/LMAR=0/XLINE=3/FONT#0=BAXL30/FONT#1=BAXM30/FONT#2=BASB30/FONT#3=SUB/FONT#4=SUP/FONT#5=BASL35/FONT#6=NGR25/FONT#7=MATH30/FONT#8=FIX25/FONT#9=GRKB30/FONT#10=ZERO30
␈↓ α∧␈↓␈↓ u1


␈↓ α∧␈↓α␈↓ ∧→SOME EXPERT SYSTEMS NEED COMMON SENSE

␈↓ α∧␈↓α␈↓ ¬
John McCarthy, Stanford University

␈↓ α∧␈↓α␈↓ ¬.Copyright 1983, John McCarthy

␈↓ α∧␈↓␈↓ αTAn␈α␈↓↓expert␈αsystem␈↓␈αis␈αa␈αcomputer␈αprogram␈αintended␈αto␈αembody␈αthe␈αknowledge␈αand␈αability␈αof

␈↓ α∧␈↓an␈α
expert␈α
in␈α
a␈α
certain␈α
domain.␈α
 The␈αideas␈α
behind␈α
them␈α
and␈α
several␈α
examples␈α
have␈αbeen␈α
described

␈↓ α∧␈↓in␈α∞other␈α∂lectures␈α∞in␈α∂this␈α∞symposium.␈α∂ Their␈α∞performance␈α∂in␈α∞their␈α∂specialized␈α∞domains␈α∂are␈α∞often

␈↓ α∧␈↓very␈α⊂impressive.␈α⊂ Nevertheless,␈α⊂hardly␈α∂any␈α⊂of␈α⊂them␈α⊂have␈α∂certain␈α⊂␈↓↓common␈α⊂sense␈↓␈α⊂knowledge␈α∂and

␈↓ α∧␈↓ability␈α
possessed␈α∞by␈α
any␈α
non-feeble-minded␈α∞human.␈α
 This␈α∞lack␈α
makes␈α
them␈α∞"brittle".␈α
 By␈α∞this␈α
is

␈↓ α∧␈↓meant␈α∀that␈α∀they␈α∀are␈α∀difficult␈α∀to␈α∀extend␈α∀beyond␈α∀the␈α∀scope␈α∀originally␈α∀contemplated␈α∀by␈α∀their

␈↓ α∧␈↓designers,␈α∞and␈α∂they␈α∞usually␈α∂don't␈α∞recognize␈α∂their␈α∞own␈α∂limitations.␈α∞ Many␈α∂important␈α∞applications

␈↓ α∧␈↓will␈α∩require␈α⊃common␈α∩sense␈α⊃abilities.␈α∩ The␈α⊃object␈α∩of␈α⊃this␈α∩lecture␈α⊃is␈α∩to␈α⊃describe␈α∩common␈α⊃sense

␈↓ α∧␈↓abilities and the problems that require them.

␈↓ α∧␈↓␈↓ αTCommon␈α
sense␈α
facts␈αand␈α
methods␈α
are␈αonly␈α
very␈α
partially␈αunderstood␈α
today,␈α
and␈αextending

␈↓ α∧␈↓this understanding is the key problem facing artificial intelligence.

␈↓ α∧␈↓␈↓ αTThis␈αisn't␈αexactly␈αa␈αnew␈αpoint␈αof␈αview.␈α I␈αhave␈αbeen␈αadvocating␈α"Computer␈αPrograms␈αwith

␈↓ α∧␈↓Common␈αSense"␈αsince␈αI␈αwrote␈αa␈αpaper␈αwith␈αthat␈αtitle␈αin␈α1958.␈α Studying␈αcommon␈αsense␈αcapability

␈↓ α∧␈↓has␈α⊂sometimes␈α⊂been␈α⊂popular␈α⊂and␈α⊃sometimes␈α⊂unpopular␈α⊂among␈α⊂AI␈α⊂researchers.␈α⊂ At␈α⊃present␈α⊂it's

␈↓ α∧␈↓popular,␈α⊗perhaps␈α⊗because␈α⊗new␈α⊗AI␈α⊗knowledge␈α∃offers␈α⊗new␈α⊗hope␈α⊗of␈α⊗progress.␈α⊗ Certainly␈α∃AI

␈↓ α∧␈↓researchers␈αtoday␈αknow␈αa␈αlot␈αmore␈αabout␈αwhat␈αcommon␈αsense␈αis␈αthan␈αI␈αknew␈αin␈α1958␈α-␈αor␈αin␈α1969

␈↓ α∧␈↓when␈α∂I␈α∞wrote␈α∂another␈α∞paper␈α∂on␈α∞the␈α∂subject.␈α∞ However,␈α∂expressing␈α∞common␈α∂sense␈α∂knowledge␈α∞in

␈↓ α∧␈↓formal␈α
terms␈α
has␈α
proved␈α
very␈α
difficult,␈α
and␈α
the␈α
number␈α
of␈α
scientists␈α
working␈α
in␈α
the␈α
area␈α
is␈αstill␈α
far

␈↓ α∧␈↓too small.

␈↓ α∧␈↓␈↓ αTOne␈α⊂of␈α⊂the␈α⊂best␈α⊂known␈α⊂expert␈α∂systems␈α⊂is␈α⊂Mycin␈α⊂(Shortliffe␈α⊂1976;␈α⊂Davis,␈α⊂Buchanan␈α∂and

␈↓ α∧␈↓Shortliffe␈α
1977),␈α∞a␈α
program␈α
for␈α∞advising␈α
physicians␈α∞on␈α
treating␈α
bacterial␈α∞infections␈α
of␈α∞the␈α
blood
␈↓ α∧␈↓␈↓ u2


␈↓ α∧␈↓and␈α∞meningitis.␈α∞ It␈α∞does␈α∞reasonably␈α∞well␈α∞without␈α∞common␈α∞sense,␈α∞provided␈α∞the␈α∞user␈α∞has␈α∞common

␈↓ α∧␈↓sense and understands the program's limitations.

␈↓ α∧␈↓␈↓ αTMycin␈α∞conducts␈α
a␈α∞question␈α∞and␈α
answer␈α∞dialog.␈α
 After␈α∞asking␈α∞basic␈α
facts␈α∞about␈α∞the␈α
patient

␈↓ α∧␈↓such␈α∂as␈α∂name,␈α∂sex␈α∂and␈α∂age,␈α∂Mycin␈α∂asks␈α∂about␈α∂suspected␈α∂bacterial␈α∂organisms,␈α∂suspected␈α⊂sites␈α∂of

␈↓ α∧␈↓infection,␈α∪the␈α∪presence␈α∪of␈α∩specific␈α∪symptoms␈α∪(e.g.␈α∪fever,␈α∩headache)␈α∪relevant␈α∪to␈α∪diagnosis,␈α∩the

␈↓ α∧␈↓outcome␈αof␈αlaboratory␈αtests,␈α
and␈αsome␈αothers.␈α It␈αthen␈α
recommends␈αa␈αcertain␈αcourse␈α
of␈αantibiotics.

␈↓ α∧␈↓While␈α⊂the␈α∂dialog␈α⊂is␈α∂in␈α⊂English,␈α∂Mycin␈α⊂avoids␈α∂having␈α⊂to␈α∂understand␈α⊂freely␈α∂written␈α⊂English␈α∂by

␈↓ α∧␈↓controlling␈α⊂the␈α⊂dialog.␈α⊃ It␈α⊂outputs␈α⊂sentences,␈α⊃but␈α⊂the␈α⊂user␈α⊂types␈α⊃only␈α⊂single␈α⊂words␈α⊃or␈α⊂standard

␈↓ α∧␈↓phrases.␈α Its␈α
major␈αinnovations␈α
over␈αmany␈α
previous␈αexpert␈αsystems␈α
were␈αthat␈α
it␈αuses␈α
measures␈αof

␈↓ α∧␈↓uncertainty␈α⊂(not␈α⊂probabilities)␈α∂for␈α⊂its␈α⊂diagnoses␈α⊂and␈α∂the␈α⊂fact␈α⊂that␈α⊂it␈α∂is␈α⊂prepared␈α⊂to␈α⊂explain␈α∂its

␈↓ α∧␈↓reasoning to the physician, so he can decide whether to accept it.

␈↓ α∧␈↓␈↓ αTOur␈αdiscussion␈αof␈αMycin␈αbegins␈αwith␈αits␈αontology.␈α The␈αontology␈αof␈αa␈αprogram␈αis␈αthe␈αset␈αof

␈↓ α∧␈↓entities that its variables range over.  Essentially this is what it can have information about.

␈↓ α∧␈↓␈↓ αTMycin's␈α
ontology␈α
includes␈αbacteria,␈α
symptoms,␈α
tests,␈α
possible␈αsites␈α
of␈α
infection,␈αantibiotics␈α
and

␈↓ α∧␈↓treatments.␈α
 Doctors,␈αhospitals,␈α
illness␈α
and␈αdeath␈α
are␈α
absent.␈α Even␈α
patients␈α
are␈αnot␈α
really␈α
part␈αof

␈↓ α∧␈↓the␈α⊂ontology,␈α⊂although␈α∂Mycin␈α⊂asks␈α⊂for␈α⊂many␈α∂facts␈α⊂about␈α⊂the␈α∂specific␈α⊂patient.␈α⊂ This␈α⊂is␈α∂because

␈↓ α∧␈↓patients␈α⊂aren't␈α⊂values␈α⊂of␈α⊃variables,␈α⊂and␈α⊂Mycin␈α⊂never␈α⊃compares␈α⊂the␈α⊂infections␈α⊂of␈α⊃two␈α⊂different

␈↓ α∧␈↓patients.  It would therefore be difficult to modify Mycin to learn from its experience.

