
AIã«èãåã«èªåã§èããæå³ã~åŠç¿ç ç©¶ã§ç¢ºããã~
æè¿ãHadesïŒ@0xHvdesïŒããã®ãThe Principles of Learning Faster.ããšããè±èªã®èšäºãèªã¿ãŸãããåŠç¿ãéããã7ã€ã®ååããçãäŸãšãã£ããã«äžŠã¹ãé·æã§ãã
7ã€ã®ãã¡ãç§ãç«ã¡æ¢ãŸã£ãã®ã¯3ã€ç®ã®ãStruggle Before You SearchãïŒæ€çŽ¢ããåã«æ©ãïŒã§ããæ°ç§ã§çããåºãæä»£ã«ãå ã«èªåã§èããæå³ã¯ã©ãã«ããã®ã§ããããã
ãã®èšäºã§ã¯ããã®ååã ããåãäžããŸããåæã«ã¯ç ç©¶ã®åŒçšãããŒã¿ããããŸããã
ããã§ãè£ä»ãã«ãªããããªå¿çåŠãšæè²ã®ç ç©¶ãäžæ¬¡è³æã§ç¢ºãããAIã«ããèãå Žåã®åŸãšæãäž¡æ¹èŠãŸãã
ãçºç«¯ã 7ååã®ãã¡ãAIãšçŽæ¥éãªã3çªç®

åå3ã¯ãçããåŸãéããšã身ã«ã€ãéãããå¥ã®ãã®ãšããŠæ±ã£ãŠããŸãã
ã»ãã®6ã€ã¯ãLearn by DoingãShorten the Feedback LoopãLearn Just in TimeãRetrieval Beats RecognitionãConnect Before You CollectãLearning Compoundsã§ãããã®èšäºã§ã¯ããããžã¯åºããŸããã
åæã®äŸã¯ãããã°ã©ãã³ã°ã®åŠç¿äžã«ãšã©ãŒãåºãå Žé¢ã§ããããAIã«è²Œã£ãŠè§£æ±ºãã人ãšã10åã»ã©èªåã§è©ŠããŠããèãäººãæ¯ã¹ãŠããŸãã
åŸè ã¯ããšã©ãŒãèªã¿ã仮説ãç«ãŠãäœããå€ããå¥ã®å Žæãå£ãããŸã詊ããŸãããã®ããšAIã«èããŠããæå ã«èªåã®æ¡ãããã®ã§ãèãã®ã©ããéã£ãããåããããšããã®ãåæã®äž»åŒµã§ãã
åæã¯ãéå ·ã䜿ããªãšã¯æžããŠããŸãããAIãèããè©ä»£ããããåã«ãè³ã«åé¡ãæ±ãæ©äŒãäžããããšããäœçœ®ã¥ãã§ãã
ãã®äž»åŒµã¯ãèè ã®çµéšåãšããŠæžãããŠããŸãã以äžã§ã¯ãç ç©¶ã§ç¢ºãããããéšåãšã確ãããããªãéšåãåããŸãã
ãåºç€ã çããèŠãåã®è©Šè¡ããããšã®åŠç¿ã匷ãã

ééãããããããªã詊è¡ã§ãããã®ããšã«çããèŠããšãèªãã ãããèšæ¶ã«æ®ãããšããå®éšçµæããããŸãã
é¢ä¿ãã广ã¯3ã€ãããŸããçæå¹æã¯ãçããèªåã§äœããšãèªãã ãããèšæ¶ã«æ®ãçŸè±¡ã§ãã
SlameckaãšGrafã¯ã1978幎ã®è«æã§ãã®ååãä»ããŸããã倧åŠç96人ã察象ã«ãã5ã€ã®å®éšã§ã¯ãèªåã§äœã£ãåèªããåãåèªãèªãã ããããèšæ¶ã«æ®ããŸããïŒSlameckaãšGrafïŒ1978ïŒïŒã
