Counteracting the negative effect of form auto-completion on the privacy calculus
1. Counteracting the negative effect of form
auto-completion on the privacy calculus
Bart Knijnenburg, Alfred Kobsa, Hongxia Jin
2. Form auto-completion: the bright side
Modern browsers offer an
auto-completion feature
that reduces the effort of
filling out web forms
Bart
2
3. Form auto-completion: the bright side
Modern browsers offer an
auto-completion feature
that reduces the effort of
filling out web forms
!
Imagine such tools could
fill out any form, on any
website
3
4. Form auto-completion: the bright side
Modern browsers offer an
auto-completion feature
that reduces the effort of
filling out web forms
!
Imagine such tools could
fill out any form, on any
website
!
This would be particularly
useful for mobile browsers
4
5. Form auto-completion: the dark side
Preibusch et al. (2012) warn that such auto-completion tools
may cause users to complete more form fields than they
intended
6. Form auto-completion: the dark side
Preibusch et al. (2012) warn that such auto-completion tools
may cause users to complete more form fields than they
intended
!
Auto-completion tools make it so easy to submit a fully
completed form that users may skip weighing the benefits
and risk of disclosing a certain piece of information in a
specific situation (privacy calculus!)
7. Form auto-completion: the dark side
Preibusch et al. (2012) warn that such auto-completion tools
may cause users to complete more form fields than they
intended
!
Auto-completion tools make it so easy to submit a fully
completed form that users may skip weighing the benefits
and risk of disclosing a certain piece of information in a
specific situation (privacy calculus!)
!
Can we overcome these problems with a better
auto-completion tool?
8. Research outline
Please tell us more about yourself
BlogHeroes will assign a "guild" to you based on the information you provide below. Note that none
of the fields are required, but our classification will be better if you provide more information.
General info about me
We introduce two new tools
that we hypothesize will
reinstate the privacy calculus
Please provide some background info to get our matching process started.
Name (first):
> For employers
E-mail address:
> For inves tors
Gender:
> Contac t Age (years):
!
Address:
> About us
City:
(last):
John
Smith
Please enter your information
john@smith.com
I WRK will find jobs based on the information you enter on this form.
Male
None of the items on the form are required, but if you provide more
23
information the jobs will be a better match.
123 Main St.
State:
New York
G ENERA L
NY
Zip:
12345
A ND C ONTA C T I NFO
Specifically, we compare three
What I do for a living
auto-completion tools:
General and contact information
Some guilds write about their jobs. Tell us more about yours, and we can provide a better match.
FIRST NAME
Employment status:
– Auto FormFiller (automatically fills
fields, users can remove manually)
– Remove FormFiller (same but
users can click to remove eachMy health
field)
– Add FormFiller (no automatic
filling, users can click a button to fill
each field)
Experience (years):
Employed for wages
John
LAST NAME
clear
Smith
Enter your details, please
5
AGE
Your personal Codacare health insurance policy will be based on the
23
Current/previous job: information you provide. Please note that Education the items are
Researcher
Sector:
none of / training / library
required, but the insurance will be better tailored to your needs if you
Income level:
between $50K and $100K/year
GENDER
provide more information.
Education:
clear
Male
Doctoral
General information
E-MAIL ADDRESS
Please provide your general information.
john@smith.com
Name (first):
clear
(last):
Some guilds write about their health. Providing us with some info will help us match them to you.
ADDRESS
CITY
STATE
ZIP
Physical health:
Dietary restrictions:
Birth control usage:
clear
123 Main St.
About average
Address:
allergic to nuts
City:
ORK EXPERI ENC E
None
W
Gender:
Please
New York
State:
NY
12345
Zip:
tell us about your education and work experience, so that we
can find a suitable job for you.
