Image source: cornell.edu clicdata.com
University of Toronto Department of Computer Science and Institute
CSC311 Embedded Ethics Module
Copyright By PowCoder代写 加微信 powcoder
Introduction to Embedded Ethics (Online)
• The goal of this module is not to tell you what to think about ethical issues, but to make you more comfortable in identifying and discussing them.
• Feel free to use the chat! We want to hear your thoughts and reactions, and will do our best to respond to a few of them in real time.
• Be respectful, but don’t hesitate to disagree with one another. In the chat and breakout groups, address your comments to the person’s views or arguments, rather than the person themselves.
The Facebook Papers
In October, a number of internal Facebook documents were made public by a whistleblower named .
These documents showed that Facebook was aware of many of the ethically dubious consequences of their social media platforms.
Source: https://edition.cnn.com/2021/10/25/tech/facebook-papers/index.html
Source: bloomberg.com 3
The Ethics of Recommender Systems
• Our goal will be to introduce you to some ethical concepts that might be relevant to thinking about what lessons can be drawn from the Facebook Papers – and about the ethics of recommender systems in general.
• Note that this is different from the legality of these issues!
Issue 1: Content Moderation
“Facebook researchers documented how its platform has contributed to divisive, inter- religious conflict in India, according to internal records.
Employees flagged that human traffickers in the Middle East used the site to lure women into abusive employment situations. They warned that armed groups in Ethiopia used the site to incite violence against ethnic minorities. They sent alerts to their bosses about organ selling, pornography and government action against political dissent, according to the documents. They also show the company’s response, which in many instances is inadequate or nothing at all.” (Source: Wall Street Journal, https://www.wsj.com/articles/the-facebook-files- 11631713039)
Issue 1: Content Moderation
To maximize
{ Clicks } Viewing time Engagement
Choose which of these to present to the user
Issue 1: Content Moderation
To maximize
{ Clicks } Viewing time Engagement
Choose which of these to present to the user, when to present them, in what order….
The problem of content moderation:
are there any posts that should not be
presented under any circumstances?
Discussion Question
Instructions for building an untraceable assault rifle
False information claiming that COVID vaccines cause autism
Which of these posts, if any, should be prohibited on a social media platform?
A call to “take back” Parliament through force
Mentimeter poll
In the chat (or raise your hand to speak):
What principle (if any) guided your judgments about which content to prohibit?
Discussion Question
Issue 1: Content Moderation
Mill’s Harm Principle
“The object of this Essay is to assert one very simple principle, as entitled to govern absolutely the dealings of society with the individual in the way of compulsion and control, whether the means used be physical force in the form of legal penalties, or the moral coercion of public opinion. That principle is, that the sole end for which mankind are warranted, individually or collectively, in interfering with the liberty of action of any of their number, is self- protection. That the only purpose for which power can be rightfully exercised over any member of a civilised community, against his will, is to prevent harm to others.” (Mill, On Liberty)
Issue 1: Content Moderation
Mill thought that the Harm Principle implied very strong protections for speech:
“If all mankind minus one were of one opinion, and only one person were of the contrary opinion, mankind would be no more justified in silencing that one person, than he, if he had the power, would be justified in silencing mankind.” (Mill, 18)
Issue 1: Content Moderation
Mill’s argument for this conclusion rests on the importance of free speech and debate in coming to know the truth:
“If the opinion is right, they are deprived of the opportunity of exchanging error for truth: if wrong, they lose, what is almost as great a benefit, the clearer perception and livelier impression of truth, produced by its collision with error.” (Mill, 18)
He thinks that speech should only be prohibited in cases where it results in a direct harm to a specific person (e.g. inciting a crowd to harm them, or lying about them in a way that harms their reputation)
Discussion Question
In the chat (or raise your hand to speak):
Going back to our three examples:
1. Gunmanufacturinginstructions 2. False COVID information
3. Acallto“takeback”Parliament
How compelling do you find Mill’s argument for each of these three examples?
Discussion Question
In the chat (or raise your hand to speak):
Let’s now assume that there are no failures of content moderation.
What other ethical issues might arise with recommender systems?
