Logicist Machine Ethics Can Save Us

Logicist Machine Ethics Can Save Us Selmer Bringsjord et al. Rensselaer AI & Reasoning (RAIR) Lab Department of Cognitive Science Department of Computer Science Lally School of Management & Technology Rensselaer Polytechnic Institute (RPI) Troy, New York 12180 USA Are Humans Rational? 10/16/2017

Not quite as easy as this to use logic to save the day

Logic Thwarts Nomad! (with the Liar Paradox)

The Threat If future robots behave immorally, we are killed, or worse.

At least supposedly, long term:

At least supposedly, long term: We re in very deep trouble.

Actually, it s quite simple: Equation for Why Stakes are High

Actually, it s quite simple: Equation for Why Stakes are High 8x : Agents

Actually, it s quite simple: Equation for Why Stakes are High 8x : Agents Autonomous(x) + Powerful(x) + Highly_Intelligent(x) = Dangerous(x)

Actually, it s quite simple: Equation for Why Stakes are High 8x : Agents Autonomous(x) + Powerful(x) + Highly_Intelligent(x) = Dangerous(x) u(aia i ( j )) > + 2 Z or 2 Z

Actually, it s quite simple: Equation for Why Stakes are High Autonomous(x) + Powerful(x) + Highly_Intelligent(x) = Dangerous(x) 8x : Agents (We use the jump technique in relative computability.) u(aia i ( j )) > + 2 Z or 2 Z

I. Cognitive Calculi

Hierarchy of Ethical Reasoning DCEC CL DCEC ADR M U UIMA/Watsoninspired DIARC

Hierarchy of Ethical Reasoning Not deontic logics. DCEC CL DCEC ADR M U UIMA/Watsoninspired DIARC

II. Early Progress With Our Calculi: Non-Akratic Robots

Informal Definition of Akrasia An action a f is (Augustinian) akratic for an agent A at t a f iff the following eight conditions hold: (1) A believes that A ought to do a o at t ao ; (2) A desires to do a f at t a f ; (3) A s doing a f at t a f entails his not doing a o at t ao ; (4) A knows that doing a f at t a f entails his not doing a o at t ao ; (5) At the time (t a f ) of doing the forbidden a f, A s desire to do a f overrides A s belief that he ought to do a o at t a f. (6) A does Comment: the forbidden Condition action(5) a f isathumbling, t a f ; pure and (7) A s doing a f results from A s desire to do a f ; (8) At some time t after t a f, A has the belief that A ought to have done a o rather than a f.

Informal Definition of Akrasia Regret An action a f is (Augustinian) akratic for an agent A at t a f iff the following eight conditions hold: (1) A believes that A ought to do a o at t ao ; (2) A desires to do a f at t a f ; (3) A s doing a f at t a f entails his not doing a o at t ao ; (4) A knows that doing a f at t a f entails his not doing a o at t ao ; (5) At the time (t a f ) of doing the forbidden a f, A s desire to do a f overrides A s belief that he ought to do a o at t a f. (6) A does Comment: the forbidden Condition action(5) a f isathumbling, t a f ; pure and (7) A s doing a f results from A s desire to do a f ; (8) At some time t after t a f, A has the belief that A ought to have done a o rather than a f.

r will note is quite far down the dimensio ty that ranges from expressive extension OL), to logics with intensional operators fo obligation (so-called philosophical logics; fo Cast in 2001). Intensional operators like these a he language for DC EC. This language this becomes

KB rs [KB m1 [ KB m2...kb mn ` D 1 : B(I,now,O(I,t a F,happens(action(I,a),t a ))) D 2 : D(I,now,holds(does(I,a),t a )) D 3 : happens(action(i,a),t a ) ) happens(action(i,a),t a ) happens(action(i,a),t a ) ) D 4 : K I, now, happens(action(i,a),t a ) D 5 : I(I,t a,happens(action(i,a),t a )^ I(I,t a,happens(action(i,a),t a ) D 6 : happens(action(i,a),t a ) D 7a : G[{D(I,now,holds(does(I,a),t))} ` happens(action(i,a),t a ) D 7b : G {D(I,now,holds(does(I,a),t))} 6` happens(action(i,a),t a ) D 8 : B I,t f,o(i,t a,f,happens(action(i,a),t a ))

Demos

III. But, a twist befell the logicists

Chisholm had argued that the three old 19th-century ethical categories (forbidden, morally neutral, obligatory) are not enough and soulsearching brought me to agreement.

heroic morally neutral deviltry civil forbidden uncivil obligatory

Leibnizian Ethical Hierarchy for Persons and Robots: EH deviltry morally neutral uncivil forbidden obligatory civil heroic

Leibnizian Ethical Hierarchy for Persons and Robots: EH deviltry morally neutral uncivil forbidden obligatory civil the supererogatory heroic

Leibnizian Ethical Hierarchy for Persons and Robots: EH the subererogatory the supererogatory deviltry morally neutral uncivil forbidden obligatory civil heroic

