January2026

Srinivasan(Cheenu)

Venkatachary

8min.

BuildingPersonalIntelligence:

Asteptowards

trulypersonalAI

1IntroductionPage03

2HowPersonalIntelligenceworksPage05

3ResponsibleDevelopmentPage08

4ResearchDirectionsPage12

5VisionfortheFuturePage13

03BuildingPersonalIntelligence:AsteptowardstrulypersonalAI

Introduction

Makingourproductsmorehelpfulbymakingthempersonal

AtGoogle,we’vespentyearsrefiningourproductstobenotjusthelpful,buthelpfultoyou.

Welearnedearlyonthatpersonalcontextmatters:when

someonesearchesfor“runningshoes”,theyaren’tusually

lookingforalistofgenericbestsellers,butforthespecific

brandsandstylestheyprefer.Thatinsight—thatyourhistoryandpreferencesshouldinformyourresults—enhanced

GoogleSearch,andintheyearssince,we’vebroughtthatsamelevelofpersonalizationtomanyofourproducts.

Interactingwithyourpersonalinformationhashistoricallymeantnavigatingindependentproductexperiences:

searchingforyourflightreservationrequiredopeningGmail,whilefindingamemorymeantsearchingorscrollingthrough

GooglePhotos.

Inthelastfewyears,webeganbridgingthesegapswith

featuresthatallowedforexplicitinformationretrievalacrossproductswithyourpermission—liketagging@GmailtofindflightdetailsintheGeminiapp.Wealsointroducedfeaturesthatpersonalizeyourexperienceinaspecificproduct—

likeusingyourpastchatsintheGeminiapptogiveyouamoretailoredresponse.Butthesestilldon’tprovideafullypersonalizedexperience.

Introduction

04BuildingPersonalIntelligence:AsteptowardstrulypersonalAI

IntroducingPersonalIntelligence

PersonalIntelligencemarks

theshifttoAIthatcantruly

understandyourpersonalcontext.

Itbringstogethertheinformationyoushareandtheactivitywithintheproductsyouuse,whileconnectingthedotsacrossyourGoogleappstotomakeAI

uniquelyhelpfulforyou.

WithPersonalIntelligence,youcanchoosetoconnectcertainGoogleappstogetmoretailoredresponses

—allwhileyourinformationremainssafeguarded.

Today,youcanuseabetaversionofPersonal

IntelligenceintheGeminiapp.Itsecurelyconnects

informationfromappslikeGmailandGooglePhotostomakeGeminiuniquelyhelpful.PersonalIntelligenceiscomingsoontoAIModeinSearch.

Thatmeansthe

Geminiapp

canhelpwithmore

tailoreddiscoveryandcomplexplanning—like

helpingyouplanforspringbreakbasedonplaces

you’vealreadybeenorproactivelyfindingtheperfectsetoftiresforyourspecificcarmakeandmodel.

Ifyouturnthisbetafeatureon,youcontrolwhichappstolink,andeachonesuperchargesthe

experience.

Thisisafoundationalsteptowardmovingbeyond

genericassistancetoAIthatworksforyou.It’sstill

earlydaysforthistechnology,andwe’recontinuingtoworkthroughknowntechnicalissuesandlimitations.Aswecontinuetotestandlearn,weareeagertohearyourfeedback.

05BuildingPersonalIntelligence:AsteptowardstrulypersonalAI

HowPersonalIntelligenceWorks

TheTechnicalChallenge:

SolvingtheContextPackingProblem

PersonalIntelligenceunlocksnewutilitybysolving

forthecontextpackingproblem:enablingour

Geminimodelstosafelyandaccuratelyreasonoverdisparateandvastamountsofpersonaldatasourcesinreal-timewithoutcompromisinguserprivacy.