Age:
HIGHEST DEGREE EARNED
Doctoral
E-‐mail:
fill
clear
fill
fill
fill
8
clear
9. Research outline
People base th
eir information
disclosure dec
the perceived
ision on
risk and relev
ance of the inf
ormation
e)
thus disclosur
(and
st
nd relevance
rceived risk a
y of the reque
Pe
cit
rpose-specifi
e pu
depend on th
The effect of ris
k and relevance
is moderated b
type: disclosure
y tool
is more “calcula
ted” when the a
remove tools ar
dd and
e used
15. What causes information disclosure?
Typical answer: privacy calculus (Laufer and Wolfe 1977)
15
16. What causes information disclosure?
Typical answer: privacy calculus (Laufer and Wolfe 1977)
!
Privacy calculus is like utility maximization (Li 2012): people
trade off the different aspects and then choose the option
that maximizes their utility (Bettman et al. 1998; Simon
1959)
16
17. What causes information disclosure?
Typical answer: privacy calculus (Laufer and Wolfe 1977)
!
Privacy calculus is like utility maximization (Li 2012): people
trade off the different aspects and then choose the option
that maximizes their utility (Bettman et al. 1998; Simon
1959)
!
So what are these aspects that people trade off?
17
18. What causes information disclosure?
Risk
!
Operationalization:
Providing my [item] to [site] is:
(-3 = very risky; +3 = very safe)
Risk
Disclosure
!
!
Relevance
Relevance
!
Operationalization:
The fact that [site] asked for my [item] was:
(-3 = very inappropriate; +3 very appropriate)
18
21. But what causes risk and relevance?
On web forms, users selectively disclose different types of
information in different extent to different types of websites
(Hsu 2006)
22. But what causes risk and relevance?
On web forms, users selectively disclose different types of
information in different extent to different types of websites
(Hsu 2006)
!
Social media users also share selectively with others,
depending on the purpose of the information (Olson et al.
2005)
23. But what causes risk and relevance?
On web forms, users selectively disclose different types of
information in different extent to different types of websites
(Hsu 2006)
!
Social media users also share selectively with others,
depending on the purpose of the information (Olson et al.
2005)
!
Does purpose-specificity play a role in commercial privacy as
well?
24. But what causes risk and relevance?
In our experiment, participants:
– entered a wide range of info into an auto-completion tool
– tested the tool on one of three websites
Create a Profile
Please create your profile by entering your information below.
Note that FormFiller will store the information locally on your device, and only for the duration of
this study. We will never submit any forms automatically or disclose this information to others
without your active involvement.
Please tell us more about yourself
BlogHeroes will assign a "guild" to you based on the information you provide below. Note that none
of the fields are required, but our classification will be better if you provide more information.
General info about me
Please provide some background info to get our matching process started.
Name (first):
John
E-mail address:
About you:
john@smith.com
Gender:
(last):
Age (years):
First name:
State:
NY
123 Main St.
City:
Gender:
23
Address:
Last name:
Smith
Male
New York
Zip:
12345
What I do for a living
Some guilds write about their jobs. Tell us more about yours, and we can provide a better match.
Age:
> For employers
Employment status:
Experience (years):
> For inves tors
Current/previous job:
Address:
City:
E-‐‑mail:
Phone:
> Contac t
Income level:
State:
Zip:
Education:
> About us
My health
Employed for wages
5
Please enter your information
I WRK will find jobs based on the information you enter on this form.
Sector:
Education / training library
None of the items on the form are required,/ but if you provide more
Researcher
between $50K andthe jobs will
information $100K/year
be a better match.
Doctoral
G ENERA L
A ND C ONTA C T I NFO
General and contact information
Some guilds write about their health. Providing us with some info will help us match them to you.
