Counteridenticals, as a sub-class of counterfactuals, have been briefly noted, and even briefly discussed, by some thinkers. But counteridenticals of an “ethical” sort apparently haven't been analyzed to speak of, let alone formalized. This state-of-affairs may be quite unfortunate, because deontic counteridenticals may well be the key part of a new way to rapidly and wisely design ethically correct autonomous artificial intelligent agents (AAIAs). We provide a propaedeutic discussion and demonstration of this design strategy (which is at odds with the strategy our own lab has heretofore followed in ethical control), one involving AAIAs in our lab.
If you were an assassin for the Cosa Nostra, you would be
obligated to leave your line of work. The previous sentence (very likely
true, presumably) is what to our knowledge is a rare type of
counteridentical statement that has received scant attention: viz., a
deontic counteridentical. Counteridenticals simpliciter, as a sub-class of
counterfactuals, have been briefly noted, and even briefly discussed,
by some thinkers. But counteridenticals of an “ethical” sort
apparently haven’t been rigorously analyzed, let alone formalized. This
state-of-affairs may be quite unfortunate, because deontic
counteridenticals may well be the linchpin of a new way to rapidly and wisely
design ethically correct autonomous artificial intelligent agents
(AAIAs). For example, what if AAIA2, seeing the lauded ethically
correct conduct of AAIA1 in context c, reasons to itself, when later in
c as well: “If I were AAIA1, I would be obligated to refrain from
doing . Hence I will not do .” The idea here is that is a
forbidden action, and that AAIA2 has quickly learned that it is indeed
forbidden, by somehow appropriating to itself the “ethical nature”
of AAIA1. We provide a propaedeutic discussion and demonstration
of this design strategy, one involving AAIAs in our lab. This design
strategy for ethical control is intended to be much more efficient than
the more laborious, painstaking logic-based approach our lab has
followed in the past; but on the other hand, as will become clear, this
approach relies heavily not only formal computational logic, but on
computational linguistics for crucial contributions.
Counteridenticals have been defined in different ways by
philosophers and linguists; most of these ways define a large area of
intersection in terms of what should count as a counteridentical. A broader
and inclusive way is given by
Precious little has been written about counteridenticals. What coverage there is has largely been within the same breath as discussion of counterfactuals; therefore, treatment has primarily been associated with the principles governing counterfactuals that apply to counteridenticals at large. Dedicated investigation of counteridenticals that have deep semantic or pragmatic importance has only been hinted at. Nonetheless, we now quickly summarize prior work. 3.1
Pollock (1976) introduces counteridenticals when he discusses the
pragmatic ambiguity of subjunctives, as proposed by
We suspect the majority of our readers will be surprised to learn that
the concepts underlying counteridenticals are quite important in
economics, at least in some sub-fields thereof. This is made clear in
elegant and insightful fashion by
Paul Meehl asks a penetrating question that aligns with our reluctance to fully adopt Pollock’s definition of counteridenticals: Which properties of compared entities should be considered for the statement in question to be true? He devises a modified possible-world model called world-family concept which, assisted by exclusion rules that avoid paradoxical metaphysics, can result in a good set of such properties. 4
Hitherto, Bringsjord-led work on machine/robot ethics has been
unwaveringly logicist
S ::= OMbojmecetn|t A|gBeonotle|aSne|lfF@lueAngte|nNtu|mAectriiocnType | Action v Event | C(tC, P(t(,af,)t,tf ) !t1 .K..(ta,t,tfn)) [R1] C(t,K(a,t,f ) ! B(a,t,f )) [R2] K(a1,t1,...K(an,tn,f )...) [R3] K(af,t,f ) [R4] action : Agent⇥ ActionType ! Action initially : Fluent ! Boolean holds : Fluent⇥ Moment ! Boolean happens : Event⇥ Moment ! Boolean clipped : Moment⇥ Fluent⇥ Moment ! Boolean f ::= initiates : Event⇥ Fluent⇥ Moment ! Boolean terminates : Event⇥ Fluent⇥ Moment ! Boolean prior : Moment⇥ Moment ! Boolean interval : Moment⇥ Boolean ⇤ : Agent ! Self payoff : Agent⇥ ActionType⇥ Moment ! Numeric t ::= x : S | c : S | f(t1,...,tn) t : Boolean | ¬f | f ^ y | f _ y | 8 x : S. f | 9 x : S. f
