Understanding large language models demands distinguishing human projection from machine cognition
{"Current":[0],"efforts":[1],"to":[2,21,93],"understand":[3],"Large":[4],"Language":[5],"Models":[6],"(LLMs)":[7],"are":[8],"largely":[9],"metaphorical.":[10],"Researchers":[11],"map":[12],"LLMs":[13,75],"onto":[14],"familiar":[15],"domains,":[16],"from":[17,82,88,100],"physics":[18],"and":[19,23,39,42],"neuroscience":[20],"psychology":[22],"sociology,":[24],"each":[25],"illuminating":[26],"specific":[27],"facets":[28],"while":[29],"obscuring":[30],"others.":[31],"We":[32],"chart":[33],"these":[34],"metaphors":[35],"across":[36],"mechanistic,":[37],"behavioral,":[38],"interactive":[40],"scales":[41],"delineate":[43],"their":[44,77,95],"explanatory":[45],"boundaries.":[46],"Crucially,":[47],"this":[48,101],"metaphorical":[49],"projection":[50],"creates":[51],"a":[52],"recursive":[53],"loop":[54],"of":[55,80],"anthropomorphism,":[56],"fueling":[57],"the":[58],"\\"genuine":[59],"understanding\\"":[60],"versus":[61],"\\"pattern":[62],"matching\\"":[63],"impasse.":[64],"As":[65],"an":[66],"alternative":[67],"approach,":[68],"we":[69],"propose":[70],"machine":[71],"experientialism,":[72],"positing":[73],"that":[74,98],"build":[76],"own":[78],"form":[79],"understanding":[81],"training":[83],"corpora.":[84],"The":[85],"priority":[86],"shifts":[87],"cataloging":[89],"LLMs'":[90],"human-like":[91],"traits":[92],"uncovering":[94],"distinct":[96],"logic":[97],"emerges":[99],"text-based":[102],"world.":[103]}
Authors
- Yan Teng (ORCID: https://orcid.org/0000-0002-7069-4728)
- Lingyu Li (ORCID: https://orcid.org/0000-0003-3031-8223)
- Yue Wang (ORCID: https://orcid.org/0000-0003-0098-5359)
- Xia Hu
Institutions
- Beijing Academy of Artificial Intelligence (CN)
- Shanghai Artificial Intelligence Laboratory (CN)
Publication Details
- Journal
- Communications Psychology
- Published
- 2026-07-16
- DOI
- https://doi.org/10.1038/s44271-026-00508-6
- Primary Topic
- Artificial Intelligence in Healthcare and Education
- Type
- article
- Field-Weighted Citation Impact
- 0.00