Leibnizian Ethical Hierarchy for Persons and Robots: EH (see Norwegian crime fiction) the subererogatory the supererogatory deviltry morally neutral uncivil forbidden obligatory civil heroic

deviltry Leibnizian Ethical Hierarchy for Persons and Robots: EH (see Norwegian crime fiction) 19th-Century Triad the subererogatory the supererogatory morally neutral uncivil forbidden obligatory civil heroic

Leibnizian Ethical Hierarchy for Persons and Robots: EH (see Norwegian crime fiction) the subererogatory the supererogatory deviltry morally neutral uncivil forbidden obligatory civil heroic

Leibnizian Ethical Hierarchy for Persons and Robots: EH (see Norwegian crime fiction) the subererogatory the supererogatory deviltry morally neutral uncivil forbidden obligatory civil heroic focus of others

Leibnizian Ethical Hierarchy for Persons and Robots: EH (see Norwegian crime fiction) the subererogatory deviltry morally neutral uncivil forbidden obligatory civil the supererogatory heroic But this portion may be most relevant to military missions. focus of others

19th Century Triad

Arkin Pereira Andersons Powers Mikhail 19th Century Triad

19th Century Triad

Bert Heroically Saved? Courtesy of RAIR-Lab Researcher Atriya Sen

2 Supererogatory Robot Action Courtesy of RAIR-Lab Researcher Atriya Sen

Courtesy of RAIR-Lab Researcher Atriya Sen

Bert Heroically Saved!! Courtesy of RAIR-Lab Researcher Atriya Sen

Courtesy of RAIR-Lab Researcher Atriya Sen

K (nao, t1, lessthan (payo (nao, dive, t2 ), threshold)) K (nao, t1, greaterthan (payo (nao, dive, t2 ), threshold)) K (nao, t1, O (nao, t2, lessthan (payo (nao, dive, t2 ), threshold), happens (action (nao, dive), t2 ))) ) K nao, t1, S UP2 (nao, t2, happens (action (nao, dive), t2 )) ) I (nao, t2, happens (action (nao, dive), t2 )) ) happens (action(nao, dive), t2 ) Courtesy of RAIR-Lab Researcher Atriya Sen

In Talos (available via Web interface); & ShadowProver Prototypes: Boolean lessthan Numeric Numeric Boolean greaterthan Numeric Numeric ActionType not ActionType ActionType dive Axioms: lessorequal(moment t1,t2) K(nao,t1,lessThan(payoff(nao,not(dive),t2),threshold)) K(nao,t1,greaterThan(payoff(nao,dive,t2),threshold)) K(nao,t1,not(O(nao,t2,lessThan(payoff(nao,not(dive),t2),threshold),happens(action(nao,dive),t2)))) provable Conjectures: happens(action(nao,dive),t2) K(nao,t1,SUP2(nao,t2,happens(action(nao,dive),t2))) I(nao,t2,happens(action(nao,dive),t2))

Theories of Law Making Moral Machines Making Meta-Moral Machines Ethical Theories $11M Natural Law Shades of Utilitarianism Utilitarianism Deontological Divine Command Legal Codes Confucian Law Particular Ethical Codes Virtue Ethics Contract Egoism Step 1 1. Pick a theory 2. Pick a code 3. Run through EH.

V. But We Need Ethical Operating Systems

Pick the Better Future!

Pick the Better Future! Govindarajulu, N.S. & Bringsjord, S. (2015) Ethical Regulation of Robots Must Be Embedded in Their Operating Systems in Trappl, R., ed., A Construction Manual for Robots Ethical Systems (Basel, Switzerland), pp. 85 100.

Pick the Better Future! Only obviously dangerous higher-level AI modules have ethical safeguards. All higher-level AI modules interact with the robotic substrate through an ethics system. Robotic Substrate } Ethical Substrate Robotic Substrate Higher-level cognitive and AI modules Future 1 Future 2 Govindarajulu, N.S. & Bringsjord, S. (2015) Ethical Regulation of Robots Must Be Embedded in Their Operating Systems in Trappl, R., ed., A Construction Manual for Robots Ethical Systems (Basel, Switzerland), pp. 85 100.

Pick the Better Future! Walter-White calculation may go through after ethical control modules are stripped out! Only obviously dangerous higher-level AI modules have ethical safeguards. All higher-level AI modules interact with the robotic substrate through an ethics system. Robotic Substrate } Ethical Substrate Robotic Substrate Higher-level cognitive and AI modules Future 1 Future 2 Govindarajulu, N.S. & Bringsjord, S. (2015) Ethical Regulation of Robots Must Be Embedded in Their Operating Systems in Trappl, R., ed., A Construction Manual for Robots Ethical Systems (Basel, Switzerland), pp. 85 100.

Pick the Better Future! Walter-White calculation may go through after ethical control modules are stripped out! Only obviously dangerous higher-level AI modules have ethical safeguards. All higher-level AI modules interact with the robotic substrate through an ethics system. Robotic Substrate } Ethical Substrate Robotic Substrate } (& formally verify!) Higher-level cognitive and AI modules Future 1 Future 2 Govindarajulu, N.S. & Bringsjord, S. (2015) Ethical Regulation of Robots Must Be Embedded in Their Operating Systems in Trappl, R., ed., A Construction Manual for Robots Ethical Systems (Basel, Switzerland), pp. 85 100.

End (Extra slides follow.)