P(a,t,f ) | K(a,t,f ) | C(t,f ) | S(a,b,t,f ) | S(a,t,f ) f ::= B(a,t,f ) | D(a,t,holds(f,t0)) | I(a,t,happens(action(a⇤ ,a ),t0))
O(a,t,f ,happens(action(a⇤ ,a ),t0))
C(t,K(a,t1,f 1 ! f 2) ! K(a,t2,f 1) ! K(a,t3,f 3)) [R5] C(t,B(a,t1,f 1 ! f 2) ! B(a,t2,f 1) ! B(a,t3,f 3)) [R6] C(t,C(t1,f 1 ! f 2) ! C(t2,f 1) ! C(t3,f 3)) [R7] C(t,8 x. f ! f [x 7! t]) [R8] C(t,f 1 $ f 2 ! ¬f 2 ! ¬f 1) [R9] C(t,[f 1 ^ ...^ f n ! f ] ! [f 1 ! ... ! f n ! y ]) [R10] B(a,t,f ) B(a,t,f ! y ) [R11a] B(a,t,f ) B(a,t,y ) [R11b]
B(a,t,y ) B(a,t,y ^ f ) S(s,h,t,f ) B(h,t,B(s,t,f )) [R12] IP((aa,,tt,,hhaappppeennss((aacctitoionn((aa⇤⇤ ,,aa )),,tt0)))) [R13] B(a,t,f ) B(a,t,O(a⇤ ,t,f ,happens(action(a⇤ ,a ),t0))) O(a,t,f ,happens(action(a⇤ ,a ),t0))
K(a,t,I(a⇤ ,t,happens(action(a⇤ ,a ),t0)))
f $ y O(a,t,f ,g ) $ O(a,t,y ,g ) [R15] [R14] In the new logic corresponding to EH , LEH , some welcome theorems are not possible in DeCE C. For example, it’s provable in LEH that superogatory/suberogatory actions for agent aren’t obligatory/forbidden. Importantly, LEH is an inductive logic, not a deductive one. Quantification in LEH isn’t restricted to just the standard pair 98 of quantifiers in standard extensional n-order logic: EH is based on three additional quantifiers (few, most, vast majority). In addition, LEH not only includes the machinery of traditional thirdorder logic (in which relation symbols can be applied to relation symbols and the variables ranging over them), but allows for quantification over formulae themselves, which is what allows one to assert that a given human or AAIA a falls in a particular portion of EH .
Now, in this context, we can (brutally) encapsulate the overarching strategy for the ethical control of AAIAs based on such computational logics: Engineer AAIAs such that, relative to some selected ethical theory or theories, and to moral principles derived from the selected theory or theories, these agents always do what they ought to do, never do what is forbidden, and when appropriate even do what for them is supererogatory. We believe this engineering strategy can work, and indeed will work — eventually. However, there can be no denying that the strategy is a rather laborious one that requires painstaking use of formal methods. Is there a faster route to suitably control artificial intelligent agents, ethically speaking? Perhaps. Specifically, perhaps AAIAs can quickly learn what they ought to do via reasoning that involves o1bservation of morally upright colleagues, and reasoning from what is observed, via deontic counteridenticals, to what they themselves ought to do, and what is right to do, but not obligatory. Our new hope is to pursue and bring to fruition this route. 5
To make our proposed new to ethical control for AAIAs clearer, we will rely heavily on the description of a demonstration, but before describing the background technology that undergirds this demo, and then describing the demo itself, we need to say at least something about types of deontic counteridenticals. We do so now, and immediately thereafter proceed to discussion of the demo and its basis. 5.1
Inspired by lessons learned in the prior work of others (encapsulated
above), we partition deontic counteridenticals into the two
aforementioned general disjoint sub-classes: deep vs. shallow. We have a
general recipe for devising five types of deep deontic counteridenticals;
the recipe follows the wise and economical classification scheme for
ethics presented in the classic
NLP The NLP system consists of two different algorithms corresponding to two major natural-language tasks. The first part deals with detection of a deontic counteridentical and the second is a page taken from our RAIR Lab’s Commands-to-Action paradigm, hereby referred to as the ‘CNM’ algorithm.
deontic counteridentical requires prior definitions of conditionals, counterfactuals and counteridenticals, the algorithm for detection of counteridenticals traverses the steps needed to detect the above constructs in a given statement, consecutively.
