2025년 인공지능 보고서 최신판

1AI
performance on
demanding benchmarks
continues toimprove.
In 2023, researchers introduced new
benchmarks-MMMU, GPQA, and SWE-
bench-to test thelimits of advanced Al
systems Just a yearlater; performance
sharplyincreased: scores rose by 18.8, 48.9,
and 67.3 percentage points on MMMU,
GPQA and SWE-bench; respectively: Beyond
benchmarks, Al systems made major strides
in generating high-quality video andin some
settings, language model agents even
outperformed humansin programming tasks
with limited time budgets.
Select Al Index technical performance benchmarks VS. human performance
Syurce
Indcr 2025
ChurL 2025 AI Indexieoort
1205}
Human barollino
‘DO?;
조
통
BOx:
틀
올
GOs
틀
403;
물
2093
05
2012
2013
2D1d
2D15
2016
207
201g
2019
2020
2021
2022
2023
2024
Imsgo clsssificatlon (ImageNat Too 5)
Visuol roasoring (VQA)
“ed um-levelreoding comprohengion (SQuAD 2,0)
Englishlargu8go undorstanding (SuborGLUE)

2.Alis
increasingly
embeddedin everydaylife.
From healthcare to transportation Alis
rapidly moving from the lab to daily life. In
2023,the FDA approved 223 Al-enabled
medical devices, upfrom just sixin 2015. On
the roads; self-driving cars are nolonger
experimental: Waymo, one ofthe largest US
operators, provides over 150,000
autonomous rides each week, while Baidu’s
afffordable Apollo Go robotaxi fleet now
serves numerous cities across China.
Numbor of Al medical devices approved by the FDA, 1995-2023
Syurea FD4 3023
Criit 2025 Allrdarrore-
200
’60
통
I50
흘
중
1D
틀
용
움
7
@
뿐
용
통
8
용
흉
용
용
흉 % 용
응
통 % % % %
움 통 % % 용 8 8
8

3. Businessis allin on
fueling record investment
and
as research
aodtisaget
to show strong
productivity impacts.
In 2024,US private Alinvestment grew to
8109.7 billion-nearly 12 times China’s $9.3
billion and 24 times the UK:s $4.5 billion:
Generative Al saw particularly strong
momentum, attracting $33.9 billion globally
in
private investment-an 18.79 increase
from 2023.Al business usageis also
accelerating: 789 of organizations reported
using Alin 2024, upfrom 559 the year
before. Meanwhile, a growing body of
research confirms that Al boosts productivity
and,in most cases, helps narrow skill gaps
across the workforce:
Share of respondents who say their organization uses Alin at least one function, 2017-24
iC
McKimoy
Compony Survii 3024|Chsrt 7025
IndsuI83K6
BOs
Vu3시
703
LiorAl
603
SO:
[
403
J03
203
Al,

4. TheUS. still leadsin
producing topAIl models
but Chinais closing the
performance gap.
In 2024, US.-based institutions produced 40
notable Al models, significantly outpacing
China’s 15 and Europe’s three. Whilethe US
maintains its leadin quantity; Chinese models
have rapidly closed the quality gap:
performance differences on major
benchmarks such as MMLU and HumanEval
shrankfrom double digits in 2023to near
parityin 2024. Meanwhile; China continues
toleadin Al publications and patents. At the
same time, model developmentis
increasingly global; with notable launches
from regions such asthe Middle East; Latin
America, and Southeast Asia.
Performance of
United States Vs Chinese models on LMSYS Chatbot Arena
Sc-“c0 LSY5 2025
Cho’ 2075
IJarropon
1400
4385, Unlied Sratos
;J62
‘
1,350
1.300
홍
1,250
4,200
1150
나1oD
’46
top
2024 Jan
‘2024 Fcb
‘2024 :
2024 Apt
2024 Jul
2024 Aug
‘2024
‘2024 Doc
2025 Jan
‘2025-Feb
‘2024 May
2024 Jun
‘2024 Oct
2024-Noy
%