Issue 2: The Feed
• Recommender systems don’t just give users the opportunity to look up certain content they are interested in.
• They also choose which content to amplify (by posting it more prominently or more often).
To maximize
{ Clicks } Viewing time Engagement Logins
……. ……. Choose between….
Issue 2: The Feed
comments and emotion emojis, which
were viewed as more meaningful than
Wall Street Journal, “Facebook Tried to Make Its Platform a Healthier Place. It Got Angrier Instead”
• In deciding which posts to present to users, Facebook has an explicit formula describing the relative weights of certain factors.
• Facebook introduced this formula in order to drive more meaningful interactions.
• “The goal of the algorithm change was to reverse the decline in comments, and other forms of engagement, and to encourage more original posting. It would reward posts that garnered more
likes, the documents show.”
Issue 2: The Feed
“While the FB platform offers people the opportunity to connect, share and engage, an
, often before we can catch it and mitigate its effects,”
he wrote. “Political operatives and publishers tell us that they rely more on negativity and sensationalism for distribution due to recent algorithmic changes that favor reshares.” (Internal Facebook Memo, quoted by the Wall Street Journal)
unfortunate side effect is
that harmful and misinformative
content can go viral
Breakout Group Exercise
We will look at four cases showing different ways that social media platforms might influence their users.
For each case presented:
1) Discuss each case as a group. Rank them from most ethical to least ethical. Submit one answer per group via google form, attempting to capture the consensus in your group. [10 minutes]
Google LINK:
1) We will then take a vote on how unethical you think the influence is, on a scale from 1 (raises no ethical problems) to 5 (very unethical, to the point that it should be legally prohibited).
Mentimeter
Breakout Group Exercise
1. A recommender algorithm that presents users with advertisements similar to the posts they have clicked on in the past.
2. A recommender system that presents advertisements for expensive luxuries when it judges that a user is stressed or feels like a failure, based upon the language of their posts, messages and click history.
3. A recommender system that presents users with posts that many would find offensive, in order to determine whether the user would like to see that content (like a bandit algorithm), but does not show them again unless the user engages with them.
1. A recommender system that presents users with offensive posts based upon its estimate of the gender, race, sexuality, etc of the user.
The Ethics of Recommender Systems
Average value on a scale of 1 (no ethical problems) to 5 (highly unethical)
A recommender algorithm that presents users with advertisements similar to the posts they have clicked on in the past.
A recommender system that chooses advertisements based on whether it judges that a user is stressed or feels like a failure.
A recommender system that presents users with shocking content in order to determine whether the user would like to see that content (like a bandit algorithm).
A recommender system that presents users pictures of cute animals over and over again, to the exclusion of other content, when the user has a compulsion to look at pictures of cute animals.
Issue 2: The Feed
• One concept that might help make sense of your intuitions on the previous questions is manipulation.
Example 1: Conditioning
Conditioning is an attempt to get someone to adopt a pattern of behaviour by rewarding or punishing their actions.
Example 2: The Guilt Trip
• A guilt trip is using an inappropriate amount of guilt to influence someone to do something.
Example 3: Gaslighting
• Gaslighting is an attempt to get someone to do something by (falsely) persuading them that their judgment is generally flawed or even delusional.
Defining Manipulation
• Adefinitionofmanipulationwouldexplain what all of these cases have in common with each other.
• Inthechat:Whatdothepreviousthree examples have in common with each other that make them count as ‘manipulation’?
Defining Manipulation
• Whatdotheseactionshaveincommonwith each other that make them count as ‘manipulation’?
• Onetheory:manipulativeactionisthe intentional attempt to get someone’s beliefs, desires, or emotions to violate their norms or ideals, from the perspective of the manipulator. ( , “Manipulative Actions: A Conceptual and Moral Analysis”)
Defining Manipulation
• Whatnormsoridealsguideourbeliefs, desires and emotions? Noggle:
• Beliefs:
“Believe only the truth.”
• Desires:
“Desire only what you judge that you have reason to desire.”
• Emotions:
“Base your emotions on true beliefs.”
