The Logic of Large Language Models'Language Security Risks From the Perspective of a Holistic Approach to National Security
Language is the thread that weaves together the political and cultural community of a na-tion-state,forming the cornerstone of a nation's cultural institution.In the theoretical framework of a holistic approach to national security,cultural security functions as a guarantee element of holistic national security,with language security being subordinate to the security of representa-tional cultural symbols within cultural security.It also serves as an important indicator for assess-ing the state of cultural security.In the age of artificial intelligence,through the supervision of training data and application of fine-tuning data by architects of Large Language Models(LLMs),models will be able to carry specific collective values and ideological attributes,infil-trating through massive interactions with society,which will generate two kinds of risk.The di-rect risk is that the dominance of natural language in the LLMs triggers disputes over cultural sovereignty.Major economies around the world are committed to developing localized LLMs,in-tensifying"language securitization".The indirect risk is that the LLMs with ideological attri-butes will subtly erode the mainstream social values and core value ideology,thus weakening the national identification with the lingua franca and causing"language insecurity".The risks of lan-guage security are characterized by long-term,hidden accumulation,and once they become a threat to cultural security,they are difficult to mitigate.Therefore,we need to adopt a bottom-line mindset,and actively promote the development and application of the local LLMs.Then we should strengthen the top-level design of AI,improve the laws and regulations in the field of AI,remain vigilant against the risk of value bias,and enhance our ability to value alignment,in order to effectively safeguard the national language security and cultural sovereignty.
a holistic approach to national securitylanguage securitycultural securityartifi-cial intelligence securityLarge Language Models(LLMs)
NETL
NSTL
万方数据
