LevelsofAGI
ployment measures should be taken at each level. Cur- asts(enjoyment),fordriver’slicensingexams(assessment),
rent SOTA generativeAIs are classified as an ASL-2 risk. orinconditionswheresensorscannotberelieduponsuch
Including items matched to ASL capabilities in any AGI as technology failures or extreme weather events (safety).
benchmarkwouldconnectpointsin ourAGItaxonomyto While Level 5 Self-Driving (SAEInternational, 2021) ve-
specificrisksandmitigations. hicles would likely be a Level 4 or 5 Narrow AI under
our taxonomy, the same considerations regarding human
6.2.Capabilitiesvs. Autonomy vs.computerautonomyapplytoAGIs.Wemaydevelopan
AGI,butchoosenottodeployit autonomously,orchoose
While capabilities provide prerequisites for AI risks, AI
todeployitwithdifferentiatedautonomylevelsindistinct
systems (including AGI systems) do not and will not op-
circumstancesasdictatedbycontextualconsiderations.
erate in a vacuum. Rather, AI systems are deployedwith
particularinterfacesandusedtoachieveparticulartasksin Certainaspectsofgeneralitymayberequiredtomakepar-
specific scenarios. These contextual attributes (interface, ticular interaction paradigms desirable. For example, the
task,scenario,end-user)havesubstantialbearingonrisk. AutonomyLevels3,4,and5(“Collaborator,”“Expert,”and
“Agent”)mayonlyworkwellifanAIsystemalsodemon-
Consider,forinstance,theaffordancesofuserinterfacesfor
strates strong performance on certain metacognitive abili-
AGI systems. Increasing capabilities unlock new interac-
ties(learningwhentoaskahumanforhelp,theoryofmind
tionparadigms,butdonotdeterminethem. Rather,system
modeling, social-emotional skills). Implicit in our defini-
designers and end-users will settle on a mode of human-
tionofAutonomyLevel5(“AIasanAgent”)isthatsucha
AI interaction (Morrisetal., 2023) that balances a variety
fullyautonomousAIcanactinanalignedfashionwithout
ofconsiderations,includingsafety. We proposecharacter-
continuoushumanoversight,butknowswhentoconsulthu-
izing human-AI interaction paradigms with six Levels of
mans(Shahetal.,2021).Interfacesthatsupporthuman-AI
Autonomy,describedinTable2.
alignmentthroughbettertaskspecification,thebridgingof
TheseLevelsofAutonomyarecorrelatedwiththeLevelsof processgulfs,andevaluationofoutputs(Terryetal.,2023)
AGI.Higherlevelsofautonomyare“unlocked”byAGIca- areavitalareaofresearch.
pabilityprogression,thoughlowerlevelsofautonomymay
be desirable for particular tasks and contexts even as we 6.3.Human-AIInteractionandRiskAssessment
reach higher levels of AGI. Carefully considered choices
Table2illustratestheinterplaybetweenAGILevel,Auton-
aroundhuman-AIinteractionarevitaltosafeandresponsi-
omyLevel,andrisk. Advancesinmodelperformanceand
bledeploymentoffrontierAImodels.
generality unlock additional interaction paradigm choices
Unlike prior taxonomies of computer automation (including full autonomy). These interaction paradigms
(Sheridanetal., 1978; Sheridan&Parasuraman, 2005; in turn introduce new classes of risk. The interplay of
Parasuramanetal., 2000) that take a computer-centric modelcapabilitiesandinteractiondesignwillenablemore
perspective (framing automation in terms of how much nuanced risk assessments and responsible deploymentde-
control the designer relinquishes to computers), we char- cisionsthanconsideringmodelcapabilitiesalone.
acterize the concept of autonomy through the lens of the
Table2alsoprovidesconcreteexamplesofeachofoursix
natureofhuman-AIinteractionstyle;further,ourontology
proposed Levels of Autonomy. For each level of auton-
considers how AI capabilities may enable particular
omy,we indicatethecorrespondinglevelsofperformance
interactionparadigmsandhowthecombinationoflevelof
andgeneralitythat“unlock”thatinteractionparadigm(i.e.,
autonomyandlevelofAGImayimpactrisk. Shneiderman
the level of AGI at which it is possible or likely for that
(Shneiderman, 2020) observes that automation is not a
paradigmtobesuccessfullydeployedandadopted).
zero-sum game, and that high levels of automation can
co-exist with high levels of human control; this view is Our predictions regarding “unlocking” levels tend to re-
compatiblewithourperspectiveofconsideringautomation quirehigherlevelsofperformanceforNarrowthanforGen-
through the perspective of varying styles of human-AI eralAIsystems;forinstance,wepositthattheuseofAIas
partnerships. a Consultantis likely with either an ExpertNarrow AI or
anEmergingAGI.Thisdiscrepancyreflectsthefactthatfor
Weemphasizetheimportanceofthe“NoAI”paradigmfor
Generalsystems,capabilitydevelopmentislikelytobeun-
manycontexts,includingforeducation,enjoyment,assess-
even;forexample,aLevel1GeneralAI(“EmergingAGI”)
ment,orsafetyreasons.Forexample,inthedomainofself-
may have Level 2 or perhaps even Level 3 performance
drivingvehicles,whenLevel5Self-Drivingtechnologyis
acrosssomesubsetoftasks. Suchunevennessofcapability
widely available, there may be reasons for using a Level
forGeneralAIsmayunlockhigherautonomylevelsforpar-
0 (No Automation)vehicle. These include for instructing
ticulartasksthatarealignedwiththeirspecificstrengths.
a new driver (education), for pleasure by driving enthusi-
8