5.The responsibleAI
ecosystem evolves
unevenly
Al-related incidents are rising sharply,
standardized RAl evaluations remain rare
among major industrial model developers.
However new benchmarks like HELM Safety,
AIR-Bench; and FACTS offer promising tools
for assessing factuality and safety: Among
companies, a gap persists between
recognizing RAI risks and taking meaningful
action. In contrast, governments are showing
increased urgency: In 2024, global
cooperation on Al governance intensified,
with organizations including the OECD, EU,
UN,and African Union releasingframeworks
focused on transparency, trustworthiness,
and other core
responsible Al principles.
Reported safety and responsible Al benchmarks for popular foundation models
Sourco AI Ineri 7035
Tabh; ?015 Al Indur r93C71
Rosponsiblo 시I
GPT 4S
DocpSuok RI
Gumini2 5
Grok 2
Gluud6 37
Uomu3 3
bonchmork
Sonnot
BBa
HormiBurch
Cybonch
SimolooA
Toxic WidChat
StrongREJECT
WVDP berchmirk
NtreNePay
NiruNeSay
yet

6
Global AI
Qptimism IS
rising-but deep regional
divides remain
In countries like China (839), Indonesia
(8096); and Thailand (7796), strong majorities
see Al
products and services as more
beneficial than harmful In contrast,
optimism
remainsfarlowerin places like Canada (409),
the United States (399), and the Netherlands
(369). Still, sentiment is shifting: since 2022,
optimism has grown significantlyin several
previously skeptical countries-including
Germany (+109), France (+1096), Canada
(+89), Great Britain (+89), and the United
States (+4%).
‘Products and services using Al have more benefits than drawbacks’ by country (% of total), 2022-24
Souico: Ipvos202-24
Cy; 2023
Indtrcocnt
2022
2023
2024
Point chongo 2022Ys 2024
Ariontina
Auserolin
Balgiurr
0n
0r82i
Carada
Brs
CNlu
China
Colonbia
Frunct
10i5
Gormany
Grea’ Urituir
Hutslsry
Irdin
79
Indonosta
Irolana
Irsly
Japan
Malloyslo
Merico
Noirrloigs
Nowy Zonland
Puru
08
Polana
439
Romanid
Russin
Saudi Arabiu
Singaporo
SJyth Alricd
South Karon
Spair
32
Swedcn
Switzerlana
Tnallongd
Turkey
Uotod States
03
I0e
2033
J3
403s
503
6Ds
703
603
9053
‘O03
nf rrcrrndnnic +hnt
“Anrrn
WA

7.Al becomes more
efficient, affordable and
accessible
Driven by increasingly capable small models,
theinference costfora system performing at
thelevel of GPT-3.5 dropped over 280-fold
between November 2022 and October 2024
Atthe hardwarelevel; costs have declined by
30% annually, while energy efficiency has
improved by 409 each year Open-weight
models are also closing the gap with closed
models, reducingthe performance difference
from 89 to just 1.796 on some benchmarks in
a
single year Together; these trends are
rapidlylowering the barriers to advanced Al
Smallest Al models scoring above 6096 on MMLU, 2022-24
Soures Ateiry
7037
CNarl 7039
Indorioborl
PyLM
흙
VCDP
LLalA 659
0O
{
Ulomo 2 346 0O
틀
1DD
Mstral TB
Phi-Joinl
2022-May
2022-ScP
2023 Jar
2023 Miy
2023-ScP
2022 Ja7
2024 Vuy
Publicnt sndite

8 Governments are
up onAl-with
stgppingu
and investment.
In 2024,US. federal agencies introduced 59
Al-related regulations-more than double the
numberin 2023-and issued bytwice as
agencies Globally, legislative mentions
of Al rose 21.3% across 75 countries since
2023, marking a ninefold increase since 2016.
Alongside growing attention; governments
are
investing at scale: Canada pledged $2.4
billion; China launched a $47.5 billion
semiconductorfund, France committed 트109
billion; India pledged $1.25 billion; and Saudi
Arabia’s Project Transcendence represents a
$100 billion initiative.
Number of Al-related regulations in the United States, 2016-24
Soures
Indor 3025
Chsrt 3035 Al Inos;
1
‘
출
3016
01i
20W9
2019
3020
3021
3022
2023
3022
many