␈↓ α∧␈↓␈↓ αTMycin's␈α∩program,␈α∩written␈α∩in␈α∩a␈α∩general␈α⊃scheme␈α∩called␈α∩Emycin,␈α∩is␈α∩a␈α∩so-called␈α⊃␈↓↓production

␈↓ α∧␈↓↓system␈↓.␈α A␈αproduction␈αsystem␈αis␈αa␈αcollection␈αof␈αrules,␈αeach␈αof␈αwhich␈αhas␈αtwo␈αparts␈α-␈αa␈αpattern␈αpart

␈↓ α∧␈↓and␈α
an␈αaction␈α
part.␈α
 When␈αa␈α
rule␈α
is␈αactivated,␈α
Mycin␈αtests␈α
whether␈α
the␈αpattern␈α
part␈α
matches␈αthe

␈↓ α∧␈↓database.␈α If␈αso␈αthis␈αresults␈αin␈αthe␈αvariables␈αin␈αthe␈αpattern␈αbeing␈αmatched␈αto␈αwhatever␈αentities␈αare

␈↓ α∧␈↓required␈αfor␈α
the␈αmatch␈αof␈α
the␈αdatabase.␈α If␈α
not␈αthe␈αpattern␈α
fails␈αand␈αMycin␈α
tries␈αanother.␈α
 If␈αthe

␈↓ α∧␈↓match␈α∞is␈α
successful,␈α∞then␈α
Mycin␈α∞performs␈α
the␈α∞action␈α
part␈α∞of␈α
the␈α∞pattern␈α
using␈α∞the␈α
values␈α∞of␈α
the
␈↓ α∧␈↓␈↓ u3


␈↓ α∧␈↓variables␈α
determined␈αby␈α
the␈α
pattern␈αpart.␈α
 The␈αwhole␈α
process␈α
of␈αquestioning␈α
and␈αrecommending␈α
is

␈↓ α∧␈↓built up out of productions.

␈↓ α∧␈↓␈↓ αTThe␈α∂production␈α⊂formalism␈α∂turned␈α⊂out␈α∂to␈α⊂be␈α∂suitable␈α⊂for␈α∂representing␈α⊂a␈α∂large␈α⊂amount␈α∂of

␈↓ α∧␈↓information␈αabout␈αthe␈αdiagnosis␈αand␈αtreatment␈αof␈αbacterial␈αinfections.␈α When␈αMycin␈αis␈αused␈αin␈αits

␈↓ α∧␈↓intended␈αmanner␈αit␈αscores␈αbetter␈αthan␈αmedical␈αstudents␈αor␈αinterns␈αor␈αpracticing␈αphysicians␈αand␈αon

␈↓ α∧␈↓a␈α∞par␈α∞with␈α∞experts␈α∞in␈α∞bacterial␈α∞diseases␈α∞when␈α
the␈α∞latter␈α∞are␈α∞asked␈α∞to␈α∞perform␈α∞in␈α∞the␈α∞same␈α
way.

␈↓ α∧␈↓However,␈α
Mycin␈α
has␈α
not␈α
been␈α
put␈α
into␈α
production␈α
use,␈α
and␈α
the␈α
reasons␈α
given␈α
by␈α
experts␈α∞in␈α
the

␈↓ α∧␈↓area␈α⊂varied␈α⊃when␈α⊂I␈α⊂asked␈α⊃whether␈α⊂it␈α⊂would␈α⊃be␈α⊂appropriate␈α⊂to␈α⊃sell␈α⊂Mycin␈α⊂cassettes␈α⊃to␈α⊂doctors

␈↓ α∧␈↓wanting␈αto␈α
put␈αit␈α
on␈αtheir␈α
micro-computers.␈α Some␈α
said␈αit␈αwould␈α
be␈αok␈α
if␈αthere␈α
were␈αa␈α
means␈αof

␈↓ α∧␈↓keeping␈α∩Mycin's␈α∩database␈α∩current␈α⊃with␈α∩new␈α∩discoveries␈α∩in␈α∩the␈α⊃field,␈α∩i.e.␈α∩with␈α∩new␈α∩tests,␈α⊃new

␈↓ α∧␈↓theories,␈α∞new␈α∞diagnoses␈α∞and␈α∞new␈α∞antibiotics.␈α∞ For␈α∞example,␈α∞Mycin␈α∞would␈α∞have␈α∞to␈α∞be␈α∞told␈α∞about

␈↓ α∧␈↓Legionnaire's␈αdisease␈αand␈αthe␈αassociated␈α␈↓↓Legionnella␈↓␈αbacteria␈αwhich␈αbecame␈αunderstood␈αonly␈αafter

␈↓ α∧␈↓Mycin␈α→was␈α→finished.␈α~ (Mycin␈α→is␈α→very␈α→stubborn␈α~about␈α→new␈α→bacteria,␈α→and␈α~simply␈α→replies

␈↓ α∧␈↓"unrecognized response")

␈↓ α∧␈↓␈↓ αTOthers␈αsay␈α
that␈αMycin␈α
is␈αnot␈α
even␈αclose␈αto␈α
usable␈αexcept␈α
experimentally,␈αbecause␈α
it␈αdoesn't

␈↓ α∧␈↓know␈αits␈α
own␈αlimitations.␈α I␈α
suppose␈αthis␈αis␈α
partly␈αa␈αquestion␈α
of␈αwhether␈αthe␈α
doctor␈αusing␈αMycin␈α
is

␈↓ α∧␈↓trusted␈α
to␈αunderstand␈α
the␈αdocumentation␈α
about␈αits␈α
limitations.␈α Programmer's␈α
always␈α
develop␈αthe

␈↓ α∧␈↓idea␈α
that␈αthe␈α
users␈α
of␈αtheir␈α
programs␈α
are␈αidiots,␈α
so␈αthe␈α
opinion␈α
that␈αdoctors␈α
aren't␈α
smart␈αenough

␈↓ α∧␈↓not to be misled by Mycin's limitations may be at least partly a consequence of this ideology.

␈↓ α∧␈↓␈↓ αTAn␈αexample␈αof␈α
Mycin␈αnot␈αknowing␈αits␈α
limitations␈αcan␈αbe␈αexcited␈α
by␈αtelling␈αMycin␈α
that␈αthe

␈↓ α∧␈↓patient␈α⊃has␈α⊃␈↓↓Cholerae␈α∩Vibrio␈↓␈α⊃in␈α⊃his␈α∩intestines.␈α⊃ Mycin␈α⊃will␈α∩cheerfully␈α⊃recommend␈α⊃two␈α∩weeks␈α⊃of

␈↓ α∧␈↓tetracycline␈αand␈αnothing␈αelse.␈α Presumably␈αthis␈αwould␈αindeed␈αkill␈αthe␈αbacteria,␈αbut␈αmost␈αlikely␈αthe

␈↓ α∧␈↓patient␈α
will␈α
be␈α
dead␈α
of␈α
cholera␈α∞long␈α
before␈α
that.␈α
 However,␈α
the␈α
physician␈α
will␈α∞presumably␈α
know

␈↓ α∧␈↓that the diarrhea has to be treated and look elsewhere for how to do it.
␈↓ α∧␈↓␈↓ u4


␈↓ α∧␈↓␈↓ αTOn␈α
the␈αother␈α
hand␈α
it␈αmay␈α
be␈α
really␈αtrue␈α
that␈α
some␈αmeasure␈α
of␈α
common␈αsense␈α
is␈αrequired␈α
for

␈↓ α∧␈↓usefulness␈α∞even␈α∞in␈α∞this␈α∂narrow␈α∞domain.␈α∞ We'll␈α∞list␈α∞some␈α∂areas␈α∞of␈α∞common␈α∞sense␈α∂knowledge␈α∞and

␈↓ α∧␈↓reasoning␈αability␈αand␈αalso␈αapply␈αthe␈αcriteria␈αto␈αMycin␈αand␈αother␈αhypothetical␈αprograms␈αoperating

␈↓ α∧␈↓in Mycin's domain.

␈↓ α∧␈↓WHAT IS COMMON SENSE?

␈↓ α∧␈↓␈↓ αTUnderstanding␈α⊗common␈α∃sense␈α⊗capability␈α∃is␈α⊗now␈α∃a␈α⊗hot␈α∃area␈α⊗of␈α∃research␈α⊗in␈α∃artificial

␈↓ α∧␈↓intelligence,␈α
but␈α
there␈α
is␈α
not␈α
yet␈α
any␈αconsensus.␈α
 We␈α
will␈α
try␈α
to␈α
divide␈α
common␈α
sense␈αcapability␈α
into

␈↓ α∧␈↓common␈α⊃sense␈α∩knowledge␈α⊃and␈α⊃common␈α∩sense␈α⊃reasoning,␈α∩but␈α⊃even␈α⊃this␈α∩cannot␈α⊃be␈α∩made␈α⊃firm.

␈↓ α∧␈↓Namely,␈αwhat␈αone␈αman␈αbuilds␈αas␈αa␈αreasoning␈αmethod␈αinto␈αhis␈αprogram,␈αanother␈αcan␈αexpress␈αas␈αa

␈↓ α∧␈↓fact␈α
using␈αa␈α
richer␈α
ontology.␈α However,␈α
the␈α
latter␈αcan␈α
have␈αproblems␈α
in␈α
handling␈αin␈α
a␈α
good␈αway

␈↓ α∧␈↓the generality he has introduced.

␈↓ α∧␈↓COMMON SENSE KNOWLEDGE

␈↓ α∧␈↓␈↓ αTWe shall discuss various areas of common sense knowledge.

␈↓ α∧␈↓␈↓ αT1.␈α
The␈α
most␈α
salient␈α
common␈α
sense␈α
knowledge␈α
concerns␈α
situations␈α
that␈α
change␈α
in␈α
time␈α∞as␈α
a

␈↓ α∧␈↓result␈αof␈α
events.␈α The␈α
most␈αimportant␈α
events␈αare␈α
actions,␈αand␈α
for␈αa␈α
program␈αto␈α
plan␈αintelligently,␈α
it

␈↓ α∧␈↓must be able to determine the effects of its own actions.

