Privacy and Rationality:Theory and Evidence

Alessandro Acquisti

Heinz School, Carnegie Mellon Universityacquisti@andrew.cmu.edu

Who should protect your privacy?

It is true that there are potential costs of using Gmail for email storage […] The question is whether consumers should have the right to make that choice and balance the tradeoffs, or whether it will be preemptively denied to them by privacy fundamentalists out to deny consumers that choice. -- Declan McCullagh (2004)

Privacy attitudes vs. behavior…

• Attitudes: usage– Top reason for not going online (Harris [2001]). 78% would

increase Internet usage given more privacy (Harris [2001])

• Attitudes: shopping– $18 billion in lost e-tail sales (Jupiter [2001]). Reason for 61% of

Internet users to avoid ECommerce (P&AB [2001]). 73% would shop more online with guarantee for privacy (Harris [2001])

• Behavior– Anecdotic evidence: DNA for BigMac– Experiments: Spiekermann, Grossklags, and Berendt (2001):

privacy “advocates” & cameras– Everyday examples: Dot com deathbed

Explanations

• Syverson (2003)– “Rational, after all” explanation

• Shostack (2003)– “When it matters” explanation

• Vila, Greenstadt, and Molnar (2003)– “Lemon market” explanation

• Are there other explanations?– Acquisti and Grossklags (2003): privacy and

rationality

Personal informationis a very peculiar economic good…

• Subjective • Ex-post• Context-dependent• Asymmetric• Both private and public good aspects

– Lump sum vs. negative annuity

• Buy value vs. sell value vs. expected loss • … privacy issues actually originate from two different

markets– Market for personal information– Market for privacy

Privacy and rationality

• Traditional economic view: forward looking agent, utility maximizer, bayesian updater, perfectly informed– Both in theoretical works on privacy– And in empirical studies

• Exceptions: PEW Survey 2000, Annenberg Survey 2003

Yet: privacy trade-offs

• Protect:– Immediate costs or loss of immediate benefits– Future (uncertain) benefits

• Do not protect:– Immediate benefits– Future (uncertain) costs

(sometimes, the reverse may be true)

Why is this problematic?

• Incomplete information• Bounded rationality• Psychological/behavioral distortions

Theory: Acquisti and Grossklags WEIS 04 Acquisti ACM EC 04

Empirical approach: Acquisti and Grossklags WEIS 04

1. Incomplete information

• What information has the individual access to when she takes privacy sensitive decisions? – For instance, is she aware of privacy invasions and associated

probability and magnitude of risks? – Is she aware of benefits she may miss by protecting her personal

data?– What is her knowledge of the existence and characteristics of

protective technologies?

• Privacy:– Asymmetric information

• Exacerbating: e.g., RFID, GPS

– Material and immaterial costs and benefits– Uncertainty vs. risk, ex post evaluations

2. Bounded rationality

• Is the individual able to consider all the parameters relevant to her choice? – Or is she limited by bounded rationality?

– Herbert Simon’s “mental models” (or shortcuts)

• Privacy:– Decisions must be based on several stochastic assessments

and intricate “anonymity sets”

– Inability to process all the stochastic information related to risks and probabilities of events leading to privacy costs and benefits

– E.g., HIPAA

– E.g., pervasive computing

3. Psychological/behavioral distortions

• Privacy and deviations from rationality– Optimism bias– Complacency towards large risks– Inability to deal with prolonged accumulation of small

risks– Coherent arbitrariness– “Hot/cold” theory– Attitude (generic) vs. behavior (specific)

• Fishbein and Ajzen. Attitude, Intention and Behavior: An Introduction to Theory and Research, 1975

– Hyperbolic discounting, immediate gratification

Hyperbolic discounting and Privacy

• Hyperbolic discounting (Laibson 94; O’Donoghue and Rabin 01)• Procrastination, immediate gratification• We do not discount future events in time-

consistent way• Compare:

• Privacy attitude in survey versus actual behavior• Privacy behavior when risks are more or less

temporally close

Hyperbolic discounting

Survey time vs. decision time

Survey time vs. decision time

Time consistency vs. time inconsistency

The big picture

Privacy protection

Costs

Marginal costs of privacy protection

Expected costs of privacy intrusions

Sum of costs

Survey and experiment

• Phase One: pilot• Phase Two: ~100 questions, 119 subjects from CMU

list. – Paid, online survey (CMU Berkman Fund)– Goals: check for

– Privacy attitudes / behavior

– Information, bounded rationality, and psychological distortions

• Phase Three: experiment

Clusters

• Multivariate clustering techniques (k-means)

• Privacy attitudes– privacy fundamentalists 26.1%; two medium groups (concerned about online 23.5% or

offline 20.2% identity); low concerns 27.7%

• Self reported behavior of privacy relevance– high degree of information revelation and risk exposure 64.7%; low revelation and

exposure 35.3%

• Knowledge of privacy risks – average knowledge of privacy threats 46.2%; high unawareness 31.9%; “aware” 21.9%

• Knowledge of privacy protection and security

– small group very knowledgeable 31.7%; larger group showing a blatant lack of

awareness 68.3%

Economic factors (excerpts)

• Preliminary evidence of hyperbolic discounting, bounded rationality, risk aversion

Indirect example of bounded rationality (excerpts)

You have completed a credit card purchase with an online merchant. Besides you and the merchant website, who has data about parts of your transaction? Nobody: 36.4% Credit card company: 18.7% Hackers: 15%

“Nobody, assuming an SSL transaction, without which I would not commit an online transaction using my credit card”

Behavior (excerpts)

• 74% adopted some strategy or technology or otherwise took some particular action to protect their privacy

– Encryption, PGP– Do-not-call list– Interrupt purchase– Provide fake information– […]

• However, when you look at details, percentages go down…

– 8% encrypt emails regularly– Similar results for shredders, do-not-call lists, caller-IDs, etc.

Conclusions

• Theory– Time inconsistencies may lead to under-protection and over-release of

personal information. Genuinely privacy concerned individuals may end up not protecting their privacy

• Evidence– Sophisticate attitudes, and (somewhat) sophisticate behavior – But also: evidence of overconfidence, incorrect assessment of own

behavior, incomplete information about risks and protection, buy/sell dichotomy.

• Implications– Rationality model not appropriate to describe individual privacy

behavior– Self-regulation alone, or reliance on technology and user responsibility

alone, may not work