Ethics of personalized information
filtering
Ansgar Koene, Elvira Perez, Christopher J. Carter,
Ramona Statache, Svenja Adolphs, Claire O’Malley, Tom
Rodden, and Derek McAuley
HORIZON Digital Economy Research,
University of Nottingham
Overview
•
•
•
•
Public/private data
Privacy: expressed concerns vs. expressed behaviour
Interim summary
Conditions for consent
Information Overload!
Estimated data
production in 2012
(www.domo.com)
Good information service = good filtering
Information services, e.g. internet search, news feeds etc.
• free-to-use => no competition on price
• lots of results => no competition on quantity
• Competition on quality of service
• Quality = relevance
= appropriate filtering
Why personalized filtering?
John and Jane average have:
2.43 children
0.47 dogs & 0.46 cats
0.67 houses & 0.73 cars
• John and Jane average do not exists
• Results based on population averages are crude approximations
• Personalized filtering – a natural step in the evolution of
information services
Personalized filter/recommender
systems
• Content based – similarity to past results the user liked
• Collaborative – results that similar users liked
(people with statistically similar tastes/interests)
• Community based – results that people in the same social
network liked
(people who are linked on a social network e.g. ‘friends’)
Concerns regarding personalization
• Social consequences: self-reinforcing information filtering
– the ‘filter bubble’ effect
• Privacy: personalization involves profiling of individual
behaviour/interests
• Agency: the filtering algorithm decides which segment of
available information the user gets to see
• Manipulation: people’s actions/choices are depend on the
information they are exposed to
User profiling: privacy
User profiling involves mining of data about:
• past behaviour of the user interacting with the service
• user behaviour on other services
o through ‘tracking cookies’
o data purchasing from other services
• mapping the social network of a user and monitoring
the behaviour of people within that social network
User profiling: (un)informed consent
Informed consent for profile building:
- Part of long, difficult to understand, Term & Conditions that
users click ‘accept’ on, usually without reading it.
- Same consent is applied for years without explicit renewal
User profiling: security issues
The profile summarizes user behaviour patterns
its purpose is to predict the interests of the user
Access to this information can facilitate:
- Phishing
- Social engineering for hacking
Agency: user vs. algorithm
Filter algorithms provide competitive advantage details about
them are often trade-secrets
• Users don’t know how the information they are presented
with was selected  no real informed consent
• Service users have no ‘manual’ override for the settings of
the information filtering algorithms
• It is difficult for service users to know which information
they don’t know about because it was filtered
Manipulation: conflict of interest
Information filtering, or ranking, implicitly manipulates choice
behaviour.
Many online information services are ‘free-to-use’, the
service is paid for by adverting revenue, not users directly
 Potential conflict of interest:
promote advertisement vs. match user interests
Advertising inherently tries to manipulate consumer behaviour
Personalized filtering can also be use for political spin /
propaganda etc.
Evidence of public concern

2011 FTC investigation of Goolge for search
bias

EU competition regulation vs Google

Netflix prize competition de-anonymization
Manipulation: conflicts of interest
Is privacy sensitive, need to know how it is handled
Role for regulating authoroty, but also:
Tools to probe filtering criteria -> black-box testing
User-friendly testing kit for general public -> RRI -> so people
can decide for themsleves if they are happy with a service
Conclusion





Personalized information filtering is a natural evolution in the
interaction with the user
It raises issues relating to privacy and data protection.
Lack of transparency -> concerns over agency & manipulation
Potential for covert manipulation
RRI -> researchers developing recommender algorithms have
responsibility
identifying and studying the socio-psychological impact of personalized
filtering;
helping people to understand and regulate the level of privacy intrusion they
are willing to accept for personalized information filtering;
developing a methodology to probe the subjective ‘validity’ of the information
that is provided to users based on their own interests;
engaging with corporate information service providers to reinforce ethical
practices.
Call for research programme
Technical development of tools:
Black-box testing kit for probing the characteristics of the user behavior profiles used in recommender systems.
Recommendation bias detection system for identifying user behavior manipulation
A two-layer recommender architecture that de-couples the delivery of non-personalized information by service
providers from a user owned/controlled system for personalized ranking of the information.
Psycho-social research on the impact of personalized information filtering on:
General exploration-exploitation trade-off in action selection
Attitudes towards trust and critical evaluation of information
Cybersecurity:
Protection against mal-use of personalized recommender systems for phishing related social engineering
Policy:
Development of guidelines for responsible innovation and use of recommender systems, protecting the privacy and
freedom of access to information of users.
Public engagement:
Develop educational material to help people understand how recommendations they receive from search engines, and
other recommender systems, are filtered so that they can better evaluate the information they receive.
two layer system


Data collections by service provider
Filtering by user
Acquisti et al. (2009)