Products

withPersonal

IntelligenceGeminiappSearch

(comingsoon)

AIModel

connected

toPersonal

Intelligence

Engine

PersonalIntelligenceEngine

Securedata

retrievalwith

userpermission

GmailPhotosMore

GeminimodelenablingPersonalIntelligence

HowPersonalIntelligenceWorks

06BuildingPersonalIntelligence:AsteptowardstrulypersonalAI

AdvancedGeminiModels&

ANewPersonalIntelligenceEngine

Tohandlethecomplexityofretrievingdatafrommultiplesourcessimultaneouslyandprovidesignificantlymorehelpfulresponsestoyou,webuiltanewengineforPersonalIntelligence.

PersonalIntelligencehastwocore

strengths:toolcallstoretrievespecificdetailsandreasoningacrossmany

complexsources.Itoftencombines

bothapproachesandcanworkacrosstext,photos,andvideotogiveyou

one-of-a-kindanswers.

ItenablesGeminimodelstosecurely

retrieverelevantcontext,leveraging

advancesintooluse,denseretrieval,andlongcontextcapabilitiestoreasonacrossyourpersonaldatafromGoogleproductsinreal-time.

AdvancedReasoning

Gemini3

,ourmostintelligentmodelseriestodate,

isbetteratgeneralunderstandinganddeciphering

moredepthandnuance—capabilitiesthatarecriticalforunderstandingcomplexpersonalcontext,such

asmappingfamilialrelationshipsorrecognizingyourspecificaestheticpreferences.

AdvancedToolUse

TheGeminimodelalsohassignificantadvancesintool

usecapabilitieswhichmeansitcanunderstandyourgoalandretrievemoreinformationrelatedtoyourpreferencesfromthePersonalIntelligenceengine.Thisretrievalalsobuildsonthefoundationofresearchwe’vedoneon

searchanddenseretrieval,suchas

GeminiEmbeddings

.

LongContext

Gemini3hasa1milliontokencontextwindowwhichenablesreasoningacrossavastamountofdata.

However,trulyhelpfulpersonalizationrequiresprocessingatamuchlargerscale,asauser’saccumulatedcontext

acrossemailsandphotosaloneoftenexceedsthiswindowbyordersofmagnitude.

Tobridgethisgap,weutilize“contextpacking”

—atechniquethathelpsusdynamicallyidentifyandsynthesizeappropriatepiecesofinformationintotheworkingmemoryforthemodel.

07BuildingPersonalIntelligence:AsteptowardstrulypersonalAI

WiththisnewengineandGemini3,wenowhavethecapabilities

requiredfortruepersonalization.

Forexample,whenyouaskGeminito“Planalistof

restaurantsformyupcomingtripthatarenearmy

hotel”,withPersonalIntelligence,youwon’tjustget

generictop-ratedspots.TheGeminimodelunderstandsthatthistaskrequiressynthesizingdisparatepersonal

detailsfromacrossGoogleproducts—yourhotel

reservations,flightarrivaltime,pastdininghistory,andtheaspirationalspotsyou’vesaved.

?Themodelagenticallyandsecurelyexecutes

searchesforthepersonalinformationthat’srelevanttotheresponse,lookingforyourtripinrecentemails,butalsootherrelevantinformationrelatedtothe

query,likepastrestaurantreservationstounderstandwhatyoulove.

?Finally,itdeliversatailoredsetofpersonal

recommendationsclosetoyourupcoming

accommodationsbymakingsenseofyourdatalikeyourphotosandemails,aswellasthingslikeyour

pastGeminiappchatconversations,SearchqueriesandYouTubehistory.

Thisfundamentallyshiftsourarchitectureand

ourapproachtopersonalization:wearemoving

towardsaworldwhere,withyourpermission,productsliketheGeminiappcansecurelyaccesscertaintypesofpersonalinformationasacontinuousstreamof

contexttoinformeveryinteraction–deliveringreal,tailoredhelpfulness.