␈↓ α∧␈↓␈↓ αTConsider␈α
the␈αMycin␈α
domain␈αas␈α
an␈αexample.␈α
 The␈αsituation␈α
with␈αwhich␈α
Mycin␈αdeals␈α
includes

␈↓ α∧␈↓the␈α∩doctor,␈α∩the␈α∩patient␈α∩and␈α∩the␈α∩illness.␈α∪ Since␈α∩Mycin's␈α∩actions␈α∩are␈α∩advice␈α∩to␈α∩the␈α∪doctor,␈α∩full

␈↓ α∧␈↓planning␈α⊂would␈α∂have␈α⊂to␈α∂include␈α⊂information␈α⊂about␈α∂the␈α⊂effects␈α∂of␈α⊂Mycin's␈α∂output␈α⊂on␈α⊂what␈α∂the

␈↓ α∧␈↓doctor␈αwill␈αdo.␈α Since␈αMycin␈αdoesn't␈αknow␈αabout␈α
the␈αdoctor,␈αit␈αmight␈αplan␈αthe␈αeffects␈αof␈αthe␈α
course

␈↓ α∧␈↓of␈α
treatment␈α
on␈α
the␈α
patient.␈α
 However,␈α
it␈α
doesn't␈α
do␈α
this␈α
either.␈α
 Its␈α
rules␈α
give␈α∞the␈α
recommended

␈↓ α∧␈↓treatment␈α∪as␈α∩a␈α∪function␈α∩of␈α∪the␈α∩information␈α∪elicited␈α∩about␈α∪the␈α∩patient,␈α∪but␈α∩Mycin␈α∪makes␈α∩no

␈↓ α∧␈↓prognosis␈α∞of␈α
the␈α∞effects␈α∞of␈α
the␈α∞treatment.␈α∞ Of␈α
course,␈α∞the␈α∞doctors␈α
who␈α∞provided␈α∞the␈α
information

␈↓ α∧␈↓built into Mycin considered the effects of the treatments.
␈↓ α∧␈↓␈↓ u5


␈↓ α∧␈↓␈↓ αTIgnoring␈α∩prognosis␈α∩is␈α∩possible␈α∩because␈α∩of␈α⊃the␈α∩specific␈α∩narrow␈α∩domain␈α∩in␈α∩which␈α⊃Mycin

␈↓ α∧␈↓operates.␈α
 Suppose,␈α
for␈α∞example,␈α
a␈α
certain␈α∞antibiotic␈α
had␈α
the␈α∞precondition␈α
for␈α
its␈α∞usefulness␈α
that

␈↓ α∧␈↓the␈α∂patient␈α∂not␈α∂have␈α∂a␈α∂fever.␈α∞ Then␈α∂Mycin␈α∂might␈α∂have␈α∂to␈α∂make␈α∞a␈α∂plan␈α∂for␈α∂getting␈α∂rid␈α∂of␈α∞the

␈↓ α∧␈↓patient's␈α
fever␈α
and␈α
verifying␈α
that␈α∞it␈α
was␈α
gone␈α
as␈α
a␈α
part␈α∞of␈α
the␈α
plan␈α
for␈α
using␈α
the␈α∞antibiotic.␈α
 In

␈↓ α∧␈↓other␈α∂domains,␈α∂expert␈α∞systems␈α∂and␈α∂other␈α∂AI␈α∞programs␈α∂have␈α∂to␈α∞make␈α∂plans,␈α∂but␈α∂Mycin␈α∞doesn't.

␈↓ α∧␈↓Perhaps␈αif␈αI␈αknew␈αmore␈αabout␈αbacterial␈αdiseases,␈αI␈αwould␈αconclude␈αthat␈αtheir␈αtreatment␈αsometimes

␈↓ α∧␈↓really does require planning and that lack of planning ability limits Mycin's utility.

␈↓ α∧␈↓␈↓ αTThe␈αfact␈αthat␈αMycin␈αdoesn't␈αgive␈αa␈αprognosis␈αis␈αcertainly␈αa␈αlimitation.␈α For␈αexample,␈αMycin

␈↓ α∧␈↓cannot␈αbe␈αasked␈αon␈α
behalf␈αof␈αthe␈αpatient␈α
or␈αthe␈αadministration␈αof␈α
the␈αhospital␈αwhen␈αthe␈αpatient␈α
is

␈↓ α∧␈↓likely␈α
to␈αbe␈α
ready␈αto␈α
go␈αhome.␈α
 The␈αdoctor␈α
who␈αuses␈α
Mycin␈αmust␈α
do␈αthat␈α
part␈αof␈α
the␈αwork␈α
himself.

␈↓ α∧␈↓Moreover,␈α
Mycin␈α
cannot␈α
answer␈α
a␈α
question␈α
about␈α
a␈α
hypothetical␈α
treatment,␈α
e.g.␈α
"What␈αwill␈α
happen

␈↓ α∧␈↓if␈α
I␈α
give␈α∞this␈α
patient␈α
penicillin?"␈α∞or␈α
even␈α
"What␈α
bad␈α∞things␈α
might␈α
happen␈α∞if␈α
I␈α
give␈α∞this␈α
patient

␈↓ α∧␈↓penicillin?".

␈↓ α∧␈↓␈↓ αT2.␈α⊂Various␈α⊂formalisms␈α⊂are␈α⊂used␈α⊂in␈α⊂artificial␈α⊂intelligence␈α⊂for␈α⊂representing␈α⊂facts␈α⊂about␈α⊂the

␈↓ α∧␈↓effects␈αof␈αactions␈αand␈αother␈αevents.␈α However,␈αall␈αsystems␈αthat␈αI␈αknow␈αabout␈αgive␈αthe␈αeffects␈αof␈αan

␈↓ α∧␈↓event␈α⊂in␈α⊂a␈α⊂situation␈α⊂by␈α⊂describing␈α⊂a␈α⊂new␈α∂situation␈α⊂that␈α⊂results␈α⊂from␈α⊂the␈α⊂event.␈α⊂ This␈α⊂is␈α∂often

␈↓ α∧␈↓enough,␈αbut␈αit␈αdoesn't␈αcover␈αthe␈αimportant␈αcase␈αof␈αconcurrent␈αevents␈αand␈αactions.␈α For␈αexample,␈αif

␈↓ α∧␈↓a␈α⊃patient␈α⊃has␈α∩cholera,␈α⊃while␈α⊃the␈α⊃antibiotic␈α∩is␈α⊃killing␈α⊃the␈α⊃cholera␈α∩bacteria,␈α⊃the␈α⊃damage␈α∩to␈α⊃his

␈↓ α∧␈↓intestines␈α⊂is␈α∂causing␈α⊂loss␈α∂of␈α⊂fluids␈α∂that␈α⊂are␈α⊂likely␈α∂to␈α⊂be␈α∂fatal.␈α⊂ Inventing␈α∂a␈α⊂formalism␈α⊂that␈α∂will

␈↓ α∧␈↓conveniently␈α∀express␈α∀people's␈α∪common␈α∀sense␈α∀knowledge␈α∀about␈α∪concurrent␈α∀events␈α∀is␈α∀a␈α∪major

␈↓ α∧␈↓unsolved problem of AI.

␈↓ α∧␈↓␈↓ αT3.␈α∞The␈α∞world␈α∞is␈α∞extended␈α
in␈α∞space␈α∞and␈α∞is␈α∞occupied␈α
by␈α∞objects␈α∞that␈α∞change␈α∞their␈α
positions

␈↓ α∧␈↓and␈α∂are␈α∂sometimes␈α⊂created␈α∂and␈α∂destroyed.␈α⊂ The␈α∂common␈α∂sense␈α∂facts␈α⊂about␈α∂this␈α∂are␈α⊂difficult␈α∂to

␈↓ α∧␈↓express␈αbut␈αare␈αprobably␈αnot␈αimportant␈αin␈αthe␈αMycin␈αexample.␈α A␈αmajor␈αdifficulty␈αis␈αin␈αhandling
␈↓ α∧␈↓␈↓ u6


␈↓ α∧␈↓the␈αkind␈αof␈αpartial␈αknowledge␈αpeople␈αordinarily␈αhave.␈α
 I␈αcan␈αsee␈αpart␈αof␈αthe␈αfront␈αof␈αa␈α
person␈αin

␈↓ α∧␈↓the␈α∂audience,␈α∂and␈α⊂my␈α∂idea␈α∂of␈α∂his␈α⊂shape␈α∂uses␈α∂this␈α∂information␈α⊂to␈α∂approximate␈α∂his␈α⊂total␈α∂shape.

␈↓ α∧␈↓Thus␈α∞I␈α∞don't␈α∞expect␈α∂him␈α∞to␈α∞stick␈α∞out␈α∞two␈α∂feet␈α∞in␈α∞back␈α∞even␈α∞though␈α∂I␈α∞can't␈α∞see␈α∞that␈α∂he␈α∞doesn't.

␈↓ α∧␈↓However, my idea of the shape of his back is less definite than that of the parts I can see.

␈↓ α∧␈↓␈↓ αT4.␈α∩The␈α∩ability␈α∩to␈α∪represent␈α∩and␈α∩use␈α∩knowledge␈α∪about␈α∩knowledge␈α∩is␈α∩often␈α∪required␈α∩for

␈↓ α∧␈↓intelligent␈α
behavior.␈α∞ What␈α
airline␈α∞flights␈α
there␈α
are␈α∞to␈α
Singapore␈α∞is␈α
recorded␈α
in␈α∞the␈α
issue␈α∞of␈α
the

␈↓ α∧␈↓International␈α
Airline␈α
Guide␈αcurrent␈α
for␈α
the␈αproposed␈α
flight␈α
day.␈α
 Travel␈αagents␈α
know␈α
how␈αto␈α
book

␈↓ α∧␈↓airline␈α
flights␈α
and␈α
can␈α∞compute␈α
what␈α
they␈α
cost.␈α
 An␈α∞advanced␈α
Mycin␈α
might␈α
need␈α
to␈α∞reason␈α
that

␈↓ α∧␈↓Dr. Smith knows about cholera, because he is a specialist in tropical medicine.