Physical health:
Dietary restrictions:
FIRST NAME
About average
LAST NAME
John
Smith
Enter your details, please
Birth control usage:
clear
allergic to nuts
AGE
None
23
Your personal Codacare health insurance policy will be based on the
GENDER
information you provide. Please note that none of the items are
Male
required, but the insurance will be better tailored to your needs if you
E-MAIL ADDRESS
provide more information.
john@smith.com
clear
clear
clear
25. But what causes risk and relevance?
In our experiment, participants:
– entered a wide range of info into an auto-completion tool
– tested the tool on one of three websites
Enter your details, please
Your personal Codacare health insurance policy will be based on the
26. But what causes risk and relevance?
In our experiment, participants:
– entered a wide range of info into an auto-completion tool
– tested the tool on one of three websites
Websites correspond to a particular type of info:
– blogging community ⋍ personal interest items
– job search website ⋍ job skills items
– health insurer ⋍ health record items
Enter your details, please
Your personal Codacare health insurance policy will be based on the
27. But what causes risk and relevance?
Website
Item type
Risk
Relevance
Purpose-specificity will look
like an interaction effect
between Website and Item
type
!
!
When the type of information
requested matches the
purpose of the website:
– perceived risk will be lower
– perceived relevance will be higher
– people will be more likely to
disclose the item
28. But what causes risk and relevance?
Website
Item type
Risk
Relevance
Purpose-specificity will look
like an interaction effect
between Website and Item
type
!
!
Disclosure
When the type of information
requested matches the
purpose of the website:
– perceived risk will be lower
– perceived relevance will be higher
– people will be more likely to
disclose the item
29. But what causes risk and relevance?
Model tested with 543 Mechanical Turk participants
Website
Risk
see next
slide
Item type
odds: 0.818***
Disclosure
Relevance
odds: 1.079***
2
χ (12) = 11.929, p = .451; CFI = 1.00, TLI = 1.00; RMSEA < .001, 90% CI: [.000, .011]
29
30. But what causes risk and relevance?
Counteracting the negative privacy effect of form auto-completi
Counteracting the negative privacy effect of form auto-completi
Counteracting
negative privacy effect of form auto-complet
Contact info
Contact info
Contact info
33
3
Interests
Interests
Interests
Perceived risk
Perceived risk
Perceived risk
100%
100%
100%
95%
95%
95%
90%
90%
90%
Perceived relevance
85%
85%
85%
80%
80%
80%
22
2
Job skills
Health record
record
Health record
Job skills
3
3
Perceived relevance
Perceived relevance
2
2
11
1
1
1
0
00
#
#
#
#
#
0
0
-1
-1-1
#
#
-1
-1
-2
-2
-2
-3
-3
-3
"
"
"
BlogHeroes
BlogHeroes
BlogHeroes
"
"
"
I♡WRK
I♡WRK
I♡WRK
"
"
Codacare
Codacare
Codacare
-2
-2
-3
-3
-3
BlogHeroes
BlogHeroes
BlogHeroes
I♡WRK
I♡WRK
I♡WRK
Codacare
Codacare
Codacare
Figure 3. Perceived Risk and Perceived Relevance per Website and Item type. The
Figure 3. Perceived Risk and Perceived Relevance
Figure 3. Perceived Risk and Perceived Relevance per Website and Item type. The
and Item type.
When pointtype of item matches the purpose of±the website, The
the to the matching item types. Error bars are ± 1 Standard Error.
arrows point to the matching item types. Error bars are 1 Standard Error.
arrows
arrows point to the matching item types. Error bars are ± 1 Standard Error.
people perceive lower risk and higher relevance
2
egarding perceived Risk, we find that the interaction between Website and Item type is significant (χ2 2
egarding perceived Risk, we find that the interaction between Website and Item type is significant (χ (
garding perceived Risk, we find that the interaction between Website and Item type is significant (χ (
246.41, p < .0001). Moreover, for each website, the non-matching item types are perceived
246.41,
.0001). Moreover, for each website, the non-matching item types are perceived
246.41, pp << .0001). Moreover, for each website, the non-matching item types are perceived
gnificantly more risky than the matching item type. Also, for each item type, the non-matching websit
gnificantly more risky than the matching item type. Also, for each item type, the non-matching websi
nificantly more risky than the matching item type. Also, for each item type, the non-matching websit
31. But what causes risk and relevance?
Counteracting the negative privacy effect of form auto-completion
Contact info
3
Interests
3
Perceived risk
100%
95%
90%
85%
80%
2
2
1
1
0
Website
-3
"
BlogHeroes
"
I♡WRK
Health record
#
Perceived relevance
#
#
0
Risk
-1
-2
Job skills
-1
"
-2
-3
Codacare
Counteracting the negative privacy effect of form auto-completion
Contact info
100%
BlogHeroes
Interests
I♡WRK
!
