07-13 |
| DOI：10.14093/j.cnki.cn10-1132/d.2024.05.003 |
| 中文关键词: 人工智能 ChatGPT 认知战 算法 机器学习 行为实验 |
| 英文关键词: artificial intelligence, ChatGPT, cognitive warfare, algorithms, mobilization, machine learning, behavioral experiment |
| 基金项目: |
|
| 摘要点击次数: 515 |
| 全文下载次数: 218 |
| 中文摘要: |
| 作者基于自然语言处理与机器学习，针对1950—2024年173例美国对外政策与行为，构造以ChatGPT为对象的行为实验与对照研究，涵盖美国制裁禁运、安理会否决、军事干涉、政权更迭、海外作战等五类事件。虽然目前大语言模型未在明面上被用于谋杀、军事破坏或系统性认知攻防与渗透，也没有大语言模型认知攻防作战公开明确的战略政策执行或文本宣示，但OpenAI的ChatGPT模型展现出显著的认知安全风险。这种认知威胁具有长期性、隐蔽性、精准性等特点，其潜在认知杀伤与认知“殖民”能力不容忽视。ChatGPT能够选择性呈现信息与伪造事实，并塑造、扭转用户认知与思维模式，引导用户产生对美国对外政策与行为正当性的积极情感，进而实现认知操纵。ChatGPT生成的内容在文本、叙事与情感等三个层面表现出正负面极化情感动员趋势，通过“言过其实”“无罪推定”“以退为进”等方式放大己方正当性、美化自身形象，通过“栽赃陷害”“无中生有”“有罪推定”污名化、妖魔化竞争对手形象。基于大语言模型的人工智能认知战成为一种隐蔽、精准、低成本、高收益的非动能作战工具。人工智能认知战利用人类的认知弱点和思维偏差，潜移默化地实施思维诱导与认知操纵，实现了“认知武器化”，对认知域国家安全构成了重大威胁。 |
| 英文摘要: |
| Based on natural language processing and machine learning, this research program conducts a behavioral experiment on ChatGPT that focuses on 173 instances of US foreign policies and actions from 1950 to 2024. Five categories of events have been included in the research, namely US sanctions and embargoes, UN Security Council vetoes, military interventions, regime changes, and overseas operations. The results of the experiment demonstrate that, although large language models are not currently being overtly used for killing, military sabotage, or systemic cognitive warfare and infiltration, OpenAI’s ChatGPT models have exhibited significant cognitive security risks. This cognitive threat is characterized by its long-term, covert and precise nature, and therefore its potential cognitive harm and cognitive “colonization” capabilities should not be overlooked. ChatGPT possesses the power to be a “cognitive bomb” that is capable of achieving cognitive manipulation by emotionally mobilizing users, subtly and selectively presenting information, fabricating facts, shaping and reversing users’ cognition and thinking patterns, and luring them to develop positive emotions towards the legitimacy of US foreign policies and actions. The contents generated by ChatGPT show a trend of emotional polarization at the textual, narrative, and emotional levels, amplifying the legitimacy and beautifying the image of their own side by such means as “exaggeration”, “presumption of innocence” and “strategic retreat” while stigmatizing and demonizing their competitors’ image through “false accusations”, “fabrication” and “presumption of guilt”. AI cognitive warfare based on large language models has become a covert, precise, low-cost and high-yield non-kinetic combat tool. It is evident that AI cognitive warfare can significantly exploit human cognitive weaknesses and thinking biases, and subtly implement thought induction and cognitive manipulation so as to achieve “cognitive weaponization”, thus posing grave national security risks and threats to cognitive domain security. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|