␈↓ α∧␈↓␈↓ αT5.␈αA␈αprogram␈αthat␈αmust␈αco-operate␈αor␈αcompete␈αwith␈αpeople␈αor␈αother␈αprograms␈αmust␈αbe␈αable

␈↓ α∧␈↓to␈α∞represent␈α∂information␈α∞about␈α∞their␈α∂knowledge,␈α∞beliefs,␈α∞goals,␈α∂likes␈α∞and␈α∞dislikes,␈α∂intentions␈α∞and

␈↓ α∧␈↓abilities.␈α∩ An␈α∩advanced␈α⊃Mycin␈α∩might␈α∩need␈α∩to␈α⊃know␈α∩that␈α∩a␈α∩patient␈α⊃won't␈α∩take␈α∩a␈α∩bad␈α⊃tasting

␈↓ α∧␈↓medicine unless he is convinced of its necessity.

␈↓ α∧␈↓␈↓ αT6.␈α∩Common␈α∪sense␈α∩includes␈α∪much␈α∩knowledge␈α∩whose␈α∪domain␈α∩overlaps␈α∪that␈α∩of␈α∪the␈α∩exact

␈↓ α∧␈↓sciences␈α∂but␈α∞differs␈α∂from␈α∂it␈α∞epistemologically.␈α∂ For␈α∂example,␈α∞if␈α∂I␈α∞spill␈α∂the␈α∂glass␈α∞of␈α∂water␈α∂on␈α∞the

␈↓ α∧␈↓podium,␈αeveryone␈α
knows␈αthat␈αthe␈α
glass␈αwill␈αbreak␈α
and␈αthe␈αwater␈α
will␈αspill.␈α Everyone␈α
knows␈αthat

␈↓ α∧␈↓this␈α
will␈α
take␈α
a␈α
fraction␈α
of␈α
a␈α
second␈α
and␈α
that␈α
the␈α
water␈α
will␈α
not␈α
splash␈α
even␈α
ten␈α
feet.␈α However,

␈↓ α∧␈↓this␈α∞information␈α∞is␈α∂not␈α∞obtained␈α∞by␈α∞using␈α∂the␈α∞formula␈α∞for␈α∞a␈α∂falling␈α∞body␈α∞or␈α∂the␈α∞Navier-Stokes

␈↓ α∧␈↓equations␈αgoverning␈αfluid␈αflow.␈α We␈αdon't␈αhave␈αthe␈α
input␈αdata␈αfor␈αthe␈αequations,␈αmost␈αof␈αus␈α
don't

␈↓ α∧␈↓know␈α
them,␈α
and␈αwe␈α
couldn't␈α
integrate␈αthem␈α
fast␈α
enough␈α
to␈αdecide␈α
whether␈α
to␈αjump␈α
out␈α
of␈αthe␈α
way.

␈↓ α∧␈↓This␈α⊃common␈α⊃sense␈α⊃physics␈α⊃is␈α⊃contiguous␈α⊃with␈α⊃scientific␈α⊃physics.␈α⊃ In␈α⊃fact␈α⊃scientific␈α∩physics␈α⊃is

␈↓ α∧␈↓imbedded␈α∂in␈α∂common␈α∂sense␈α∂physics,␈α∂because␈α∂it␈α∂is␈α∂common␈α∂sense␈α∂physics␈α∂that␈α∂tells␈α∂us␈α⊂what␈α∂the

␈↓ α∧␈↓equation

␈↓ α∧␈↓␈↓ αT␈↓↓s = 1/2 g t␈↓∧2␈↓
␈↓ α∧␈↓␈↓ u7


␈↓ α∧␈↓means.␈α∞ If␈α∞Mycin␈α∞were␈α∞extended␈α∞to␈α∞be␈α∞a␈α∞robot␈α∞physician␈α∞it␈α∞would␈α∞have␈α∞to␈α∞know␈α∞common␈α
sense

␈↓ α∧␈↓physics and maybe also some scientific physics.

␈↓ α∧␈↓␈↓ αTIt␈α
is␈α
doubtful␈α∞that␈α
the␈α
facts␈α
of␈α∞the␈α
common␈α
sense␈α
world␈α∞can␈α
be␈α
represented␈α∞adequately␈α
by

␈↓ α∧␈↓production␈αrules.␈α Consider␈αthe␈αfact␈αthat␈αwhen␈αtwo␈αobjects␈αcollide␈αthey␈αoften␈αmake␈αa␈α
noise.␈α This

␈↓ α∧␈↓fact␈α
can␈α
be␈α
used␈α
to␈α
make␈αa␈α
noise,␈α
to␈α
avoid␈α
making␈α
a␈α
noise,␈αto␈α
explain␈α
a␈α
noise␈α
or␈α
to␈α
explain␈αthe

␈↓ α∧␈↓absence␈α⊃of␈α∩a␈α⊃noise.␈α∩ It␈α⊃can␈α∩also␈α⊃be␈α∩used␈α⊃in␈α∩specific␈α⊃situations␈α∩involving␈α⊃a␈α∩noise␈α⊃but␈α∩also␈α⊃to

␈↓ α∧␈↓understand␈α∞general␈α∞phenomena,␈α
e.g.␈α∞should␈α∞an␈α∞intruder␈α
step␈α∞on␈α∞the␈α∞gravel,␈α
the␈α∞dog␈α∞will␈α∞hear␈α
it

␈↓ α∧␈↓and␈αbark.␈α A␈αproduction␈αrule␈αembodies␈αa␈αfact␈α
only␈αas␈αpart␈αof␈αa␈αspecific␈αprocedure.␈α Typically␈α
they

␈↓ α∧␈↓match␈αfacts␈α
about␈αspecific␈α
objects,␈αe.g.␈α
a␈αspecific␈αbacterium,␈α
against␈αa␈α
general␈αrule␈α
and␈αget␈α
a␈αnew

␈↓ α∧␈↓fact about those objects.

␈↓ α∧␈↓␈↓ αTMuch␈α
present␈αAI␈α
research␈α
concerns␈αhow␈α
to␈αrepresent␈α
facts␈α
in␈αways␈α
that␈αpermit␈α
them␈α
to␈αbe

␈↓ α∧␈↓used for a wide variety of purposes.

␈↓ α∧␈↓COMMON SENSE REASONING

␈↓ α∧␈↓␈↓ αTOur␈α∂ability␈α∞to␈α∂use␈α∞common␈α∂sense␈α∞knowledge␈α∂depends␈α∞on␈α∂being␈α∞able␈α∂to␈α∞do␈α∂common␈α∞sense

␈↓ α∧␈↓reasoning.

␈↓ α∧␈↓␈↓ αTMuch␈αartificial␈αintelligence␈αinference␈αis␈αnot␈αdesigned␈αto␈αuse␈αdirectly␈αthe␈αrules␈αof␈αinference␈α
of

␈↓ α∧␈↓any␈α
of␈α
the␈α∞well␈α
known␈α
systems␈α
of␈α∞mathematical␈α
logic.␈α
 There␈α
is␈α∞often␈α
no␈α
clear␈α
separation␈α∞in␈α
the

␈↓ α∧␈↓program␈α∀between␈α∀determining␈α∪what␈α∀inferences␈α∀are␈α∪correct␈α∀and␈α∀the␈α∪strategy␈α∀for␈α∀finding␈α∪the

␈↓ α∧␈↓inferences␈α∀required␈α∀to␈α∀solve␈α∀the␈α∀problem␈α∀at␈α∀hand.␈α∀ Nevertheless,␈α∀the␈α∀logical␈α∀system␈α∪usually

␈↓ α∧␈↓corresponds␈αto␈αa␈αsubset␈αof␈αfirst␈αorder␈αlogic.␈α
 Systems␈αprovide␈αfor␈αinferring␈αa␈αfact␈αabout␈αone␈αor␈α
two

␈↓ α∧␈↓particular␈α∂objects␈α∞from␈α∂other␈α∞facts␈α∂about␈α∞these␈α∂objects␈α∞and␈α∂a␈α∞general␈α∂rule␈α∂containing␈α∞variables.

␈↓ α∧␈↓Most expert systems, including Mycin, never infer general statements, i.e. quantified formulas.

␈↓ α∧␈↓␈↓ αTHuman␈αreasoning␈αalso␈α
involves␈αobtaining␈αfacts␈αby␈α
observation␈αof␈αthe␈αworld,␈α
and␈αcomputer

␈↓ α∧␈↓programs␈α
also␈α
do␈α
this.␈α
 Robert␈α
Filman␈α
did␈α
an␈α
interesting␈α
thesis␈α
on␈α
observation␈α
in␈α
a␈α
chess␈α
world
␈↓ α∧␈↓␈↓ u8


␈↓ α∧␈↓where␈α⊃many␈α∩facts␈α⊃that␈α∩could␈α⊃be␈α∩obtained␈α⊃by␈α⊃deduction␈α∩are␈α⊃in␈α∩fact␈α⊃obtained␈α∩by␈α⊃observation.

␈↓ α∧␈↓Mycin's␈αdoesn't␈αrequire␈αthis,␈α
but␈αour␈αhypothetical␈αrobot␈α
physician␈αwould␈αhave␈αto␈αdraw␈α
conclusions

␈↓ α∧␈↓from a patient's appearance, and computer vision is not ready for it.

␈↓ α∧␈↓␈↓ αTAn␈αimportant␈αnew␈αdevelopment␈αin␈αAI␈α(since␈αthe␈αmiddle␈α1970s)␈αis␈αthe␈αformalization␈αof␈αnon-

␈↓ α∧␈↓monotonic reasoning.