08BuildingPersonalIntelligence:AsteptowardstrulypersonalAI

ResponsibleDevelopment

It’scriticaltodeveloptechnologylikePersonalIntelligence

responsibly.Inaccordancewith

ourAIprinciples

,weare

focusedonbuildingthistechnologysecurely,whileprotectinguserprivacyandsettingguardrailsforsensitivetopics.

Forexample,themodelaimstoavoidmakingproactiveassumptionsaboutsensitivedatalikeyourhealth,thoughitwilldiscussthisdataifyouask.

Privacy&

PersonalIntelligence

AkeyareaoffocusforuswhenbuildingPersonalIntelligencewasprotecting

userprivacyandsecurelyconnectingdatasourcesacrossGoogleapps.

Usercontrolsbydesign

Youcanchoosewhetherornottoturnthesefeaturesonorof.IntheGeminiapp,youcan

manageallyourpreferences

directlyinyoursettings

,likechoosingwhichservices—

suchasGoogleWorkspace,GooglePhotos,YouTube,

andSearch—youwanttoconnectasapartofPersonalIntelligenceinthe“ConnectedApps”settings.These

connectedappsettingsareofbydefault.

Securelyconnecteddatasources

Westartbybuildingonourbest-in-classsecurity

infrastructureandimplementadditionalindustry-leadingsafeguardstoensurethatthisdataremainsprotected

evenasitpowersnewAIexperiences.Forexample,

userdataisencryptedatrestbydefaultandprotected

intransitbetweenoursystemsusingApplicationLayer

TransportSecurity(ALTS).We’vealsodoneworkto

increaseresistancetopromptinjectionsandhaveimprovedprotectionagainstmisuseviacyberattacks.

LimitedgenerativeAItraining

Ourgoalistoimproveyourexperiencewhilekeepingyourdatasecureandunderyourcontrol.Builtwithprivacyin

mind,GeminiAppsdon’ttraindirectlyonyourGmailinboxorGooglePhotoslibrary.Toimprovefunctionalityover

time,wetrainoninfolikepromptsandresponsesinGeminiaswellassummaries,excerptsandinferencesusedtohelpansweryourprompts.Tolearnmore,visitthe

HelpCenter

.

ResponsibleDevelopment

09BuildingPersonalIntelligence:AsteptowardstrulypersonalAI

LikemanyemergingAIfeatures,PersonalIntelligenceisstill

evolving.Thistechnologymaymakemistakes—likemisinterpretingcontextormakingincorrectassumptionsaboutyouractivity.

KnownTechnical

Limitations&Issues

YoucancorrectGeminiifitmakesmistakesdirectlyviaaprompt(e.g.

rememberthatIdon’teatmeat).

We’realsoworkingtoaddressknown

issuesthroughrigorousinternaltestingandmodeltuning,butweknownew

challengeswillarise.Your

feedback

iskeytoidentifyingthemandmakingthistechnologyashelpfulaspossible.

Herearesomeofthekeytechnical

challengeswe’reworkingtoaddress:

?Overpersonalizationbasedonyourinterests

Aknownchallengeisthetendencyforthemodeltorelytooheavilyonapersonalizedinferencewhereit’snot

appropriate—aphenomenonwecall“tunnelvision”.

Example:YoumightbeabigfanofcofeeshopsandthemodelunderstandsthataspartofPersonalIntelligence.Whenyouaskitto“planatriptoAustralia”,itmay

inadvertentlyplanatripwheretheitineraryisfocusedoncofeeshops.

Example:Ifyouhaveanemailinyourinboxaboutyouremployment,itmightstartanchoringyourresponsesaroundthefactthatyou’reasoftwareengineer.

Example:Ifyouask“whatkindofsocksshouldIbuy?”,itcouldassumethatbecauseyouhaveamarathoncomingupyouareonlylookingforathleticrunningsocks.