“Ensure that emotions highlight only things that are genuinely relevant to your deliberations.”
Defining Manipulation
Does it violate the norms of
our beliefs, desires, emotions, or none of them?
Conditioning
The guilt trip
Gaslighting
A recommender system that chooses advertisements based on whether it judges that a user is stressed or feels like a failure.
A recommender system that presents users with shocking content in order to determine whether the user would like to see that content (like a bandit algorithm).
A recommender system that presents users pictures of cute animals over and over again, when the user has a compulsion to look at pictures of cute animals, to the exclusion of other content.
The Morality of Influencing Others
Three plausible options for the ethics of manipulation:
Intentional manipulation is always unethical.
Intentional manipulation is “prima facie” or “pro tanto” unethical.
• Manipulation is always unethical unless it is required to fulfill another moral duty
Intentional manipulation is not inherently unethical; it is only bad when it leads to bad outcomes.
Back to the Second Issue
The concepts discussed so far may help us to better understand our reactions to the second issue revealed by the Facebook Papers: the amplification of angry posts. Some possible reactions:
1. The amplification of angry content is ethical (or at least ethically permissible)
2. It is unethical primarily because of its bad consequences (polarization, violence, etc)
3. Itisunethicalprimarilybecauseitis manipulative.
4. It is unethical primarily for some
other reason.
, the writer of the best-known AI textbook, has hypothesized the following:
“Typically, such algorithms are designed to maximize click- through, that is, the probability that the user clicks on presented items. The solution is simply to present items that the user likes to click on, right? Wrong.
A more predictable user can be fed items that they are likely to click on, thereby generating more revenue. People with more extreme political views tend to be more predictable in which items they will click on…. Like any rational entity, the algorithm learns how to modify the state of its environment – in this case, the user’s mind – in order to maximize its own reward.” (Russell, Human Compatible, 8)
change the user’s preferences so that they become more
The solution is to
predictable.
Improving Recommender Systems?
• Consentmayleadtomoreethical recommender systems.
• Ifauserconsentstobeinfluencedinacertain way, perhaps that makes it more permissible, even if that influence involves manipulation.
• Legal requirements vs moral requirements
• Explicitconsent>implicitconsent
Improving Recommender Systems?
• Givinguserscontroloverwhattheyseein their feeds may also lead to more ethical recommender systems. (Stray, ‘Beyond Engagement’)
• E.g.‘seelessoften’or‘hidepost’functionsin feeds
Discussion Question
In the chat (or raise your hand to speak):
What sort of personal controls would you want to have over your feeds in the social
media platforms you use?
Improving Recommender Systems?
• Themostradicalchange:wemightdesign more ethical recommender systems by changing the objective function of a recommender system – what it is that the recommender system is trying to maximize. (Stray, ‘Beyond Engagement’)
• Well-beingmetrics
• IEEE7010
Other Resources
• Talks and events at:
• Centre for Ethics
• Institute
• Courses in:
• Philosophy
• History and Philosophy of Science and Technology
• Ethics, Society and Law
• Faculty of Information
• CSC 300: Computer Science and Society
• And stay tuned for more Embedded Ethics in Computer Science modules!
Acknowledgements
This module was created as part of an Embedded Ethics Education Initiative (E3I), a joint project between the Department of Computer Science1 and the Institute for Technology and Society2, University of Toronto.
Instructional Team:
, , Clure
Faculty Advisors:
1, 1, and Sheila McIlraith1,2
Department of Computer Science
Institute for Technology and Society University of Toronto
References
• Mill, . On Liberty. Accessed online: https://socialsciences.mcmaster.ca/econ/ugcm/3ll3/mill/ liberty.pdf
• Noggle, Robert. “Manipulative Actions: A Conceptual and Moral Analysis”, American Philosophical Quarterly 33(1), p.43-55
• Russell, Stuart. Human Compatible. : Viking Press, 2019
• , “Beyond Engagement” Accessed online: https://partnershiponai.org/beyond-engagement- aligning-algorithmic-recommendations-with-prosocial- goals/
程序代写 CS代考 加微信: powcoder QQ: 1823890830 Email: powcoder@163.com