␈↓ α∧␈↓␈↓ αTDeductive␈αreasoning␈αin␈αmathematical␈αlogic␈αhas␈αthe␈αfollowing␈αproperty␈α-␈αcalled␈αmonotonicity

␈↓ α∧␈↓by␈α⊃analogy␈α⊂with␈α⊃similar␈α⊃mathematical␈α⊂concepts.␈α⊃ Suppose␈α⊂we␈α⊃have␈α⊃a␈α⊂set␈α⊃of␈α⊃assumptions␈α⊂from

␈↓ α∧␈↓which␈α∞follow␈α∞certain␈α∞conclusions.␈α∞ Now␈α∞suppose␈α∞we␈α∞add␈α∞additional␈α∞assumptions.␈α∞ There␈α∞may␈α∞be

␈↓ α∧␈↓some␈α∩new␈α∪conclusions,␈α∩but␈α∩every␈α∪sentence␈α∩that␈α∪was␈α∩a␈α∩deductive␈α∪consequence␈α∩of␈α∪the␈α∩original

␈↓ α∧␈↓hypotheses is still a consequence of the enlarged set.

␈↓ α∧␈↓␈↓ αTOrdinary␈αhuman␈α
reasoning␈αdoes␈αnot␈α
share␈αthis␈α
monotonicity␈αproperty.␈α If␈α
you␈αknow␈α
that␈αI

␈↓ α∧␈↓have␈αa␈αcar,␈αyou␈αmay␈αconclude␈αthat␈αit␈αis␈αa␈αgood␈αidea␈αto␈αask␈αme␈αfor␈αa␈αride.␈α If␈αyou␈αthen␈αlearn␈αthat

␈↓ α∧␈↓my␈αcar␈αis␈αbeing␈αfixed␈α(which␈αdoes␈αnot␈αcontradict␈αwhat␈αyou␈αknew␈αbefore),␈αyou␈αno␈α
longer␈αconclude

␈↓ α∧␈↓that␈α
you␈α
can␈α
get␈α∞a␈α
ride.␈α
 If␈α
you␈α∞now␈α
learn␈α
that␈α
the␈α∞car␈α
will␈α
be␈α
out␈α∞in␈α
half␈α
an␈α
hour␈α∞you␈α
reverse

␈↓ α∧␈↓yourself again.

␈↓ α∧␈↓␈↓ αTSeveral␈α
artificial␈α
intelligence␈αresearchers,␈α
for␈α
example␈αMarvin␈α
Minsky␈α
(1974)␈α
have␈αpointed

␈↓ α∧␈↓out␈α∞that␈α∞intelligent␈α
computer␈α∞programs␈α∞will␈α∞have␈α
to␈α∞reason␈α∞non-monotonically.␈α∞ Some␈α
concluded

␈↓ α∧␈↓that therefore logic is not an appropriate formalism.

␈↓ α∧␈↓␈↓ αTHowever,␈α∞it␈α
has␈α∞turned␈α
out␈α∞that␈α
deduction␈α∞in␈α
mathematical␈α∞logic␈α
can␈α∞be␈α∞supplemented␈α
by

␈↓ α∧␈↓additional␈αmodes␈αof␈αnon-monotonic␈αreasoning,␈αwhich␈αare␈αjust␈αas␈αformal␈αas␈αdeduction␈αand␈αjust␈αas

␈↓ α∧␈↓susceptible␈α∀to␈α∪mathematical␈α∀study␈α∀and␈α∪computer␈α∀implementation.␈α∀ Formalized␈α∪non-monotonic

␈↓ α∧␈↓reasoning␈α⊃turns␈α⊂out␈α⊃to␈α⊃give␈α⊂certain␈α⊃rules␈α⊂of␈α⊃conjecture␈α⊃rather␈α⊂than␈α⊃rules␈α⊂of␈α⊃inference␈α⊃-␈α⊂their

␈↓ α∧␈↓conclusion␈α∞are␈α∞appropriate,␈α∞but␈α∞may␈α∞be␈α∂disconfirmed␈α∞when␈α∞more␈α∞facts␈α∞are␈α∞obtained.␈α∂ One␈α∞such

␈↓ α∧␈↓method is ␈↓↓circumscription␈↓, described in (McCarthy 1980).
␈↓ α∧␈↓␈↓ u9


␈↓ α∧␈↓␈↓ αTA␈α
mathematical␈α∞description␈α
of␈α
circumscription␈α∞is␈α
beyond␈α∞the␈α
scope␈α
of␈α∞this␈α
lecture,␈α∞but␈α
the

␈↓ α∧␈↓general␈αidea␈αis␈αstraightforward.␈α
 We␈αhave␈αa␈αproperty␈αapplicable␈α
to␈αobjects␈αor␈αa␈αrelation␈α
applicable

␈↓ α∧␈↓to␈α∂pairs␈α∂or␈α⊂triplets,␈α∂etc.␈α∂of␈α∂objects.␈α⊂ This␈α∂property␈α∂or␈α∂relation␈α⊂is␈α∂constrained␈α∂by␈α⊂some␈α∂sentences

␈↓ α∧␈↓taken␈αas␈αassumptions,␈αbut␈αthere␈αis␈αstill␈αsome␈αfreedom␈αleft.␈α Circumscription␈αfurther␈α
constrains␈αthe

␈↓ α∧␈↓property or relation by requiring it to be true of a minimal set of objects.

␈↓ α∧␈↓␈↓ αTAs␈α∩an␈α∩example,␈α∩consider␈α∩representing␈α∩the␈α∪facts␈α∩about␈α∩whether␈α∩an␈α∩object␈α∩can␈α∩fly␈α∪in␈α∩a

␈↓ α∧␈↓database␈α∩of␈α∩common␈α∩sense␈α∩knowledge.␈α∩ We␈α⊃could␈α∩try␈α∩to␈α∩provide␈α∩axioms␈α∩that␈α∩will␈α⊃determine

␈↓ α∧␈↓whether␈α~each␈α~kind␈α→of␈α~object␈α~can␈α~fly,␈α→but␈α~this␈α~would␈α→make␈α~the␈α~database␈α~very␈α→large.

␈↓ α∧␈↓Circumscription␈α∂allows␈α∂us␈α∂to␈α∂express␈α∂the␈α⊂assumption␈α∂that␈α∂only␈α∂those␈α∂objects␈α∂can␈α∂fly␈α⊂for␈α∂which

␈↓ α∧␈↓there␈α⊂is␈α⊃a␈α⊂positive␈α⊃statement␈α⊂about␈α⊂it.␈α⊃ Thus␈α⊂there␈α⊃will␈α⊂be␈α⊂positive␈α⊃statements␈α⊂that␈α⊃birds␈α⊂and

␈↓ α∧␈↓airplanes␈αcan␈αfly␈α
and␈αno␈αstatement␈α
that␈αcamels␈αcan␈α
fly.␈α Since␈αwe␈α
don't␈αinclude␈αnegative␈α
statements

␈↓ α∧␈↓in␈α⊂the␈α⊂database,␈α⊃we␈α⊂could␈α⊂provide␈α⊂for␈α⊃flying␈α⊂camels,␈α⊂if␈α⊂there␈α⊃were␈α⊂any,␈α⊂by␈α⊃adding␈α⊂statements

␈↓ α∧␈↓without␈αremoving␈αexisting␈αstatements.␈α This␈αmuch␈αis␈αoften␈αdone␈αby␈αa␈αsimpler␈αmethod␈α-␈αthe␈α␈↓↓closed

␈↓ α∧␈↓↓world␈αassumption␈↓␈αdiscussed␈αby␈αRaymond␈αReiter.␈α However,␈αwe␈αalso␈αhave␈αexceptions␈αto␈αthe␈α
general

␈↓ α∧␈↓statement␈α∂that␈α∂birds␈α⊂can␈α∂fly.␈α∂ For␈α⊂example,␈α∂penguins,␈α∂ostriches␈α⊂and␈α∂birds␈α∂with␈α⊂certain␈α∂feathers

␈↓ α∧␈↓removed␈α∀can't␈α∃fly.␈α∀ Moreover,␈α∃more␈α∀exceptions␈α∃may␈α∀be␈α∃found␈α∀and␈α∃even␈α∀exceptions␈α∃to␈α∀the

␈↓ α∧␈↓exceptions.␈α∃ Circumscription␈α⊗allows␈α∃us␈α∃to␈α⊗make␈α∃the␈α∃known␈α⊗exceptions␈α∃and␈α∃to␈α⊗provide␈α∃for

␈↓ α∧␈↓additional exceptions to be added later - again without changing existing statements.

␈↓ α∧␈↓␈↓ αTNon-monotonic␈α
reasoning␈αalso␈α
seems␈α
to␈αbe␈α
involved␈α
in␈αhuman␈α
communication.␈α
 Suppose␈αI

␈↓ α∧␈↓hire␈α∞you␈α∂to␈α∞build␈α∞me␈α∂a␈α∞bird␈α∂cage,␈α∞and␈α∞you␈α∂build␈α∞it␈α∞without␈α∂a␈α∞top,␈α∂and␈α∞I␈α∞refuse␈α∂to␈α∞pay␈α∂on␈α∞the

␈↓ α∧␈↓grounds␈α
that␈α
my␈α
bird␈α
might␈α
fly␈α
away.␈α
 A␈α
judge␈α
will␈α
side␈α
with␈α
me.␈α
 On␈α
the␈α
other␈α∞hand␈α
suppose

␈↓ α∧␈↓you␈αbuild␈αit␈αwith␈αa␈αtop,␈αand␈αI␈αrefuse␈αto␈αpay␈αfull␈αprice␈αon␈αthe␈αgrounds␈αthat␈αmy␈αbird␈αis␈αa␈αpenguin,

␈↓ α∧␈↓and␈αthe␈αtop␈αis␈αa␈αwaste.␈α Unless␈αI␈αtold␈αyou␈αthat␈αmy␈αbird␈αcouldn't␈αfly,␈αthe␈αjudge␈αwill␈αside␈αwith␈αyou.