?Mistakinganotherperson’spreferencesforyourown

Intestingwe’vealsoseenachallengearoundthemodelconflatingsubjects—forinstance,attributingafamilymember’sintereststoyou.WhenyoushareahouseholdaccountforserviceslikeYouTube,ordoresearchor

makeapurchaseforafriendorfamilymember,the

systemmaymistakeothers’preferencesasyourown.

Example:Basedonareceiptinyouremail,themodel

mightthinkyouenjoylisteningtoheavymetalandofersuggestionsonconcertsnearby,whenyouactually

purchasedtheticketsasabirthdaygiftforyourbrother.

?Incompleteinformation

Therearesomeinstanceswhereyoumightnotseeall

ofyourpersonalinformationifyouask.Alltherelevantinformationmaynothavebeenretrievedoroursystemsmightmakeinaccurateinferencesbasedonthe

informationavailable.

Example:Ifyouaskforasummaryoflastmonth’s

activities,wemightonlyhaveinformationforafractionofthem.

ResponsibleDevelopment

10BuildingPersonalIntelligence:AsteptowardstrulypersonalAI

KnownTechnical

Limitations&Issues

?Mixinguptimelines

TemporalrelevanceisaknownissueforAImodelsgenerallyandaddingpersonalhistoryaddsadditionalcomplexityandmakesittrickytomakerelevantconnections.

Example:Themodelmaymixuptiming,notingthata

graduateprogramapplicationdeadlinefromyouremailisinthepast,wheninfactit’sstillupcoming.

?Misinterpretingrelationships

Themodelalsohaschallengesunderstandingandgraspingthenuanceofrelationshipsandcomplexdynamics,

sometimesmisidentifyingfamilyroles.

Example:Itcanmisidentifyamotherforagrandmotherbasedonambiguoustextinemailsorlabelasiblingasafriend.

?Missingmajorlifechanges

Themodelwon’talwaysknowwhenamajorchangeinyourlifehasoccurred,suchasadivorceoradeathinthefamily.

Example:Themodelmightsuggestananniversarydinnerreservationforapartneryouarenolongerwith.

?Incorrectassumptions

Themodeloftenassumesthatatransactionrecordequalsacompletedaction.Itmayassumeyouboughtanitemorattendedaneventbasedonaconfirmationemail,missingthesubsequentcontextthatyoureturnedtheitemor

cancelledthereservation.

Example:Themodelcouldrecommendafollow-upbookinaseriesbecauseyouboughtthefirstone,failingtorealizeyoureturneditthenextday.

?Overlookingcorrections

Ifyoucorrectthemodelaboutyourpersonalinformation,itmightbemissedsometimes.Thisoftenhappenswithmoreambiguousprompts.

Example:Youtellthemodel,“Idon’tusuallyeatsteak,”butitsuggestsasteakhouserecommendationagainaweek

latereventhoughyougenerallydon’tprefersteak.

ResponsibleDevelopment

11BuildingPersonalIntelligence:AsteptowardstrulypersonalAI

OtherTechnicalChallenges?Balancingspeedanddepth

Deliveringtruepersonalizationrequiresmorecomplexprocessing,sowe’reconstantlynavigatingthedelicatebalancebetweenlatencyandqualitytradeofs.

Toensurethebestuserexperience,thesystem

distinguishesbetweengeneralquerieswithout

personalizationandcomplexpersonalrequestsandwe’llcontinuetoiterateonthisasourmodelsandtechnologyevolve.

Forgeneralquestionswithoutpersonalization,you’llgetafasterresponse.

Withmorepersonalquestions,youmayseea“thinking”

indicatorintheproductwhichmaysay“Personalizing”or

“PersonalIntelligence”.Thisvisualizestheprocessingstepswhilethemodelsecurelyretrievesandreasonsoveryour

personalinformationtoprovideathoroughanswer.

IntheGeminiapp,youcanalsochoosetheoption“Answernow”togetafasterresponseifyou’reusingthe“Thinking”or“Pro”models.

?Ta