LevelsofAGI
Table2.MorecapableAIsystemsunlocknewhuman-AIinteractionparadigms. Thechoiceofappropriateautonomylevelneednotbe
themaximumachievablegiventhecapabilitiesoftheunderlyingmodel.Oneconsiderationinthechoiceofautonomylevelareresulting
risks.Thistable’sexamplesillustratetheimportanceofcarefullyconsideringhuman-AIinteractiondesigndecisions.
AutonomyLevel ExampleSystems Unlocking ExampleRisks
AGILevel(s) Introduced
AutonomyLevel0: Analogue approaches (e.g., sketching NoAI n/a(statusquorisks)
NoAI withpencilonpaper)
humandoeseverything
Non-AI digital workflows (e.g., typ-
ing in a text editor; drawing in a paint
program)
AutonomyLevel1: Information-seeking with the aid of a Possible: de-skilling
AIasaTool searchengine EmergingNarrowAI (e.g.,over-reliance)
human fully controls task
and uses AI to automate Revising writing with the aid of a Likely: disruptionof
mundanesub-tasks grammar-checkingprogram CompetentNarrowAI established
industries
Readingasignwitha
machinetranslationapp
AutonomyLevel2: Relyingonalanguagemodeltosumma- Possible: over-trust
AIasaConsultant rizeasetofdocuments CompetentNarrowAI
AItakesona radicalization
substantive role, but only Accelerating computer programming Likely:
wheninvokedbyahuman withacode-generatingmodel ExpertNarrowAI; targeted
EmergingAGI manipulation
Consuming most entertainment via
asophisticatedrecommendersystem
AutonomyLevel3: Training as a chess player through Possible: anthropomorphization
AIasa interactions with and analysis of a EmergingAGI (e.g.,parasocial
Collaborator chess-playingAI relationships)
co-equal human-AI collab- Likely:
oration;interactivecoordi- Entertainment via social interactions ExpertNarrowAI; rapidsocietalchange
nationofgoals&tasks withAI-generatedpersonalities CompetentAGI
AutonomyLevel4: UsinganAIsystemtoadvancescientific Possible: societal-scaleennui
AIasanExpert discovery(e.g.,protein-folding) VirtuosoNarrowAI
AI drives interaction; hu- masslabor
man provides guidance & Likely: displacement
feedback or performs sub- ExpertAGI
tasks decline of human ex-
ceptionalism
AutonomyLevel5: AutonomousAI-powered Likely: misalignment
AIasanAgent personalassistants VirtuosoAGI;
fullyautonomousAI (notyetunlocked) ASI concentration
ofpower
Considering AGI systems in the context of use by people 7. Conclusion
allows us to reflect on the interplay between advances in
Artificial General Intelligence is a conceptof both aspira-
models and advances in human-AI interaction paradigms.
tionalandpracticalconsequences.Weanalyzedninedefini-
The role of model building research can be seen as help-
tionsofAGI,identifyingstrengthsandweaknesses. Based
ing systems’ capabilities progress along the path to AGI
onthisanalysis,weintroducedsixprinciplesforaclear,op-
in their performance and generality, such that an AI sys-
erationalizabledefinitionofAGI:focusingoncapabilities,
tem’sabilitieswilloverlapanincreasinglylargeportionof
notprocesses;focusingongeneralityandperformance;fo-
humanabilities. Conversely,theroleofhuman-AIinterac-
cusing on cognitive and metacognitive (rather than phys-
tion research can be viewed as ensuring new AI systems
ical) tasks; focusing on potential rather than deployment;
are usable by and useful to people such that AI systems
focusing on ecologicalvalidity for benchmarking; and fo-
successfullyextendpeople’scapabilities(i.e.,“intelligence
cusingonthepathtoAGIratherthanasingleendpoint.
augmentation”(Brynjolfsson,2022;Englebart,1962)).
With these principles in mind, we introduced our Levels
9