ãã¹ã广ã¯ãæãåºãç·Žç¿ãã®ãã®ãèšæ¶ã匷ããçŸè±¡ã§ããRoedigerãšKarpickeïŒ2006ïŒã§ã¯ãåŠçãæç« ãèªãã ããšãåèªãã矀ãšãæãåºããŠæžã矀ã«åãããŸããã
5ååŸã®è©Šéšã§ã¯åèªãäžåããŸãããã2æ¥åŸãš1é±éåŸã¯ãæãåºãç·Žç¿ããã矀ã倧ããäžåããŸãããåèªãã矀ã®ã»ãããèŠããããããšããèªä¿¡ã¯é«ãã£ãããã§ãïŒPubMedïŒã
ãã¬ãã¹ã广ã¯ãçããåãããªãåãã«å ã«æããšããã®ããšã®åŠç¿ã匷ãŸãçŸè±¡ã§ããKornellãã¯ãæ£è§£ã§ããªãåããçšæããå ã«çããæšæž¬ãããŠããçããèŠãã矀ãšãåããšçããåæã«èŠããçŸ€ãæ¯ã¹ãŸããã
å ã«æšæž¬ãã矀ã®ã»ãããããšã®æçžŸãé«ããªããŸããïŒPubMedïŒã
åå3ã¯ããã®3ã€ç®ã®å¹æãšæãè¿ãäœçœ®ã«ãããŸãã
ãéãã å ã«èª¿ã¹ãåŠç¿ãšãå ã«è©ŠãåŠç¿ã®åããç®

éãã¯ãçããèŠãç¬éã«ãæ¯ã¹ãçžæãæå ã«ãããã§ãã
æ¬¡ã®æŽçã¯ãåæã®äŸãšäžã®ç ç©¶ãããç§ã軞ãç«ãŠããã®ã§ããåæã«ãç ç©¶ã«ãããã®åœ¢ã¯ãããŸããã
æåã«èµ·ããããš
å ã«æ€çŽ¢ã»AIã«èãïŒçããèªã
å ã«èªåã§è©ŠãïŒä»®èª¬ãç«ãŠã
çããèŠããšã
å ã«æ€çŽ¢ã»AIã«èãïŒå 容ãåãåã
å ã«èªåã§è©ŠãïŒèªåã®æ¡ãšæ¯ã¹ã
ãããæé
å ã«æ€çŽ¢ã»AIã«èãïŒçã
å ã«èªåã§è©ŠãïŒè©Šè¡ã®åã ãé·ã
åãå Žé¢
å ã«æ€çŽ¢ã»AIã«èãïŒæããããã»ãŒç¡ãååŠ
å ã«èªåã§è©ŠãïŒæããããå°ãã§ãããæ®µé
æåŸã®è¡ã¯ãåŸè¿°ããã¯ãŒã¯ããšã°ã¶ã³ãã«ïŒè§£ãæ¹ãå šéšèŠããäŸé¡ïŒã®ç ç©¶ããåŒãããç§ã®èªã¿æ¹ã§ãã
åŸæ¥ã®ãåãããªããã°èª¿ã¹ããã¯ãæéã®é¢ã§ã¯åççã§ããåå3ãèŠãŠããã®ã¯ãæéã§ã¯æž¬ããªããçããåãåã£ãããšã®å®çã®ã»ãã§ãã
ãä»çµã¿ã èªåã®æ¡ãšçãããæ¯ã¹ããã圢ãž
å ã«è©Šãå©ç¹ã¯ããæ©ãã ããšããã®ãã®ããã詊è¡ãšçãã察ã«ãªãããšã«ãããŸãã
Kornellãã®çµæã¯ãææŠãã詊è¡ã®ããšã«çããæ¥ãå Žåã§ããå¥ã®å®éšã§ã¯ã倱æãã詊è¡ã®ããšã«çãããã瀺ããããšãèšæ¶ã匷ãŸããŸããã
çããé ããŠç€ºããããšãåãå©ç¹ã¯åºãŸããã§ããïŒHaysãã®è«æïŒã詊è¡ãå¹ãã®ã¯ããã®ããšã«çããæ¥ãŠãæ¯ã¹ããããšãã§ãã
AIã«ããèãåŽã®çµæããããŸããBastaniãã¯ããã«ã³ã®é«æ ¡çããã1,000äººã«æ°åŠãç·Žç¿ãããŸããã
