2026负责任AI进展报告.docxVIP

ProgressReport

PublishedinFebruary2026

2

Foreword:

Theopportunity

oftheAIera

If2024wasdefinedbybuildingoutthefoundationsforanAIfuture,2025markedAI’sshiftintoahelpful,proactivepartner,capableofreasoningand

navigatingtheworldwithusers.Asmodels

growevenmoresophisticated,weseeusersandbusinessesaroundtheglobetransitioningfromexplorationtointegration,findingnewwaystoputthesetoolstoworkintheirdailylives.Fromfoundationaladvancesinscientificdiscovery

andclinicalmilestonesinhealthcaretotheriseofagenticsystemsandnewtoolstosupportcreativitysuchas‘vibecoding’andgenerativemedia,thetransformationalpotentialofthesetoolsiscomingmoreclearlyintofocus.

Sincewestartedpublishingthesereports,

ourresponsibleAIdevelopmentapproachhas

continuedtomatureandisnowfullyembedded

withourproductdevelopmentandresearch

lifecycles.In2025,asmodelsbecamemorecapable,personalized,andmulti-modal,werelieduponrobustprocessesfortestingandmitigatingrisks,and

deepenedtherigoroussafeguardsbuiltintoour

products.TomeetthischallengeatthespeedandscaleofGoogle,wehavepaired25yearsofuser

trustinsightswithacomprehensivetestingstrategythatisdrivenbyhumanexpertiseandsupportedbyAI-enabledautomation.

ThisworkcontinuestobeguidedbyourAIPrinciples,whichweupdatedlastyeartorelectourlatest

understandingoftheopportunitiesandrisks

presentedbythisplaformshift.Today’sreport

detailsourmulti-layeredapproachtoresponsible

AIgovernance,andfocusesinparticularonagenticandfrontierrisksfromincreasinglysophisticated

models.Insuchadynamicenvironment,italso

showshowoursystemsarebuilttobeabletodetectandthenadapttoemergingrisks.Whetherwearehardeningagenticsystemsagainstadversarial

manipulationorembeddingprovenancesignals

intoeverysyntheticoutput,ourgoalremainsclea