The Price of Intelligence.
Large language models (LLMs) exhibit three inherent risks—hallucination, indirect prompt injection, and jailbreaks—that stem from their probabilistic foundations and linguistic flexibility. Because training and generation both rely on stochastic processes, outputs can be unpredictable, occasionally...
| Publicado en: | Communications of the ACM Vol. 68; no. 9; pp. 46 - 54 |
|---|---|
| Autores principales: | , , , |
| Formato: | Artículo |
| Publicado: |
Association for Computing Machinery
Sep2025
|
| Materias: | |
| Acceso en línea: | Ver este registro en EBSCOhost |
| fields | @attributes: recordID: 1 pdfLink: plink: https://search.ebscohost.com/login.aspx?direct=true&db=hlh&AN=187621004&site=ehost-live header: @attributes: shortDbName: hlh uiTerm: 187621004 longDbName: Humanities International Complete uiTag: AN controlInfo: bkinfo: jinfo: jid: 00010782 ACM jtl: Communications of the ACM issn: 00010782 maglogo: N pubinfo: dt: Sep2025 vid: 68 iid: 9 pid: 68 pub: Association for Computing Machinery artinfo: ui: 187621004 10.1145/3749447 ppf: 46 ppct: 8 formats: tig: atl: The Price of Intelligence. aug: au: Russinovich, Mark Salem, Ahmed Zanella-Béguelin, Santiago Zunger, Yonatan affil: Microsoft Azure, Bellevue, WA, USA Microsoft Security Response Center, Redmond, WA, USA Microsoft Research, Cambridge, United Kingdom Microsoft, Mountain View, CA, USA su: Language models Hallucinations (Artificial intelligence) Computer network security Probabilistic generative models Stochastic processes Artificial intelligence in business Misinformation sug: subj: Language models Hallucinations (Artificial intelligence) Computer network security Probabilistic generative models Stochastic processes Artificial intelligence in business Misinformation ab: Large language models (LLMs) exhibit three inherent risks—hallucination, indirect prompt injection, and jailbreaks—that stem from their probabilistic foundations and linguistic flexibility. Because training and generation both rely on stochastic processes, outputs can be unpredictable, occasionally incorrect, or vulnerable to manipulation. These challenges pose particular concerns for high-stakes applications in areas like healthcare, law, and finance, where reliability and safety are critical. While such risks cannot be eliminated, layered mitigation strategies—spanning alignment methods, system-level safeguards, and human oversight—offer pathways toward responsible and trustworthy deployment. pubtype: Periodical doctype: Article src: R language: English refInfo: copyright: @attributes: flag: Y dt: @attributes: year: 2025 holdings: @attributes: islocal: N |
|---|