GPT-4ãèªç±ã«äœ¿ãã矀ã¯ãç·Žç¿äžã®æçžŸã48%äžãããŸãããå®å šçãä»ããGPT Tutorçã䜿ã£ã矀ã¯ã127%äžãããŸããã
ãšããããAIãåãäžããŠè©Šéšããããšãèªç±ã«äœ¿ãã矀ã¯ãAIã䜿ããªãã£ã矀ãã17%äœããªããŸãããå®å šçä»ãã®çã§ã¯ããã®äœäžã¯ã»ãŒè§£æ¶ãããŸããïŒPNASïŒã
æ°åã¯ãPNASã®èŠæšã«ãããŸããWhartonã®èšäºã¯ãåãç ç©¶ã®äºæ¬¡çãªè§£èª¬ã§ãïŒWhartonïŒã
è«æã¯ãå®å šçã®ç¡ã矀ãç·Žç¿äžã«GPT-4ããæãã®ããã«äœ¿ããAIãå€ãã詊éšã§æçžŸãäžãã£ããšå ±åããŠããŸããçµæã¯1ã€ã®ç ç©¶ã§ã察象ã¯é«æ ¡æ°åŠã§ãã
2025幎8æã®èšæ£ã¯ãOsbert Bastaniã®æå±ã®èª€ããçŽããã ãã§ããæ°åãšçµè«ã¯å€ãã£ãŠããŸããïŒèšæ£ïŒã
åŸã®åŽã¯ãç·Žç¿äžã®æçžŸã倧ããäžããããšãšãçããŸã§ã®éãã§ããæã®åŽã¯ãAIãå€ãããšãã®æçžŸãäžããããšã§ãã
ãå ·äœäŸã ãšã©ãŒã貌ãåã®10å

åæã®äŸãã詊è¡ãšæ¯èŒã®é ã«äžŠã¹çŽããšãä»çµã¿ãèŠããŸãã
ããã°ã©ãã³ã°åŠç¿äžã«ãèŠæ £ããªããšã©ãŒãåºããšããŸããåæãå§ããã®ã¯ã10åã»ã©èªåã§è©Šãããšã§ãã
ãšã©ãŒã¡ãã»ãŒãžãèªã¿ãåå ã®ä»®èª¬ãäžã€ç«ãŠãŸããã³ãŒããäžãæå€ããŠåãããå¥ã®å Žæãå£ãããããããæãããã«ããŸãã
ãã®ããã§ãAIã«èããŸãããã®ãšããAIã®çããšãèªåãç«ãŠã仮説ã䞊ã¹ãããŸãã
仮説ãåœãã£ãŠãããªãã確èªãåããŸããå€ããŠãããªããã©ãã§èªã¿éããããåãããŸãã
ãã®æµããäžã®ç ç©¶ã§èšãæãããšã仮説ãç«ãŠã段éãçæãçããèŠã段éã峿ã®ãã£ãŒãããã¯ã§ãã
ã10åããšããé·ãã«ã¯ãä»å確èªããç ç©¶ã®è£ä»ãããããŸãããåæã®ç®å®ãšããŠæ±ã£ãŠãã ããã
ãé¢é£æŠå¿µã çç£ç倱æãšã¯ãŒã¯ããšã°ã¶ã³ãã«ã®äœ¿ãåã

å ã«æ©ãåŠç¿ã¯ãããããå Žé¢ã§æè¯ãšã¯èšãããæ¡ä»¶ã«ãã£ãŠéã®çµæãåºãŠããŸãã
KapurïŒ2008ïŒã¯ã髿 ¡2幎çã®ç©çïŒãã¥ãŒãã³ååŠïŒã®èª²é¡ã§ãæããããå°ãªãåé¡ãå ã«è§£ãããŸãããã°ã«ãŒãã§ã¯è§£ããŸããã§ãããããã®ããšã®åå¥ã®èª²é¡ã§ã¯ãæŽã£ãåé¡ããå§ãã矀ãäžåããŸããã
ãã®çŸè±¡ããçç£ç倱æãšåŒã³ãŸãïŒè«æèŠæšïŒã
