2026当AI构建AI：AI研发自动化研讨会成果报告（英文）.pdfVIP

WorkshopReport

WhenAIBuildsAI

FindingsFromaWorkshopon

AutomationofAIRD

Authors

HelenToner*

KathleenFisher

KendreaBeers*

KellerScholl

SteveNewman*

PeterWildeford

SaifKhan*

RyanGreenblatt

ColinShea-Blymyer*

SamuelAlbani

EvelynYee*

StephanieBallard

AshwinAcharya

ThomasLarsen

*Workshopco-organizerand/orleaddrafter

January2026

ExecutiveSummary

Fordecades,scientistshavespeculatedaboutthepossibilityofmachinesthatcan

improvethemselves.Today,artificialintelligence(AI)systemsareincreasinglyintegral

partsoftheresearchpipelineatleadingAIcompanies.Someobserversseethisas

evidencethatfullyautomatedAIresearchanddevelopment(RD)isontheway,

potentiallyleadingtoarapidaccelerationofAIcapabilitiesandimpairedabilityfor

humanstounderstandandcontrolAI.OthersseetheuseofAIforresearchasa

mundaneextensionofexistingsoftwaretools.

ThisWorkshopReportsharesfindingsandconclusionsfromanexpertworkshopCSET

hostedinJuly2025.Theworkshopcoveredarangeofissuesrelatedtoautomationof

AIRD.Inthisreport,‘AIRD’referstoscientificandengineeringworkthatimproves

thecapabilitiesofAIsystemsand‘AIRDautomation’refersbroadlytoanyuseofAI

thatacceleratesprogressinAIRD.

Keytakeawaysfromtheworkshopwereasfollows:

1.IncreasinglyautomatedAIRDisapotentialsourceofmajorstrategic

surprise.Whileexpertsdisagreeonlikelihood,scenariosarepossibleinwhich

AIRDbecomeshighlyautomated,thepaceofAIRDaccelerates

dramatically,andtheresultingsystemsposeextremerisks.Thiswarrants

preparatoryactionnow.

2.FrontierAIcompaniesarealreadyusingAItoacc