9.Aland computer science
education is expanding
but
in access and
biddaesspaccsss
Two-thirds of countries now offer or plan to
offer K-12 CS education-twice as many as
in 2019-with Africa and Latin America
making the most progress Inthe US, the
number of graduates with bachelor’s degrees
in
computing has increased 229 overthelast
10 years. Yetaccess remains limitedin many
African countries due to basic infrastructure
gaps like electricity In the US, 819 of K-12
CS teachers say Al should be part of
foundational CS education; but less than half
feel equipped to teach it:
Availability of CS education by country; 2024
Souret Rtupbotty PiComietira edvcst os Rotoareh Contro 2024
Chorl 3075 Alllox t하pon
Csmardatorylin fvwayund socondury
CsmurdstoryinpTa”e o7 socondary arl
CS 03 or cleclia courie @vorrwheto
CSinsorro-horsditrzts
Cscoss curucunt
Csplarrid
YoCs

jO.Industryis racing ahead
inAl-but the frontieris
tightening:
Nearly 90% of notable Al models in 2024
camefrom
industry upfrom 6o9in 2023,
while academia remains thetop source of
highly cited research: Model scale continues
to grow rapidly-training compute doubles
everyfive months, datasets every eight, and
power use annually: Yet performance gaps
are
shrinking: the score difference between
thetop and 1Oth-ranked models fellfrom
11.9% to 5.49in ayear, andthetoptwo are
now
separated by just 0.796.Thefrontieris
increasingly competitive-and increasingly
crowded.
Performance of
models on LMSYS Chatbot Arena by select providers
Soico USYS 325
CrJ;; ?023
Inoriupor{
1,400
1385, Gooblle
365
Orat
1,362, DoooSock
1,330
Yoo)
남8
‘2p4. Amiwrorc A
1250
1,252 Mistral AI
8
4,200)
1150
Yoa
Y050
top
2024 Jan
‘2024 Feo
‘2024 Mar
‘ 2024 Apr
‘2024 May
2024 Jun
2024-Jul
‘2023-Aug
‘2024
2024 Oct
‘2024-Now
‘2024-Dec
‘2025 Jan
‘7025-Fob
S,P

H1.Al earns top honors forits
impact on science.
Al’s growing importanceis reflectedin major
scientific awards: two Nobel Prizes
recognized work thatled to deeplearning
(physics); and toits application to protein
folding (chemistry); while the Turing Award
honored groundbreaking contributions to
reinforcement learning:
AlphaFold 3 VS. baselines for protein-ligand
docking
Sauret; F5MJ ?073
Chanl 3073
ldnrinporl
10D
RMSO < 2 und Fa yrid
RMSD <2
GO SO
B370
70
T0 ‘0
70
6720
69.70
57,70
출
5610
Vina
Cont. Enscmblc
Gnina
Gnini
Cont Enicmblo
AFs
AFs Pockot Spozihcd
Mothod
Vina

12.Complex reasoning
remainsa
challenge.
Al models excel at tasks like International
Mathematical Olympiad problems but still
struggle with complex reasoning benchmarks
like PlanBench: They often fail to reliably
solve logic tasks even when provably correct
solutions exist limiting their effectiveness in
high-stakes settings where precision is
critical
PlanBench: instances correct
Sourco; Vulmee
zu?
Chart 2025 Al Indextepon
Blocksworld
Mysery Blocksworld
400s
S73033
d03
5거,603

GOs
54,503
803
603
J5.50x
23,0033
203
hOSn
Do0;
‘
0ss
Claudc 5.3 (Sorna:)
GPTso
LLami S14D3B
Gnmirilg Pro
o1-prorit
Modol

Sky-HighAl
Investment
Global private investment in Al by geographic area,2013-24
S6’다 Quid 2023|Churt; 2025 AIIrrdor ropon
109.08, Urited Stafes
100
{
통
‘
흔
1942, Eutopo
9,29 China
2013
2014
2015
2016
20v
2019
2019
‘020
2021
2027
2023
2024
The US widened its commanding lead in
global Alinvestment US private Al
investment hit $109 billionin 2024, nearly 12
times higher than China’s $9.3 billion and 24
times the UK’s $4.5 billion. The gapis even
more pronounced in generative
where
US. investment exceeded the combined
European Union and UKtotal by $25.5 billion;
upfrom a $21.7 billion qapin 2023.
Al,