SNS는 뇌를 썩게 한다는 것을 AI가 보여주다.

LLMs CAN GET
BRAIN ROT”
Shuo
Junyuan Hongl
Yifan Wangt?
Runjin Chent?, Zhenyu Zhang? ,
Ananth Grama >, Zhengzhong Tu’ , Zhangyang Wang?
ITcxas AcM Univcrsity -Univcrsity of Tcxas at Austin,
Purduc University
Model & Codc: https
11Lm-brain-roC
gichub
107
ABSTRACT
We proposc and (cst (he LLM Brain Rot Hypothesis: continual cxposurc to
junk wcb tert induccs lasting cognitivc dcclinc in largc languagc modcls (LLMs):
8
To causally isolatc data quality. wC run controllcd cxpcrimcnts on rcal TwittcrlX
corpora. constructing junk and rcvcrscly controllcd datascts via [WO orthogonal
opcrationalizations: MI (cngagcmcnt dcgrcc) and M2 (scmantic quality) with
8
matchcd tokcn scalc
training opcrations across conditions. Contrary to thc
control group
continual prc-training of 4 LLMs on thc junk datasct causcs non
trivial dcclincs (Hcdgcs’
0.3) on rcasoning
contcrt
undcrstanding. safely
9
and inflating
dark traits ‘
(c.g . psychopathy narcissism). Thc gradual mixturcs of
junk and control datascts also yicld dosc-rcsponsc cognition dccay: for cxamplc,
under MI
ARC-Challcngc with Chain Or Thoughts drops 74.9
57.2 and
RULER-CWE 81.1
52.3 as junk ratio riscs from 0% t0 IoD9o.
3
Error forcnsics rcvcal scvcral
insights
First
WC
idcntify thought-skipping
as the primary lesion:
modcls incrcasingly truncatc or skip rcasoning chains,
호
cxplaining most ol thc
crror
growth
Sccond, partial but incompletc’ hcaling
is obscrvcd:
instruction tuning and clcan data prc
-training improvc the
declincd cognition Yet cannot rcstorc basclinc capability; suggcsting pcrsistcnt
rcprescntational drift rathcr than format mismatch. Finally wC discovcr that thc
popularity
non-semantic mctric
(Wccl is d bcttcr indicator of thc Brain
Rot cffcct (han thc lcngth in MI
Togcthcr thc rcsults providc significant multi
돌
perspcctivc cvidcncc that data quality is
cawyal driver of LLM capability decay.
rcframing curation for continual pretraining as
traming -Timlc sajety
problem and
motivating routinc “cognitivc hcalth chccks
for dcploycd LLMs
INTRODUCTION
통
2024
thc tcrm
‘Brain Rot” was namcd the Oxford word of ycar (Oxford Univcrsity Press. 2024)
whcn it drcw
incrcasing conccrn in modcrn socicty Brain rot is dcfincd as thc dclctcrious cffcct on
human cognition that comcs from consuming largc Yolumcs of trivial and unchallenging onlinc contcnt
(or junk data) duc to Intcrnct addiction
Thc
cognitivc impact of Intcrnet addiction havc bccn found
(0 be
significant (Firlh ct a:
2019)
thrcc dimcnsions; (i) Attcntional capacitics
thc constant
strcam Or onlinc information often undcrmincs thc ability to sustain focus on
rcading articlcs or
challcnging problcms (Haliti-Sylaj & Sadiku. 2024); (ii) Mcmory proccsscs
thc abundancc
or onlinc information altcrs how individuals storc. rctricyc, and prioritizc knowlcdgc (Vedcchkina
Borgonovi, 2021); and (iii) Social
onlinc intcractions mimic rcal-world social dynamics.
rcshaping sclf-conccpts and influcncing sclf-cstccm (Youscf ct al , 2025).
Bcyond for (hcsc cognitivc
impacts
rcccnt study in Turkish population (Satici cL al , 2023) found that Internct addiction (mainly
on Xcom) is associatcd with highcr psychological distrcss and changcs in pcrsonality including
rclationship with conscicntiousncss, cxtrovcrsion, and agrccablencss.
ds wcll as
significant
positivc rclationship with ncuroticism.
In parallcl t0 (hc risc of Brain Rot in human
cognition, artificial intclligcncc, rcprcscntcd by Largc
Languagc Modcls (LLMs) grows to
human-likc
cognition (Binz & Schulz , 2023) via Icarning
Concspsondence to Jyhongeutexas
edi. a-_aswangeutexas
edu
‘Lead authors with equal
contributions
Core contributors
Xingl
ald
long .
8
kcy
scaling
along
solving
cognition
ncgative
gain

최근 공개된 한 논문이 시선올 사로잡앗다
제목부터 엄청나다: 거대 언어모델(LLM)도
인간처럼 뇌가 썩폭다는 것이다.
시가 단순히 데이터지 많이 배운다고 똑똑해지는
것이 아니라; 어떤 데이터지 학습햇나에 따라 인지
능력이 오히려 퇴행할 수도 있음올 처음으로
실험적으로 입종한 논문으로 보인다:
연구진은 거대 언어모델올 대상으로 지속
사전학습(Continual Pretraining) 실험올 설계햇다.
즉 모델이 이미 훈련된 상태에서 추가로 고참여도
웬 텍스트(junk; high-engagement web text), 쉽게
말해, 짧고 자극적인 SNS 글을 계속 학습하게
해빛다고 한다:
결과는 시간이 지날수록 모델은 점점 더
추론력(reasoning)올 원고 긴 문맥(long-
context)올 이해하지 못하여, 안전성(safety)마저
저하락다. 인간으로 치면 기억력과 집중력 그리고
공감 능력이 동시에 무너적다:
그런데 인간 역시 짧고 자극적인 영상에
익숙해질수록 뇌는 점점 깊이보다 즉시성올
추구하게 된다: 사유의 회로는 단축되고 인내의
신경망은 약해지다, 복잡한 문장을 해석하는
능력마저 서서히 문해지논데 이러한 현상이 거대
언어모델(LLM)에서도 동일하게 나타난다는 것’
증명한 첫 사례가 y듯하다.