␈↓ α∧␈↓We␈αcan␈α
therefore␈αregard␈αit␈α
as␈αa␈α
communication␈αconvention␈αthat␈α
if␈αa␈α
bird␈αcan␈αfly␈α
the␈αfact␈αneed␈α
not

␈↓ α∧␈↓be mentioned, but if the bird can't fly and it is relevant, then the fact must be mentioned.
␈↓ α∧␈↓␈↓ f10


␈↓ α∧␈↓References:

␈↓ α∧␈↓␈↓αDavis,␈α⊗Randall;␈α∃Buchanan,␈α⊗Bruce;␈α⊗and␈α∃Shortliffe,␈α⊗Edward␈α⊗(1977)␈↓:␈α∃"Production␈α⊗Rules␈α⊗as␈α∃a
␈↓ α∧␈↓Representation␈α
for␈α
a␈αKnowledge-Based␈α
Consultation␈α
Program,"␈α␈↓↓Artificial␈α
Intelligence␈↓,␈α
Volume␈α8,
␈↓ α∧␈↓Number 1, February.

␈↓ α∧␈↓␈↓αMcCarthy,␈α↔John␈α↔(1960)␈↓:␈α↔"Programs␈α↔with␈α↔Common␈α↔Sense,"␈α↔Proceedings␈α↔of␈α↔the␈α↔Teddington
␈↓ α∧␈↓Conference on the Mechanization of Thought Processes, Her Majesty's Stationery Office, London.

␈↓ α∧␈↓␈↓αMcCarthy,␈α∂John␈α∂and␈α∞P.J.␈α∂Hayes␈α∂(1969)␈↓:␈α∞"Some␈α∂Philosophical␈α∂Problems␈α∞from␈α∂the␈α∂Standpoint␈α∞of
␈↓ α∧␈↓Artificial␈α
Intelligence",␈α∞in␈α
D.␈α
Michie␈α∞(ed),␈α
␈↓↓Machine␈α
Intelligence␈α∞4␈↓,␈α
American␈α
Elsevier,␈α∞New␈α
York,
␈↓ α∧␈↓NY.

␈↓ α∧␈↓␈↓αMcCarthy,␈α⊃John␈α⊃(1980)␈↓:␈α⊃"Circumscription␈α∩-␈α⊃A␈α⊃Form␈α⊃of␈α⊃Non-Monotonic␈α∩Reasoning",␈α⊃␈↓↓Artificial
␈↓ α∧␈↓↓Intelligence␈↓, Volume 13, Numbers 1,2, April.

␈↓ α∧␈↓␈↓αMinsky, Marvin (1974)␈↓: "A Framework for Representing Knowledge", ␈↓↓M.I.T. AI Memo 252␈↓.

␈↓ α∧␈↓␈↓αShortliffe,␈α_Edward␈α↔H.␈α_(1976)␈↓:␈α_Computer-Based␈α↔Medical␈α_Consultations:␈α_MYCIN,␈α↔American
␈↓ α∧␈↓Elsevier, New York, NY.
␈↓ α∧␈↓␈↓ f11


␈↓ α∧␈↓Answers to questions:

␈↓ α∧␈↓1.␈α
I␈α
could␈α
have␈α
made␈αthis␈α
a␈α
defensive␈α
talk␈α
about␈αartificial␈α
intelligence,␈α
but␈α
I␈α
chose␈α
to␈αemphasize

␈↓ α∧␈↓the␈α
problems␈α
that␈α∞have␈α
been␈α
identified␈α
rather␈α∞than␈α
the␈α
progress␈α
that␈α∞has␈α
been␈α
made␈α∞in␈α
solving

␈↓ α∧␈↓them.␈α∞ Let␈α
me␈α∞remind␈α
you␈α∞that␈α
I␈α∞have␈α∞argued␈α
that␈α∞the␈α
need␈α∞for␈α
common␈α∞sense␈α
is␈α∞not␈α∞a␈α
truism.

␈↓ α∧␈↓Many useful things can be done without it, e.g. Mycin and also chess programs.

␈↓ α∧␈↓2.␈αConsider␈αyour␈α20␈αyears.␈α If␈αanyone␈αhad␈αknown␈αin␈α1963␈αhow␈αto␈αmake␈αa␈αprogram␈αlearn␈αfrom␈αits

␈↓ α∧␈↓experience␈α∞to␈α∞do␈α∞what␈α∞a␈α∞human␈α∞does␈α∞after␈α∞20␈α∞years,␈α∞they␈α∞might␈α∞have␈α∞done␈α∞it,␈α∞and␈α∞it␈α∞might␈α∞be

␈↓ α∧␈↓pretty␈α∩smart␈α∪by␈α∩now.␈α∪ Already␈α∩in␈α∩1958␈α∪there␈α∩had␈α∪been␈α∩work␈α∩on␈α∪programs␈α∩that␈α∪learn␈α∩from

␈↓ α∧␈↓experience.␈α
 However,␈α
all␈α
they␈α∞could␈α
learn␈α
was␈α
to␈α
set␈α∞optimal␈α
values␈α
of␈α
numerical␈α∞parameters␈α
in

␈↓ α∧␈↓the␈α∂program,␈α∞and␈α∂they␈α∞were␈α∂quite␈α∞limited␈α∂in␈α∞their␈α∂ability␈α∞to␈α∂do␈α∞that.␈α∂ Arthur␈α∂Samuel's␈α∞checker

␈↓ α∧␈↓program␈α∂learned␈α∂optimal␈α∂values␈α⊂for␈α∂its␈α∂parameters,␈α∂but␈α∂the␈α⊂problem␈α∂was␈α∂that␈α∂certain␈α⊂kinds␈α∂of

␈↓ α∧␈↓desired␈αbehavior␈αdid␈αnot␈αcorrespond␈αto␈αany␈αsetting␈αof␈αthe␈αparameters,␈αbecause␈αit␈αdepended␈αon␈αthe

␈↓ α∧␈↓recognition␈α
of␈α
a␈α
certain␈α
kind␈α
of␈α
strategic␈αsituation.␈α
 Thus␈α
the␈α
first␈α
prerequisite␈α
for␈α
a␈α
program␈αto␈α
be

␈↓ α∧␈↓able␈α⊃to␈α⊃learn␈α⊃something␈α⊃is␈α⊂that␈α⊃it␈α⊃be␈α⊃able␈α⊃to␈α⊂represent␈α⊃internally␈α⊃the␈α⊃desired␈α⊃modification␈α⊂of

␈↓ α∧␈↓behavior.␈α∞ Simple␈α∞changes␈α∞in␈α∞behavior␈α∞must␈α∞have␈α∞simple␈α∞representations.␈α∂ Turing's␈α∞universality

␈↓ α∧␈↓theory␈α∂convinces␈α∂us␈α∂that␈α∂arbitrary␈α∂behaviors␈α∂can␈α∞be␈α∂represented,␈α∂but␈α∂they␈α∂don't␈α∂tell␈α∂us␈α∂how␈α∞to

␈↓ α∧␈↓represent␈αthem␈α
in␈αsuch␈αa␈α
way␈αthat␈αa␈α
small␈αchange␈αin␈α
behavior␈αis␈αa␈α
small␈αchange␈αin␈α
representation.

␈↓ α∧␈↓Present methods of changing programs amount to education by brain surgery.

␈↓ α∧␈↓3.␈α
In␈α
the␈α
use␈α
of␈α
Mycin,␈α
the␈α
physician␈αis␈α
supposed␈α
to␈α
supply␈α
the␈α
common␈α
sense.␈α
 The␈α
question␈αis

␈↓ α∧␈↓whether␈αthe␈αprogram␈αmust␈αalso␈αhave␈αcommon␈αsense,␈αand␈αI␈αwould␈αsay␈αthat␈αthe␈αanswer␈αis␈αnot␈αclear

␈↓ α∧␈↓in␈αthe␈αMycin␈αcase.␈α Purely␈αcomputational␈αprograms␈αdon't␈αrequire␈αcommon␈αsense,␈αand␈αnone␈αof␈αthe

␈↓ α∧␈↓present␈α∞chess␈α∞programs␈α∞have␈α
any.␈α∞ On␈α∞the␈α∞other␈α
hand,␈α∞it␈α∞seems␈α∞clear␈α
that␈α∞many␈α∞other␈α∞kinds␈α
of

␈↓ α∧␈↓programs require common sense to be useful at all.
␈↓ α∧␈↓␈↓ f12


␈↓ α∧␈↓PANEL DISCUSSION

␈↓ α∧␈↓McCarthy:␈α⊃The␈α⊃question␈α⊃is␈α⊃whether␈α⊃AI␈α∩has␈α⊃illuminated␈α⊃human␈α⊃intelligence,␈α⊃and␈α⊃I␈α∩think␈α⊃the

␈↓ α∧␈↓answer␈α⊃is␈α⊃obviously␈α⊃yes.␈α⊂ AI␈α⊃and␈α⊃psychology␈α⊃influenced␈α⊂by␈α⊃AI␈α⊃are␈α⊃responsible␈α⊃for␈α⊂destroying

␈↓ α∧␈↓behaviorism␈α
as␈α
a␈α
serious␈α
approach␈α
to␈α
psychology␈α
and␈α
turning␈α
psychologists␈α∞towards␈α
information

␈↓ α∧␈↓processing␈α⊃models.␈α⊃ Presumably␈α⊃a␈α⊃psychologist␈α⊃would␈α⊃be␈α⊃more␈α⊃competent␈α⊃to␈α⊃speak␈α⊃about␈α⊂that

␈↓ α∧␈↓influence than an AI person.