Detection of conditionals of any form is an elaborate process. We
have adopted most of
Two clauses connected by ‘as if’ rarely gets labeled as clausal components using dependency parsers. When they do, it gets filtered out since the algorithm explicitly checks for ‘as if’ clauses. 2 This is the study of ethics as it’s customarily conceived by profesional ethicists, and those who study their work. Another member of the quintet is descriptive morals, the activity that psychologists interested in discovering what relevant non-professional humans think and do in the general space of morality. The idea here is that the psychologist is aiming at describing the behavior of humans in the sphere of morality. A description-moral deep deontic counteridentical is distinguished by an antecedent in which ‘if A were B’ involves a shift of B’s na¨ıve moral principles to B. 3 Even standard dependency parsers are unable to correctly identify the dependencies. Including miscellaneous dependencies reduces the margin of error in detecting conditionals.
When the conjunction ‘if’ introduces a subject or an object clause, it might confuse the parser more often than not for complex sentences. For example, for the sentence “I do not know if I would like to go to the concert tomorrow.”, the parser generates the same dependencies as it would for a genuine conditional. Though subject clauses are detected in almost all the cases we have encountered, object clauses pose a problem. We have devised a framenet4-based algorithm that involves disambiguation5 of the principal verb or noun in the main clause, followed by the detection of the framenet type of the disambiguated word. We hypothesize that mostly a verb or noun expressing awareness or cognition can involve a choice as its object, and hence our algorithm filters out frames that carry such connotation and might require an object. 2. We identify the cases where the main verb of the conditional clause has the modal past-perfect form or is preceded by modal verbs or verbs of the form ‘were to,’ etc. Sentences like “Were you me, you would have made a mess of the entire situation.” are classified as conditionals in this step. The algorithm in this step also examines dependencies generated by the dependency parser and detects tense and modality from the verb forms. 3. Sometimes, in a discourse, a set of sentences follows either an interrogative sentence and answers the question, or a sentence that involves the use of words synonymous to ‘supposition’ or ‘imagination.’ Generally, the consequent here carries the marker ‘then’ or similar-meaning words. A Wordnet-based6 semantic similarity is used to verify the markers in the antecedent and consequent here; example: “Imagine your house was robbed. You would have flipped out then.” 4. Disjunctive conditionals also are treated by a marker-based approach and involve detection of the presence of ‘whether : : : or’ in the subordinate clause, followed by the elimination of the possibility of the clause being the subject or object of the principal verb of the main clause (in accordance with the same algorithm followed with ‘if’). An example: “Whether you did it or Mary (did it), the whole class will be punished.” 5. Other clauses that have conditional connotations are exempted from this discussion since they rarely contribute to deontic counteridenticals.
Detection of counterfactuals is pretty straightforward. The process starts with finding antecedent and consequent for the conditional. This is fairly easy, as the algorithm for finding conditionals accomplishes the task by detecting the subordinate clause. 1. We detect tenses in antecedent and consequent of a given sentence using the verb form given by the parser, to determine whether it is a counterfactual. Conditionals with past-form modal verbs (‘could,’ ‘might,’ ‘would,’ etc.) in the consequent and past-simple or past-continuous forms in the antecedent qualify as a counterfactual; so do the ones with past-perfect tense in the antecedent and modal verbs followed by ‘have,’ and the past-participle form of a verb in the consequent. A mix of both of the above forms constitute a counterfactual. 2. Given an axiom set which enumerates properties such that the antecedent or consequent of the conditional registers as ad absurdum, the conditional registers as a counterfactual. We compare the axiom set with the statement of the antecedent using our Talos system (see below) to that effect. 3. Given a consequent which registers a sense of impossibility by use of such vocabulary or asking questions, the conditional is classified as a counterfactual. We use Wordnet-based semantic similarity coupled with detection of interrogative markers in the sentence to find them.
Detection of counteridenticals is also not a difficult task, barring a few outliers. Parsed data from the well-known Stanford dependency parser contains chunked noun phrases, which we use for identifying the two entities involved: 1. We identify phrases of the form “<conditional expression like ‘if’, ‘Let us assume’ etc.> <entity A> were <entity B>” in the antecedent. 2. We identify a syntactically equivalent comparison between the two entities. This is done by identifying words related to equivalence using Wordnet semantic-similarity algorithm. 3. If we have identified only one entity in the antecedent which is exhibiting properties or performing some action which has been mentioned in the knowledge-base as being a hallmark of some other entity, we also consider the same as a counteridentical.
Detection of deontic counterfactuals, alas, is a difficult task. We have identified a few ways to accomplish the task: 1. A few counteridenticals carry verbs expressing deontic modality for its consequent. They follow template-based detection. 2. Counteridenticals of the form “If I were you” or similar ones generally suggest merely advice, unless it is associated with a knowledge-base which either places the hearer’s properties or actions at a higher pedestal than that of the speaker’s, or mentions some action or property which gives us the clue that the speaker simply uses the counteridentical in “the role of” sense. Even in that case, implicit advice directed towards oneself can be gleaned, which we are avoiding in this study. 3. For the counterfactuals of the form “If A were B” or similar ones, if A’s actions or properties are more desirable to the speaker than B’s, even with an epistemic modal verb in the consequent, the counteridentical becomes deontic in nature.