Job skills
Health record
Codacare
!
95%
!
Figure 3. Perceived Risk and Perceived Relevance per 90%
Website and Item type. The
arrows point to the matching item types. Error bars are ± 1 Standard Error.
85%
Contact info
3
Perceived risk
100%
95%
90%
85%
80%
2
1
Item type
0
-3
"
BlogHeroes
Interests
Job perceived Risk, we find that the interaction between Website and Item type is significant (χ2(6)
Health record
80%
Regarding skills
= 246.41, p < Perceived relevanceeach website, the non-matching item types are perceived as
.0001). Moreover, for
3
#
75%
significantly more risky than the matching item type. Also, for each item type, the non-matching websites
#
BlogHeroes
I♡WRK
Codacare
2
#
are perceived as significantly more risky than the matching website13. H3 is thus supported.
1
Likewise, the interaction between Website and Item type is also significant for Perceived Relevance (χ2(6)
0
= 913.47, p < .0001). For each website, the non-matching item types are perceived as significantly less
relevant than the matching item type, and for each item type, the non-matching websites are perceived as
-1
significantly less relevant than the matching website. H4 is thus supported as well.
-1
-2
"
I♡WRK
"
-2
H5+H6. Tool type ! Disclosure
-3
Codacare
BlogHeroes
I♡WRK
Codacare
The last row in Table 2 presents the effect of Tool type on Disclosure. Controlling for Perceived Risk,
Perceived Relevance, and Item type, the odds of Disclosure are 6.3% higher for users of the Add tool than
Figure 3. Perceived Risk and Perceived Relevance per Website and Item this effect not significant (p = .631). Surprisingly then, H5 is
for users of the Remove tool, although type. The
arrows point to the matching itemrejected:Erroris no significantStandard in Disclosure between the Add and Remove tools.
types. there bars are ± 1 difference Error.
The planned contrast between the traditional Auto tool and the alternative tools is significant (χ2(1) =
Regarding perceived Risk, we find that the interaction between Website and Item type is significant (χ2(6)
4.037, p = .045). Controlling for Perceived Risk, Perceived Relevance, and Item type, the odds of
= 246.41, p < .0001). Moreover, for each website, the non-matching item types are perceived as
Disclosure are 24.0% lower for users of the Remove tool compared to users of the Auto tool, a small (d =
significantly more risky than the matching item type. Also, for each item type, the non-matching websites
.165) but significant (p = .047) effect. The odds of Disclosure are 18.9% lower for users of the Add tool
are perceived as significantly more risky than the matching website13. H3 is thus supported.
compared to users of the Auto tool, a small (d = .107) effect that is not significant (p = .130). H6 is thus
partially is also significant for is indeed Relevance (χ2(6)
Likewise, the interaction between Website and Item typesupported: Disclosure Perceivedsignificantly higher for the Auto tool than for the Remove tool.
= 913.47, p < .0001). For each website, the non-matching item types are perceived as significantly less
32. But what causes risk and relevance?
When the type of information requested matches the purpose
of the website, people are more likely to disclose it.
Contact info
Interests
Job skills
100%
!
95%
!
Health record
!
90%
85%
80%
75%
BlogHeroes
I♡WRK
Codacare
32
34. ure?
ation disclos
auses inform
What c
perceived risk
and relevance
e?
and relevanc
es risk
ut what caus
B
purpose-spec
ificity
ence
at is the influ
Wh
tion
auto-comple
of the
tool?
35. Influence of tool type
Please tell us more about yourself
BlogHeroes will assign a "guild" to you based on the information you provide below. Note that none
of the fields are required, but our classification will be better if you provide more information.
General info about me
We tested three tools:
Please provide some background info to get our matching process started.