␈↓ α∧␈↓␈↓ αTNow␈αI␈αwant␈αto␈αdeal␈αwith␈αthe␈αissue␈αabout␈αwhether␈αa␈αmachine␈αreally␈αthinks␈αor␈αbelieves.␈α This

␈↓ α∧␈↓is␈α⊂an␈α⊂elaboration␈α⊂of␈α⊃a␈α⊂point␈α⊂I␈α⊂made␈α⊃in␈α⊂my␈α⊂lecture.␈α⊂ Namely,␈α⊂we␈α⊃will␈α⊂find␈α⊂it␈α⊂necessary␈α⊃to␈α⊂use

␈↓ α∧␈↓mentalistic␈αterminology␈αin␈α
describing␈αwhat␈αwe␈α
know␈αabout␈αmachines.␈α
 Of␈αcourse,␈αif␈αwe␈α
understand

␈↓ α∧␈↓how␈α∪a␈α∪thermostat␈α∪works,␈α∪we␈α∪don't␈α∪have␈α∪to␈α∪adopt␈α∪the␈α∪mentalistic␈α∪stance␈α∪of␈α∪saying␈α∀that␈α∪the

␈↓ α∧␈↓thermostat thinks the room is too warm.

␈↓ α∧␈↓␈↓ αTIndeed␈αI␈α
picked␈αthe␈α
thermostat␈αexample,␈α
precisely␈αbecause␈α
we␈αcan␈α
understand␈αit␈α
both␈αways␈α
-

␈↓ α∧␈↓mechanistically␈α
and␈αmentalistically.␈α
 Just␈αbecause␈α
we␈αcan␈α
understand␈αits␈α
mechanism␈αis␈α
not␈αa␈α
reason

␈↓ α∧␈↓to␈αbar␈αthe␈αuse␈αof␈αmentalistic␈αterms.␈α There's␈αan␈αilluminating␈αanalogy␈αwith␈αthe␈αnumber␈αsystem␈αand

␈↓ α∧␈↓its␈α∞historical␈α∂development.␈α∞ Suppose␈α∂someone␈α∞said␈α∞that␈α∂he␈α∞didn't␈α∂think␈α∞that␈α∞one␈α∂is␈α∞a␈α∂number␈α∞-

␈↓ α∧␈↓arguing␈α
that␈α
if␈α
you␈αhave␈α
only␈α
one␈α
thing␈αyou␈α
don't␈α
have␈α
to␈αcount.␈α
 Indeed␈α
most␈α
languages␈αtreat␈α
one

␈↓ α∧␈↓differently␈αfrom␈αthe␈αother␈αnumbers.␈α Some␈αtreat␈αtwo␈αdifferently␈αalso,␈αand␈αin␈αRussian␈αnumbers␈αup

␈↓ α∧␈↓to␈α
four␈αtake␈α
the␈α
genitive␈αcase.␈α
 The␈α
introduction␈αof␈α
zero␈α
to␈αthe␈α
number␈α
system␈αis␈α
even␈αmore␈α
recent,

␈↓ α∧␈↓and␈αI␈αbelieve␈αit␈αwas␈αcontroversial.␈α The␈αjustification␈αis␈αthat␈αthe␈αnumber␈αsystem␈αas␈αa␈αsystem␈αmakes

␈↓ α∧␈↓more␈α⊃sense␈α⊃if␈α⊂both␈α⊃zero␈α⊃and␈α⊂one␈α⊃are␈α⊃included.␈α⊂ Likewise,␈α⊃a␈α⊃systematic␈α⊂treatment␈α⊃of␈α⊃belief␈α⊂by

␈↓ α∧␈↓machines will have to start with the easy cases.

␈↓ α∧␈↓␈↓ αTA␈αmore␈αcomplex␈αcase␈αarises␈α
when␈αwe␈αsay␈αthat␈αa␈αdog␈α
wants␈αto␈αgo␈αout.␈α We␈αcannot␈α
practically

␈↓ α∧␈↓reduce␈α
this␈α
to␈α
a␈αpropensity␈α
to␈α
behave␈α
in␈α
a␈αcertain␈α
way,␈α
because␈α
we␈αmay␈α
not␈α
know␈α
what␈α
the␈αdog

␈↓ α∧␈↓will␈αdo␈αto␈αfurther␈αthis␈αdesire.␈α It␈αmay␈αscratch␈αthe␈αdoor␈αor␈αyelp␈αor␈αwhatever.␈α Secondly,␈αwe␈αmay␈αnot
␈↓ α∧␈↓␈↓ f13


␈↓ α∧␈↓know␈αthe␈αevidence␈αthat␈αthe␈αdog␈αwants␈αto␈αgo␈αout.␈α Therefore,␈αthe␈αfact␈αthat␈αthe␈αdog␈αwants␈αto␈αgo␈αout

␈↓ α∧␈↓is best treated as primary.

␈↓ α∧␈↓␈↓ αTAnother␈αuseful␈αidea␈αcomes␈αfrom␈αDan␈αDennett␈α-␈αthe␈αnotion␈αof␈αthe␈α"design␈αstance".␈α Suppose

␈↓ α∧␈↓we␈αare␈α
designing␈αa␈α
dog␈αas␈α
an␈αartificial␈α
intelligence.␈α It␈α
will␈αbe␈α
convenient␈αto␈α
design␈αin␈α
the␈αdesire␈α
to

␈↓ α∧␈↓go␈αout␈αas␈αa␈αpossible␈αstate.␈α We␈αhave␈αa␈αvariety␈αof␈αchoices␈αas␈αto␈αwhat␈αwill␈αcause␈αthis␈αstate␈αand␈αwhat

␈↓ α∧␈↓our␈αdog␈αwill␈αdo␈αto␈αrealize␈αthe␈αdesire.␈α In␈αdesigning␈αcomputer␈αsystems,␈αwe␈αwill␈αalso␈αfind␈αthis␈αnotion

␈↓ α∧␈↓of ␈↓↓wanting␈↓ a useful intermediate point.

␈↓ α∧␈↓␈↓ αTAs␈α
far␈α
as␈α
I␈α
can␈α
see,␈α
the␈α
purely␈α
intellectual␈α
terms␈α
are␈α
easier␈α
to␈α
handle␈α
for␈α
machines␈αthan␈α
some

␈↓ α∧␈↓of␈αthe␈αemotional␈αterms.␈α "It␈αbelieves"␈αis␈αeasier␈αthan␈α"it␈αhopes",␈αwhich␈αis␈αeasier␈αthan␈α"it␈αlikes␈αme"␈αor

␈↓ α∧␈↓"it␈αdoesn't␈αlike␈αme".␈α And␈αas␈αto␈αwhether␈αthe␈αmachine␈αis␈αsuffering,␈αall␈αI␈αcan␈αsay␈αis␈αthat␈αit␈α
complains

␈↓ α∧␈↓a lot.

␈↓ α∧␈↓␈↓ αTWhen␈α
we␈α
ask␈α
whether␈α
it␈α
is␈α
conscious,␈α
there␈α
are␈α
a␈α
lot␈α
of␈α
criteria␈α
for␈α
saying␈α
no.␈α
 No,␈αbecause␈α
it

␈↓ α∧␈↓doesn't␈α∞know␈α
about␈α∞its␈α
physical␈α∞body.␈α
 No,␈α∞it␈α
doesn't␈α∞even␈α
refer␈α∞to␈α
itself.␈α∞ On␈α
the␈α∞other␈α∞hand␈α
it

␈↓ α∧␈↓might␈αclaim␈α
to␈αbe␈α
alienated,␈αbut␈αit␈α
has␈αjust␈α
read␈αMarcuse.␈α Well␈α
that's␈αhow␈α
most␈αpeople␈αwho␈α
claim

␈↓ α∧␈↓to be alienated come to claim it.  It's something they read about.

␈↓ α∧␈↓McCarthy:␈αFirst␈αof␈αall␈αthe␈αChinese␈αroom.␈α
 There␈αis␈αconfusion␈αbetween␈αthe␈αsystem␈αconsisting␈αof␈α
the

␈↓ α∧␈↓person␈α∞and␈α∂the␈α∞person.␈α∂ I␈α∞agree␈α∂with␈α∞Robert␈α∂Wilensky␈α∞who␈α∂made␈α∞the␈α∂same␈α∞point␈α∂earlier.␈α∞ The

␈↓ α∧␈↓system␈αknows␈αChinese,␈αbut␈αthe␈αperson␈αwho␈αis␈αexecuting␈αthe␈αsystem␈αmay␈αnot.␈α This␈αis␈αanalogous␈αto

␈↓ α∧␈↓an␈α⊗interpreter␈α∃running␈α⊗in␈α∃a␈α⊗computer;␈α∃the␈α⊗interpreted␈α∃program␈α⊗often␈α∃has␈α⊗capabilities␈α∃the

␈↓ α∧␈↓interpreter␈α
does␈α
not.␈α
 It's␈α
just␈α
that␈α
we␈α
don't␈α
have␈α
experience␈α
with␈α
systems␈α
in␈α
which␈α
a␈α
person␈α
carries

␈↓ α∧␈↓out␈α
a␈αmental␈α
process␈α
that␈αhas␈α
properties␈α
different␈αfrom␈α
those␈αof␈α
the␈α
person␈αhimself.␈α
 We␈α
get␈αthe

␈↓ α∧␈↓same␈αconfusion␈αwith␈α
computers.␈α Someone␈αasks␈α
me␈αwhether␈αLISP␈α
can␈αdo␈αcalculus␈α
problems.␈α No,

␈↓ α∧␈↓LISP cannot do calculus, but some LISP programs can.