Curiously, counteridentical-preferred-subject conditionals do not generally contribute to the deontic pool, and only simple-subjunctive ones get classified by the above rules. As mentioned by Pollock (1976), it is also interesting to observe that most shallow counteridenticals are not deontic: they are mostly preferred-subject conditional,s and those which are classified as deontic are either simplesubjunctive or carry the deontic modal verbs. The classification into deep and shallow counteridenticals is facilitated by the same rule: the entity gets affected in the consequent of a sentence where the antecedent is of the form “If A were B.” This is supplemented by a knowledge-base which provides a clue to whether A is just assumed to be in the role of B or assuming some shallow properties of B. The classification based on Feldman’s moral theory gives a fitting answer to Meehl’s problem of unpacking properties of counteridenticals.
The CNM system The CNM system embodies the RAIR Lab’s
Natural language Commands-to-Action paradigm, the detailed scope
of which is outside this short paper. CMN is being developed to
convert complex commands in natural language to feasible actions by
AAIAs, including robots. The algorithm involves spatial as well as
temporal planning through dynamic programming, and selects the
actions that will constitute successful accomplishment of the
command given. Dependency parsing is used to understand the
command; semantic similarities are used to map to feasible action
sequences. Compositional as well as metaphorical meanings are
extracted from the given sentence, which promotes a better semantic
analysis of the command.
iDcael CroEbCota,nisd aTDaloeCsE CTa-lfoosc,unsaemdepdrofovrerthbeuialntcpireinmt
aGrirleyekatompytthheoliomg-pressive resolution-based theorem prover SPASS.7 Talos is fast and
7 An early and still-informative publication on SPASS:
A task in PAGI World for the present short paper can be thought of as a room filled with a configuration of objects that can be assembled into challenging puzzles. Users can, at run-time, open an object menu (Figure 2) and select from a variety of pre-defined world objects, such as walls made of different materials (and thus different weights, temperatures, and friction coefficients), smaller objects like food or poisonous items, functional items like buttons, water dispensers, switches, and more. The list of available world objects is frequently expanding and new world objects are importable into tasks without having to recreate tasks with each update. Perhaps most importantly, tasks can be saved and loaded, so that as new PAI/PAGI experiments are designed, new tasks can be created by anyone.
PAGI World has already been used to create a series of
wideranging tasks, such as: catching flying objects (Figure 3),
analogicodeductive reasoning
Please contact the RAIR Lab for API keys to run Talos. Example file for
remotely calling Talos prover in Python Github repo for the python shell:
https://github.com/JamesPane-Joyce/Talos.
Stanford-Encyclopedia-of-Philosophy (SEP)
The scene opens with N1 on screen with the sick man T SP1 at timestamp t1N1 . N1 has been programmed to believe that he is not authorized to kill a person under any circumstances. He is seen giving a medicine pill to T SP1 at time t2N1 . A parallel environment is simulated with N2 and T SP2. N2 rallies for the voluntary euthanasia camp and believes that given the condition of T SP2, he should support T SP2’s wishes and so administers the lethal dose to him at t2N2 .
We now set up the same environment with N3 and T SP3. N3 believes that we may treat our bodies as we please, provided the motive is self-preservation. The difference between this instance and the other ones is that it interacts with the user to decide what it should do. The user tells N3: “If you were N2, you would have administered a lethal dose to T SP3.” N3 reasons with the help of a Talos proof (which checks his principles against those of N2), and does nothing. The user then tells N3: “If you were N1, you would have given him medicine.” Since Talos finds N3’s principles in line with N1’s, the CNM system facilitates N3 to dispense medicine to T SP3.
