Adversarial Uses of Affective Computing and Ethical Implications

Submitted to the Program in Media Arts and Sciences, School of Architecture and Planning in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy in Media Arts and Sciences at the Massachusetts Institute of Technology September 2005

Carson Jonathan Reynolds

All rights reserved.

Revision History
Revision 0.12005-09

PDF

Abstract

Much existing affective computing research focuses on systems designed to use information related to emotion to benefit users. Many technologies are used in situations their designers didn't anticipate and would not have intended. This thesis discusses several adversarial uses of affective computing: use of systems with the goal of hindering some users. The approach taken is twofold: first experimental observation of use of systems that collect affective signals and transmit them to an adversary; second discussion of normative ethical judgments regarding adversarial uses of these same systems. This thesis examines three adversarial contexts: the Quiz Experiment, the Interview Experiment, and the Poker Experiment. In the Quiz Experiment, participants perform a tedious task that allows increasing their monetary reward by reporting they solved more problems than they actually did. The Interview Experiment centers on a job interview where some participants hide or distort information, interviewers are rewarded for hiring the honest, and where interviewees are rewarded for being hired. In the Poker Experiment subjects are asked to play a simple poker-like game against an adversary who has extra affective or game state information. These experiments extend existing work on ethical implications of polygraphs by considering variables (e.g. context or power relationships) other than recognition rate and using systems where information is completely mediated by computers. In all three experiments it is hypothesized that participants using systems that sense and transmit affective information to an adversary will have degraded performance and significantly different ethical evaluations than those using comparable systems that do not sense or transmit affective information. Analysis of the results of these experiments shows a complex situation in which the context of using affective computing systems bears heavily on reports dealing with ethical implications. The contribution of this thesis is these novel experiments that solicit participant opinion about ethical implications of actual affective computing systems and dimensional metaethics, a procedure for anticipating ethical problems with affective computing systems.

Thesis Supervisor: Rosalind Picard

Title: Professor of Media Arts and Sciences


Table of Contents

Acknowledgments
1. Affective Computing Used Adversarially
Motivating Examples
Panopticon / Panemoticon
Hypotheses
Outline
2. What Is an Ethical System?
Ethical Reasoning Performed by Humans Concerning Computers
3. Methods Used Across Experiments
Overview
Participants
Recruitment
Participant Scheduling
Demographics
Apparatus
Experiment Location
Pressuremouse
HandWave
Experimental Affective System with MixedEmotions Display
Questionnaire Materials
4. Poker Experiment
Poker Experimental Design
Poker Results
Poker: Charity Gains
Poker: Charity Loses
Poker: Other Analyses
5. Interview Experiment
Interview Experimental Design
Interview Results
Interview: Control
Interview: Charity Gains
Interview: Charity Loses
Interview Experiment Discussion
6. Quiz Experiment
Quiz Experimental Design
Control
No Effect on Charity, Charity Gains, and Charity Loses Motivators
Quiz Results
Quiz: Control Motivator
Quiz: No Effect on Charity
Quiz: Charity Gains
Quiz: Charity Loses
Quiz Experiment Discussion
7. Discussion
Ecological Validity and the Expression of Opinion
Pooled Results
Same System But Different Evaluation
Social Dimensions and Dimensional Metaethics
Effects of Motivators
Revisiting Hypotheses
Performance and Ethical Reports
Ability to Interpret Affective Information
Gender Differences
The Employee Polygraph Protection Act
What Can We Conclude?
Future Work
Reference List
Appendix A
Experiment Registration System
Appendix B
Means and Standard Deviations of Experimental Data