␈↓ α∧␈↓␈↓ αTThe␈α⊂example␈α⊂of␈α⊂thirst␈α⊂is␈α⊂different.␈α⊂ A␈α⊂program␈α⊂that␈α⊂simulates␈α⊂thirst␈α⊂is␈α⊂not␈α⊂going␈α⊃to␈α⊂be

␈↓ α∧␈↓thirsty.  For example, there is no way to relieve it with real water.
␈↓ α∧␈↓␈↓ f14


␈↓ α∧␈↓␈↓ αTSearle␈αhas␈α
said␈αthat␈αthe␈α
people␈αin␈αAI␈α
take␈αthe␈α
third␈αperson␈αview␈α
of␈αmental␈αqualities.␈α
 That's

␈↓ α∧␈↓correct.␈α We␈αdo,␈αand␈αwe'll␈αclaim␈αthat␈αit's␈αa␈αvirtue.␈α He␈αsays␈αwe␈αconsider␈αthe␈αproblem␈αof␈α
intelligence

␈↓ α∧␈↓as␈α⊃distinct␈α⊃from␈α⊃biology.␈α⊃ Yes,␈α⊃we␈α⊃hold␈α∩that␈α⊃intelligence␈α⊃is␈α⊃something␈α⊃that␈α⊃can␈α⊃be␈α∩dealt␈α⊃with

␈↓ α∧␈↓abstractly␈αjust␈αas␈αcomputation␈αcan␈αbe␈αdiscussed␈αand␈αdealt␈αwith␈αabstractly.␈α One␈αcan␈αask␈αwhether␈αa

␈↓ α∧␈↓computer␈α
calculates␈α
the␈α∞sum␈α
of␈α
3␈α
and␈α∞5␈α
to␈α
be␈α
8␈α∞in␈α
the␈α
same␈α
sense␈α∞as␈α
a␈α
human␈α
does.␈α∞ I␈α
suppose

␈↓ α∧␈↓Searle␈αwould␈αagree␈αthat␈α"calculate"␈αis␈αbeing␈αused␈αin␈αthe␈αsame␈αsense␈αfor␈αboth␈αhuman␈αand␈αmachine

␈↓ α∧␈↓in this case.

␈↓ α∧␈↓␈↓ αTNow␈α∞there's␈α∂the␈α∞point␈α∂Dreyfus␈α∞made␈α∂about␈α∞it␈α∂taking␈α∞300␈α∂years.␈α∞ I␈α∂have␈α∞been␈α∂saying␈α∞that

␈↓ α∧␈↓human␈α
level␈αAI␈α
will␈αtake␈α
between␈α
5␈αand␈α
500␈αyears.␈α
 The␈αproblem␈α
isn't␈α
that␈αit␈α
will␈αtake␈α
a␈αlong␈α
time

␈↓ α∧␈↓to␈α∞enter␈α∞data␈α∂into␈α∞a␈α∞computer.␈α∞ Its␈α∂rather␈α∞that␈α∞conceptual␈α∞advances␈α∂are␈α∞required␈α∞before␈α∂we␈α∞can

␈↓ α∧␈↓implement␈αhuman␈αlevel␈αartificial␈αintelligence␈α-␈αjust␈αas␈αconceptual␈αadvances␈αwere␈αrequired␈αbeyond

␈↓ α∧␈↓the situation in 1900 before we could have nuclear energy.

␈↓ α∧␈↓␈↓ αTPursuing␈α⊃the␈α∩nuclear␈α⊃energy␈α∩analogy,␈α⊃the␈α⊃question␈α∩is␈α⊃whether␈α∩the␈α⊃present␈α∩AI␈α⊃situation

␈↓ α∧␈↓corresponds␈α
to␈α
1900␈α
or␈α
to␈α
1938␈αwhen␈α
Rutherford,␈α
the␈α
leading␈α
nuclear␈α
physicist,␈α
declared␈αnuclear

␈↓ α∧␈↓energy␈α∂impossible.␈α∂ The␈α∂situation␈α∂of␈α⊂1938␈α∂is␈α∂interesting␈α∂in␈α∂that␈α∂experiments␈α⊂exhibiting␈α∂nuclear

␈↓ α∧␈↓fission␈α
had␈αalready␈α
been␈αdone␈α
but␈αhad␈α
been␈αmisinterpreted.␈α
 Perhaps␈αsomeone␈α
has␈α
already␈αdone

␈↓ α∧␈↓experimental␈α∞research␈α∂which,␈α∞when␈α∞properly␈α∂interpreted,␈α∞will␈α∞make␈α∂possible␈α∞human␈α∞level␈α∂AI.␈α∞ I

␈↓ α∧␈↓would␈αbe␈αvery␈αsurprised.␈α When␈αwe␈αtalk␈αabout␈αfuture␈αconceptual␈αadvances,␈αwe␈αdon't␈αknow␈αwhere

␈↓ α∧␈↓we stand at present.

␈↓ α∧␈↓␈↓ αTBert␈αmade␈αa␈αpoint␈αabout␈α
a␈αreservation␈αmachine␈αnot␈αknowing␈α
whether␈α6:25␈αwill␈αdo␈αas␈αa␈α
little

␈↓ α∧␈↓earlier␈αthan␈α
6:30.␈α The␈α
program␈αwould␈α
have␈αthe␈αsame␈α
problem␈αif␈α
it␈αwere␈α
making␈αa␈αreservation␈α
for

␈↓ α∧␈↓a␈αrobot.␈α Whether␈αeven␈αa␈α6:29␈αreservation␈αwill␈αdo␈αdepends␈αon␈αcircumstances.␈α So␈αthe␈αfact␈αthat␈αthe

␈↓ α∧␈↓reservation is for humans isn't the problem.

␈↓ α∧␈↓␈↓ αTFinally,␈α∂let␈α⊂me␈α∂defend␈α⊂Searle␈α∂on␈α⊂one␈α∂point.␈α∂ He␈α⊂was␈α∂discussing␈α⊂whether␈α∂a␈α⊂computer␈α∂can
␈↓ α∧␈↓␈↓ f15


␈↓ α∧␈↓think␈α∞in␈α∂the␈α∞same␈α∞sense␈α∂as␈α∞a␈α∞human␈α∂-␈α∞not␈α∞does␈α∂it␈α∞think␈α∞in␈α∂the␈α∞same␈α∞way.␈α∂ In␈α∞my␈α∂opinion␈α∞the

␈↓ α∧␈↓thermostat␈α∂thinks␈α⊂the␈α∂room␈α⊂is␈α∂too␈α⊂warm␈α∂in␈α⊂the␈α∂same␈α⊂sense␈α∂as␈α⊂a␈α∂human␈α⊂might,␈α∂and␈α⊂he␈α∂would

␈↓ α∧␈↓disagree.  Likewise about whether the dog simulation wants to go out.

␈↓ α∧␈↓***

␈↓ α∧␈↓McCarthy:␈αThe␈αquestion␈αis␈αwhat␈α
will␈αbe␈αthe␈αsituation␈αfive␈αto␈α
ten␈αyears␈αfrom␈αnow;␈αlet␈αme␈α
make␈αit

␈↓ α∧␈↓ten␈α
or␈α
fifteen.␈α
 I␈α
think␈α
there'll␈α
be␈α
a␈α
paradigm␈α
shift␈α
among␈α
the␈α
public␈α
that␈α
will␈α
give␈α
John␈αSearle␈α
the

␈↓ α∧␈↓following␈α
problem.␈α
 He␈αwill␈α
want␈α
to␈αcome␈α
to␈α
the␈α
symposium␈αto␈α
correct␈α
our␈αuse␈α
of␈α
mental␈αterms,␈α
but

␈↓ α∧␈↓he␈α
won't␈α
even␈α
get␈α
here,␈α
because␈α
he'll␈α
have␈αto␈α
correct␈α
his␈α
secretary␈α
who␈α
will␈α
tell␈α
him,␈α
"It␈αpromised␈α
to

␈↓ α∧␈↓process␈α⊂your␈α⊂travel␈α⊂advance,␈α⊂but␈α⊂I␈α⊂don't␈α⊂think␈α⊂it␈α⊂will,␈α⊂because␈α⊂it's␈α⊂puzzled␈α⊂about␈α⊂whether␈α⊂the

␈↓ α∧␈↓expenditure for flowers was intended and necessary for the business's goals".

␈↓ α∧␈↓␈↓ αTThus␈α∪in␈α∪ten␈α∪of␈α∪fifteen␈α∪years,␈α∀quite␈α∪mundane␈α∪systems␈α∪used␈α∪for␈α∪business␈α∀and␈α∪personal

␈↓ α∧␈↓purposes␈α∂will␈α∂require␈α∞the␈α∂use␈α∂of␈α∞a␈α∂certain␈α∂amount␈α∞of␈α∂mental␈α∂terminology␈α∞in␈α∂order␈α∂to␈α∂use␈α∞them

␈↓ α∧␈↓effectively.

␈↓ α∧␈↓␈↓ αTAlso␈α
let␈α
me␈α
repeat␈α
my␈α
warning␈α
to␈α
philosophers␈α
that␈α
if␈α
they␈α
insist␈α
on␈α
discussing␈αcommon␈α
sene

␈↓ α∧␈↓reasoning␈αonly␈αat␈αthe␈α
general␈αlevel␈αof␈αtoday's␈αdiscussion,␈α
they␈αwill␈αlose␈αthe␈αjurisdiction.␈α
 We␈αneed

␈↓ α∧␈↓to␈α
consider␈α
the␈α
conditions␈α
for␈α
the␈α
ascription␈α
of␈α
particular␈α
mental␈α
qualities,␈α
and␈α
this␈α
may␈αrequire

␈↓ α∧␈↓collaboration among philosophers and artificial intelligence researchers.

␈↓ α∧␈↓␈↓ αTWe␈αattempted␈αsuch␈αa␈αcollaboration␈αseveral␈αyears␈αago,␈αbut␈αI␈αthink␈αthe␈αparticular␈αattempt␈α
was

␈↓ α∧␈↓unsuccessful␈α
largely␈α
because␈α
it␈α
considered␈α∞overly␈α
general␈α
questions.␈α
 This␈α
was␈α
partly␈α∞because␈α
the

␈↓ α∧␈↓AI␈α
people␈α
succumbed␈α
to␈α
the␈α
temptation␈α
to␈α
become␈α
amateur␈α
philosophers␈α
rather␈α
than␈α
raising␈αthe

␈↓ α∧␈↓AI issues to which philosophy is relevant.