Name (first):
> For employers
E-mail address:
> For inves tors
Gender:
John
(last):
Smith
Please enter your information
john@smith.com
– Auto FormFiller (automatically fills
I WRK will find jobs based on the information you enter on this form.
None of the items on the form are required, but if you provide more
fields, users can remove manually)
information the jobs will be a better match.
– Remove FormFiller (same but users
G
can click to remove each field)
What I do for a living
General and contact information
– Add FormFiller (no automatic filling,
Enter your details, please
users can click a button to fill each
Your personal Codacare health insurance policy will be based on the
field)
information you provide. Please note that none of the items are
> Contac t Age (years):
Address:
> About us
City:
Male
23
123 Main St.
New York
State:
NY
Zip:
12345
ENERA L A ND C ONTA C T I NFO
Some guilds write about their jobs. Tell us more about yours, and we can provide a better match.
FIRST NAME
Employment status:
Experience (years):
LAST NAME
Employed for wages
John
5
AGE
Current/previous job:
Income level:
Education:
23
Researcher
clear
Smith
Sector:
Education / training / library
clear
required, but the insurance will be better tailored to your needs if you
between $50K and $100K/year
GENDER
provide more information.
clear
Male
Doctoral
General information
E-MAIL ADDRESS
My health
Please provide your general information.
john@smith.com
Name (first):
clear
(last):
Some guilds write about their health. Providing us with some info will help us match them to you.
ADDRESS
CITY
STATE
ZIP
Physical health:
Dietary restrictions:
Birth control usage:
123 Main St.
About average
Address:
allergic to nuts
City:
ORK EXPERI ENC E
None
W
Gender:
Please
New York
State:
NY
12345
Zip:
tell us about your education and work experience, so that we
can find a suitable job for you.
fill
clear
fill
fill
Age:
HIGHEST DEGREE EARNED
Doctoral
E-‐mail:
CURRENT EMPLOYMENT STATUS
Employed for wages
fill
clear
fill
clear
36. Influence of tool type
Tool type
Website
Risk
Disclosure
Item type
Relevance
2
χ (86) = 99.443, p = .152; CFI = .990, TLI = .988; RMSEA = .007, 90% CI: [.000, .013]
2
χ (86) = 95.140, p = .235; CFI = .993, TLI = .992; RMSEA = .006, 90% CI: [.000, .012]
37. Influence of tool type
!
Tool type
Auto
0.863***
Remove
0.754***
!
Add
0.811***
!
Disclosure
!
Risk
!
!
!
Relevance
Auto
0.989
Remove
1.133***
Add
1.114***
Odds are closest to 1 for the Auto tool, so Risk and Relevance
influence Disclosure the least for users of the Auto tool
38. Influence of tool type
Results:
– Disclosure was not purpose-specific for users of the Auto tool
– Disclosure was purpose-specific for users of the Remove and Add tools.
These tools help users consider the website’s purpose in their disclosure
decisions
– Additional note: Users of the Add tool were more satisfied, despite
having to click more frequently on average!
Contact info
100%
!
Auto
100%
!
!
95%
90%
Job skills
Health record
Remove
!
!
95%
Interests
!
!
Add
!
!
!
!
!
90%
85%
85%
80%
80%
75%
BlogHeroes
I♡WRK
Codacare
BlogHeroes
I♡WRK
Codacare
BlogHeroes
I♡WRK
Codacare
39. ure?
ation disclos
auses inform
What c
perceived risk
and relevance
e?
and relevanc
es risk
ut what caus
B
purpose-spec
ificity
ence
at is the influ
Wh
moderates effe
tion
auto-comple
of the
ct of risk and re
levance
tool?
40. Take-home message
People base th
eir information
disclosure dec
the perceived
ision on
risk and relev
ance of the inf
ormation
e)
thus disclosur
(and
st
nd relevance
rceived risk a
y of the reque
Pe
cit
rpose-specifi
e pu
depend on th
The effect of ris
k and relevance
is moderated b
type: disclosure
y tool
is more “calcula
ted” when the a
remove tools ar
dd and
e used