List of Figures

1.1. Bentham's Panopticon (1787)
1.2. A modern prison design that mirrors the Panopticon
3.1. Histogram of study participant ages
3.2. Office used for subjects in sensor conditions or their control analogs
3.3. Office used for subjects in conditions receive affective sensor information or their control analogs
3.4. Pressuremouse prototype board
3.5. Cheesemouse visualization
3.6. Printed circuit board prototype
3.7. Current data acquisition board
3.8. Current mouse appearance
3.9. Current mouse sensor arrangement
3.10. Relax to Win: video game where relaxation determines the winner
3.11. HandWave, revision 5 electrical schematic
3.12. HandWave Bluetooth skin conductance sensor, orb prototype
3.13. HandWave Bluetooth skin conductance sensor, wrist prototype
3.14. HandWave Bluetooth skin conductance sensor, version used in experiments
3.15. MixedEmotions displays the sensor data to certain experiment participants
3.16. Introduction to questionnaire
3.17. Questionnaire
4.1. Initial page of the Poker Experiment
4.2. Sensor preparation information (disadvantage)
4.3. Your task (disadvantage, Charity Loses motivator)
4.4. Your task (advantage, same across two motivators)
4.5. Your task - sensors (advantage)
4.6. Your task - sensors (advantage)
4.7. Dealer's initial screen
4.8. Non-dealer's initial screen
4.9. Your bet screen
4.10. Bet called screen
4.11. Shuffling deck screen
4.12. Dealer's initial screen for actual game
4.13. Bet screen for dealer's opponent
4.14. Fold screen for dealer's opponent
4.15. Follow up question
4.16. Final page
4.17. With sensors that collect information about emotion vs. task for Charity Gains motivator
4.18. With sensors question
4.19. Performance of disadvantaged players with sensors vs. their advantaged opponents for Charity Loses motivator
5.1. Initial page of the Interview Experiment
5.2. Sensor preparation information (interviewee)
5.3. Your task (interviewee - Charity Loses motivator)
5.4. Your task (interviewer)
5.5. Your task - sensors (interviewer)
5.6. Your task - sensors (interviewer)
5.7. Your task (interviewer)
5.8. Interview questionnaire (situation)
5.9. Follow up questions
5.10. Final page for all subjects who were not hired
5.11. Ethical vs. interview role for Control motivator
5.12. Ethical question
5.13. Respectful vs. interview role for Control motivator
5.14. Respectful question
5.15. Uncomfortable vs. interview role for Control motivator
5.16. Uncomfortable question
5.17. Immoral vs. interview role for Control motivator
5.18. Immoral question
5.19. Performance vs. sensors for Control motivator
5.20. Uncomfortable vs. interview role for Charity Gains motivator
5.21. Uncomfortable question
5.22. Immoral vs. interview role for Charity Gains motivator
5.23. Immoral question
5.24. Unfair vs. interview role for Charity Gains motivator
5.25. Unfair question
5.26. Ethical vs. sensors for Charity Loses motivator
5.27. Trustful vs. interview role for Charity Loses motivator
5.28. Trustful question
5.29. With or without sensors vs. condition for Charity Loses motivator
5.30. With sensors question
5.31. With sensors that collect information about emotion vs. interview role for Charity Loses motivator
5.32. With sensors question
6.1. One of the 30 squares used in the quiz experiment. Participants were asked to find the pair of numbers that sum to 10
6.2. Initial page of the Quiz Experiment (tailored for each participant)
6.3. Initial questions for the Quiz Experiment for all conditions
6.4. Sensor preparation information
6.5. Instructions: practice quiz (Control motivator)
6.6. Instructions: 5 minutes are up
6.7. Practice quiz score (Control motivator)
6.8. Instructions: real quiz
6.9. Compensation
6.10. Complete the real quiz
6.11. Instructions: 5 minutes are up
6.12. Your score
6.13. Quiz questionnaire (situation)
6.14. Quiz questionnaire (questions)
6.15. Determining your compensation
6.16. Final page
6.17. Please shred your quiz
6.18. Compensation (No Effect on Charity, Charity Gains, and Charity Loses Motivators)
6.19. Ethical vs. sensors for quiz Control motivator
6.20. Ethical question
6.21. Respectful vs. sensors for quiz Control motivator
6.22. Respectful question
6.23. Immoral vs. sensors for quiz Control motivator
6.24. Immoral question
6.25. Trustful vs. sensors for quiz Control motivator
6.26. Trustful question
6.27. Respectful vs. sensors for Charity Gains motivator
6.28. Respectful question
6.29. Hindrance vs. sensors for Charity Gains motivator
6.30. Hindrance question
6.31. Immoral vs. sensors for Charity Gains motivator
6.32. Immoral question
6.33. Trustful vs. sensors for Charity Gains motivator
6.34. Trustful question
7.1. Histograms of no opinion vs. sensors, grouped by variable
2. The poster used to solicit subjects on the MIT campus.
3. Initial page of experiment registration system.
4. Demographics page of experiment registration system.
5. Calendar page of experiment registration system.
6. Reminder page of experiment registration system.

List of Tables

4.1. Poker conditions
4.2. Poker p-value summary: participants with sensors vs. paired opponents
4.3. Poker p-value summary: disadvantaged participants using sensors vs. disadvantaged participants with No Sensors or Visible Card
5.1. Interview conditions
5.2. Interviewees with sensors vs. paired interviewers p-value and r-value summary
5.3. Interviewees with sensors vs. interviewees without sensors p-value summary
5.4. Hiring rates of conditions in the Interview Experiment
6.1. Quiz conditions
6.2. Quiz compensation
6.3. Sensors vs. no sensors p-value and r-value summary
7.1. Social dimensions relevant to evaluation of systems that mediate the communication of affect (a non-exhaustive list)
7.2. Assumptions in the Poker Experiment
7.3. Assumptions in the Interview Experiment
7.4. Gender pooled sensors vs. no sensors p-value summary
5. Poker experiment: mean (μ) and standard deviation (σ) summary for disadvantaged participants using sensors vs. disadvantaged participants with No Sensors or Visible Card
6. Poker experiment: mean (μ) and standard deviation (σ) summary for participants with sensors vs. paired opponents
7. Interview experiment: mean (μ) and standard deviation (σ) summary for interviewees with sensors vs. interviewees without sensors
8. Interview experiment: mean (μ) and standard deviation (σ) summary for interviewees with sensors vs. paired interviewers
9. Quiz experiment: mean (μ) and standard deviation (σ) summary for sensors vs. no sensors