ÃÊ·Ï ¿­±â/´Ý±â ¹öư

»ý¼ºÇü ÀΰøÁö´ÉÀ» Ȱ¿ëÇÑ ±ÝÀ¶»óǰ Ãßõ ¼­ºñ½ºÀÇ µµÀÔÀÌ º»°ÝÈ­µÇ°í ÀÖÀ¸³ª, ±×·ÎºÎÅÍ ¹ß»ýÇÏ´Â ¹ýÀû Ã¥ÀÓÀÇ ±¸Á¶¿¡ °üÇÑ Çй®Àû ³íÀÇ´Â ÃæºÐÈ÷ ÃàÀûµÇÁö ¾Ê¾Ò´Ù. º» ³í¹®Àº Çѱ¹ ±ÝÀ¶¼ÒºñÀÚº¸È£¹ý ü°è ÇÏ¿¡¼­ »ý¼ºÇü AIÀÇ ±â¼úÀû Ư¼ºÀÎ È®·üÀû Ãâ·Â, ¸Æ¶ô ÀÇÁ¸¼º, ȯ°¢ÀÌ ÀûÇÕ¼º ¿øÄ¢(Á¦17Á¶), ¼³¸íÀǹ«(Á¦19Á¶), ºÎ´ç±ÇÀ¯ ±ÝÁö(Á¦21Á¶)ÀÇ Àû¿ë¿¡ ¾î¶°ÇÑ ±¸Á¶Àû °£±ØÀ» ¾ß±âÇÏ´ÂÁö¸¦ ºÐ¼®ÇÑ´Ù. º» ³í¹®ÀÇ ÇÙ½É ³íÁö´Â Àǹ« ÀÌÇàÀÇ Æò°¡ ´ÜÀ§¸¦ Ãâ·Â ´ÜÀ§¿¡¼­ ½Ã½ºÅÛ ´ÜÀ§·Î À̵¿½Ãų Çʿ䰡 ÀÖ´Ù´Â Á¡ÀÌ´Ù. ƯÈ÷ ¼³¸íÀǹ«¿¡ °üÇÏ¿©´Â, Ãâ·Â Â÷¿øÀÇ ¼³¸í°¡´É¼ºÀÌ ±¸Á¶ÀûÀ¸·Î Á¦ÇѵǴ ¿µ¿ª¿¡¼­ °Å¹ö³Í½º Â÷¿øÀÇ ¼³¸íÀÌ ÀÚ±â°áÁ¤±Ç º¸ÀåÀ̶ó´Â ÀÔ¹ý ÃëÁö¸¦ ´Þ¼ºÇϱâ À§ÇÑ ±â´ÉÀû µî°¡¹°·Î ÀÛµ¿ÇÒ ¼ö ÀÖÀ½À» Á¦½ÃÇÑ´Ù. ÀÌ·¯ÇÑ Á¢±ÙÀº EU ÀΰøÁö´É¹ý, ¹Ì±¹ SEC ±ÔÄ¢¾È, ½Ì°¡Æ÷¸£ FEAT ¿øÄ¢ µî ÇØ¿Ü ±ÔÁ¦ µ¿Çâ°ú Á¤ÇÕÀûÀÌ´Ù. º» ³í¹®Àº ÀÌ·¯ÇÑ ºÐ¼® À§¿¡¼­ 󰡐ÇнÀ °¡´É¼º(trainability)󰡑À» ÁÖÀÇÀǹ«ÀÇ ÇÙ½É ±¸¼º¿ä¼Ò·Î Á¦½ÃÇϸç, À̸¦ ¸Å°³·Î »ç¾÷ÀÚÀÇ °Å¹ö³Í½º ¼º¼÷µµ¿¡ µû¶ó Ã¥ÀÓÀ» Â÷µîÈ­ÇÏ´Â ´Ü°èÀû ÁÖÀÇÀǹ« ±âÁØÀ» °³¼±¹æ¾ÈÀ¸·Î Á¦¾ÈÇÑ´Ù. ÀÌ´Â ¡¸±ÝÀ¶È¸»çÀÇ Áö¹è±¸Á¶¿¡ °üÇÑ ¹ý·ü¡¹»ó ³»ºÎÅëÁ¦±âÁØÀÇ ¿î¿µ ¿ø¸®¿Í µ¿ÀÏÇÑ ¹ß»ó¿¡ ±âÃÊÇϸç, ÀýÂ÷Àû ÁÖÀÇÀǹ« ¹ý¸®ÀÇ ÀÚ¿¬½º·¯¿î È®Àå¿¡ ÇØ´çÇÑ´Ù.

Generative AI-based financial product recommendation services are entering the market, yet scholarly discussion on the structure of the legal liability arising from such services remains underdeveloped. This paper analyzes how the technical characteristics of generative AI — probabilistic outputs, context sensitivity, and hallucination — create structural gaps in the application of the principal duties under Korea¡¯s Financial Consumer Protection Act: the suitability principle (Article 17), the duty to explain (Article 19), and the prohibition of unfair solicitation (Article 21). The central thesis is that the unit of evaluation must shift from the output level to the system level. With particular reference to the duty to explain, the paper argues that governance-level explanation — disclosure of the system¡¯s training and evaluation profile, known limitations, and verification pathways — operates as a functional equivalent for safeguarding the consumer¡¯s right to self-determination where output-level explainability is structurally constrained. This approach is consistent with comparative developments in the EU, the United States, and Singapore. Building on this analysis, the paper advances trainability — the capacity of the model itself to be updated through operational experience — as a core component of the duty of care, and proposes a graduated standard of care that differentiates liability according to operators¡¯ governance maturity.



Ű¿öµå¿­±â/´Ý±â ¹öư

, , , , ,

Generative AI, Financial Consumer Protection Act, Duty to Explain, Trainability, Graduated Standard of Care, Governance-Level Explanation