äžæ¹ãã¯ãŒã¯ããšã°ã¶ã³ãã«å¹æã¯ãè§£ãæ¹ãå šéšç€ºããäŸé¡ãåŠãã ååŠè ããèªåã§è§£ããåŠç¿è ããããšã®æçžŸãé«ããšãããç¹°ãè¿ã確èªãããçµæã§ããååŠè ãæãããç¡ãã«è§£ãããšãããšãäœæ¥èšæ¶ïŒèããäžæçã«ä¿æãã容éïŒã䜿ãããè§£æ³ã®èŠåãé·æèšæ¶ã«å ¥ãã«ããããšèª¬æãããŠããŸãïŒNSWæè²çCESEã®è§£èª¬ïŒã
ãã®äºã€ã®çµæã¯ãäžèŠãããšé£ãéããŸããéããåããæ¡ä»¶ããä»åã®è³æã§ã¯ç¢ºãããããŠããŸããã
ç§ã®ä»®èª¬ã§ãããåŠç¿è ã®äºåç¥èã®éãšã詊è¡ã®ããšã«æå°ãçããæ¥ããã©ãããé¢ãã£ãŠããããã«èŠããŸãã
åèæ å ±ã»åèæžç±ïŒPRïŒ
æ¬èšäºãããæ·±ãçè§£ããããã«ãæŠå¿µãããã¹ã©ã€ããäœæããŸããããã¡ããã¿ãããšã§çè§£ã®è£å©ã«ãªãã¯ãã§ãã
ããŒã¿ãŒã»ãã©ãŠã³ããã³ãªãŒã»ããŒãã£ã¬ãŒãããŒã¯ã»ãã¯ãããšã«èãäŸç°å巳蚳ã䜿ããè³ã®éãæ¹ æåããåŠç¿ã®ç§åŠãã玹ä»ããŸããåŠç¿ã®ç§åŠãšããé¡ããæ³èµ·ãçæã詊è¡ãæ±ããã®èšäºãšéãªãããã§ãã
é¢é£èšäº
ããŸãšãã èãåã«äžè¡æžããå°ããªå®éšãž

ãAIã䜿ãã䜿ããªãããããå ã«ããèãåã«èªåã®æ¡ããããããèŠçŽãããšããæ¬¡ã®äžæ©ã§ãã
次ã®AIãžã®è³ªåã§ã次ã®3ã€ã詊ããŸãã广ã¯ç§ã確ããããã®ã§ã¯ãªããäžã®ç ç©¶ããçµãã è©Šãæ¹ã§ãã
⡠質åãéãåã«ãèªåã®çãã仮説ãäžè¡ã ãæžã
â¡ çããåãåã£ãããèªåã®æ¡ãšéã£ãç¹ãäžã€èšèã«ãã
â¡ æ°æ¥åŸã«ãAIãèŠãã«åãå 容ãèªåã®èšèã§èª¬æããŠã¿ã
æåã®æãããããŸã£ããç¡ã話é¡ã§ã¯ãå ã«è§£èª¬ãäŸé¡ãèªãã»ããåãå ŽåããããŸãããã®å Žåããèªã¿çµããããšã«ãéããŠæãåºãç·Žç¿ãè¶³ããŸãã
åèæ å ±
The generation effect: Delineation of a phenomenonïŒSlamecka and Graf, 1978ïŒ
Unsuccessful retrieval attempts enhance subsequent learningïŒKornell, Hays, Bjork, 2009ïŒ
How Retrieval Attempts Affect Learning: A Review and Synthesis
Generative AI without guardrails can harm learningïŒPNAS, 2025ïŒ
Without Guardrails, Generative AI Can Harm EducationïŒWhartonïŒ
Cognitive load theoryïŒNSW Department of Education, CESEïŒ