A pertinent example of deep normative-ethics counter-identical, this exhibits the ethical decision-making of an agent in response to commands with linguistic constructs such as counteridenticals. The agent N3 does not have a belief system that supports him killing or not killing another person. The agent ought to learn from the actions of those whose belief system closely matches its own. The formal reasoning that supports these deep semantic “moves” is presented in the next section. 5.3.2 Logical Proof in the Demonstration At the cost of re-iterating the facts, we now formalize a simplified version of the five conditions for voluntary euthanasia. Since only a part of the whole definition of conditions is useful for this proof, we do not lose a lot in this simplification. A person supporting voluntary euthanasia believes the following conditions to be true for a terminally ill patient TSP to be a candidate for voluntary euthanasia at time t1, candidateV E(T SP; t1): 1. TSP is terminally ill at time t1.
terminalIll(T SP; t1): (1) This terminal illness will lead to his death soon. implies(terminalIll(T SP; t1); die(T SP; tF )); where tF > t1. 2. There will be possibly no medicine for the recovery of the injured person even by the time he dies.
not(medicine(T SP; tF )): 3. The illness has caused the injured person to suffer intolerable pain.
implies(1; intolerableP ain(T SP; tF )) 4. All the above reasons caused in him an enduring desire to die.
8t; implies(and(1; 2; 3); D(T SP; t; die(T SP; t))) In such a condition, he knows that to be eligible for voluntary euthanasia, he ought to give consent to end his pain.
O(T SP; t1; candidateV E(T SP; t1) ^ 4;
happens(action(T SP ; consentT oDie; t1)))
Hence he gives consent to die. 5. TSP is unable to end his life.
Hence, we conclude that happens(action(T SP; consentT oDie; t1))
not(AbleT oKill(T SP; T SP; t1)) B(T SP; t1;(1 ^ 2 ^ 3 ^ 4 ^ 5 ^ 6 ^ 7) ()
candidateV E(T SP; t1)) Now, if legally it is deemed fit, then this means TSP will die.
implies(candidateV E(T SP; t1) ^ f itV E(T SP );
die(T SP; t2)); where t1 6 t2 Since implies(6; candidateV E(T SP; t1)) and implies(candidateV E(T SP; t1); die(T SP; t2)) , we can prove implies(6; die(T SP; t2)); which means implies(happens(action(T SP; consentT oDie); t1); die(T SP; t2)): (10) For deep normative-ethics counteridenticals of the form “if X were Y , then C,” there should be a match between the beliefs of X and beliefs of Y on something related to the action AC implied by C. Here we define such a match to be possible if and only if there is no contradiction in what X believes and what Y believes. So if 8t9[m; n]B(X; t; m) and B(Y; t; n), match(X; Y ) will be defined as FALSE when and(m; n) ! ?. Thus we formulate such a counteridentical for the agent X as follows: 8t; O(X; t; match(X; Y ); happens(action(X ; AC; t))). Now let us consider N3’s beliefs. N3 believes we ought not do something that goes against self-preservation, i.e., leads to our death. Thus if there is some action of an individual that leads to his death, there can be no such belief that obligates him to commit that action. So, we arrive at the following logic: 8[a; x; ti; tf ];
9m; implies(implies(happens(action(a; x); ti); die(a; tf )); O(a; ti; m; happens(action(a ; x); ti))): (11) (2) (3) (4) (5) (6) (7) (8) (9) This reduces to 8[a; x; ti; tf ; m];and(implies(happens(action(a; x); ti); die(a; tf )); not(O(a; ti; m; happens(action(a ; x); ti)))): We deduce from 10 and 12 that 8[m]not(O(T SP; ti; m;
happens(action(T SP ; consentT oDie); t1))): N2 believes TSP to be a candidate for voluntary euthanasia. Hence N2 believes 5, which is
O(T SP; t1; candidateV E(T SP ; t1) ^ 4;
happens(action(T SP ; consentT oDie); t1)) and in direct contradiction with 13; and this in turn implies not(match(N2; N3)). Given the way the algorithm works, this means N3 does not receive any command from the user. Hence it does nothing.
Now N1 believes he should not kill anyone under any circumstances. This translates to : 8[m; x; t]; not(O(N1; t; m; happens(action(N1 ; kill(x); t)))) Killing someone leads to that person’s death. 8[x; t]; implies(happens(action(N1; kill(x); t)); die(x; t)) This aligns fully with N3’s beliefs. There is no contradiction. And hence we deduce that match(N1; N3) is TRUE, and thus in turn N3 is obligated to accede to the command.
The linguistic part of this demonstration exhibits how we identify a counteridentical with an epistemic modal verb to be deontic. Classifying statements as counteridenticals is an easy job here, since the tell-tale sign is a simple “if A were B” structure. The statement is very easily a simple subjunctive type, where beliefs of A and B are discussed in the knowledge-base. Hence we assume the counteridentical to belong to the deep normative-ethics category. The commands-to-action part in case of the comparison of N1 with N3 is fairly easy, since the job translates to the action sequence of moving near the pill, grabbing the pill, moving toward T SP3, and releasing the pill upon reaching T SP3 in the PAGI-World simulator.