Image courtesy Unsplash / @WilhelmGunkel
Week 9/S1/2022
Transparency:
Decisions & Processes
Copyright By PowCoder代写 加微信 powcoder
Marc of Computing and Information Systems Centre for AI & Digital Ethics
The University of Melbourne
marc.cheong [at] unimelb.edu.au
Bonus Links – in class discussion (Thanks to Maddie for compiling these)
Virtual NFT of Doge for those who answered questions☺https://wallpapercave.com/w/wp4914323 —–
Booth (2014) ”Facebook reveals news feed experiment to control emotions”. The Guardian. https://www.theguardian.com/technology/2014/jun/29/facebook-users-emotions-news-feeds
Lomas (2021). “Google misled consumers over location data settings, Australia court finds”. TechCrunch. https://techcrunch.com/2021/04/16/google-misled-consumers-over-location-data-settings-australia-court- finds/
Meineck (2020). “Six Reasons Why Google Maps Is the Creepiest App On Your Phone”. VICE News Germany. https://www.vice.com/en/article/3an84b/six-reasons-why-google-maps-is-the-creepiest-app-on-your-phone
Duffy (2019). “Google agrees to pay $13 million in Street View privacy case”. CNN Business. https://edition.cnn.com/2019/07/22/tech/google-street-view-privacy-lawsuit-settlement/index.html
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 2
Learning Outcomes
1. Distinguish between transparency and explainability, closely-related concepts in AI ethics.
2. Understand how automated decision-making (ADM) systems require transparency at every stage.
3. Understand how tech companies approach the issue of transparency, as well as concerns that are raised by stakeholders, in two areas where AI systems are deployed: social media ads and criminal justice.
4. Understand how big data research can – either positively or negatively – affect their data subjects, and why transparency is important when conducting such studies.
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 3
Related Reading
This module has two readings corresponding to the two broad themes within (plus an optional study). Recap: screenshot below ☺
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Transparency: what’s it all about.
Automated decision-making (ADM) systems: transparency at every stage?
Current issues in transparency and stakeholders’ concerns, social media advertising systems
criminal justice AI systems
Data science research and transparency: the cases of Cambridge Analytica and Covid19 Trends.
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 5
Transparency: What’s it all about?
Image courtesy Unsplash / @WilhelmGunkel
Audience activity I [~5 mins]
Facilitator: Head Tutor, Maddie.
Online: please use Canvas Chat to share your ideas.
In-person: chat with your neighbour, then share your views with the class. Incentive:
Image source: Cadbury
What is transparency?
What does it mean to you?
What does it mean when you say ‘this AI system does things transparently’?
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
A note on nomenclature (and focus)
Some academics use the term transparency to describe the “inner workings” of algorithms (e.g. Loi et al, 2020: https://link.springer.com/article/10.1007/s10676-020-09564-w)
To clarify, the focus of this module is on the overarching processes and decision-making of an algorithmic system.
We include the technical considerations about interpretability of the algorithms themselves, counterfactual analysis, etc under the banner of explainability.
(Our very own is an expert in this field).
Image source: Eliiza
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Transparency? (1/2)
Source: UNESCO news release, quoting : https://en.unesco.org/news/privacy-expert-argues-algorithmic-
transparency-crucial-online-freedoms-unesco-knowledge-cafe 9
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Transparency? (2/2)
Jake Goldenfein, ‘Algorithmic Transparency and Decision-Making Accountability: Thoughts for buying machine learning algorithms’ in Office of the Victorian Information Commissioner (ed), Closer to the Machine: Technical, Social, and Legal aspects of AI (2019). https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3445873
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 10
Example: Hypothetical Thought Experiment. Deep Learning and Your Grades!
“What if, for this unit, we decide your final grade based on a state-of-the art deep learning-based estimator/predictor that will come up with a final grade based on 1,000 factors – ranging from how many Youtube videos you watch about ethics, to your activity in weekly discussion activities, to your typing speed when asked to comment on discussion boards, to your on-Zoom reactions to Simon and the tutors when they taught you the basics of ethics, to how much you laugh at Marc’s internet memes in the modules, etc…
The final mark predictor is so advanced, it has been audited by NASA, Google, and by 10 winners! However, since some of the tensor-based algorithms used are proprietary to NVidia, who sponsored the array of Geforce RTX3090s used for deep learning, they can’t be revealed in public.
Also, the decision made by the predictor is final.”
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 11
Audience activity II [5-10 mins]
Facilitator: Head Tutor, Maddie.
Online: please use Canvas Chat to share your ideas.
In-person: chat with your neighbour, then share your views with the class. Incentive:
Image source: Cadbury
What is wrong with this thought experiment?
“What if, for this unit, we decide your final grade based on a state-of-the art deep learning-based estimator/predictor that will come up with a final grade based on 1,000 factors – ranging from how many Youtube videos you watch about ethics, to your activity in weekly discussion activities, to your typing speed when asked to comment on discussion boards, to your on-Zoom reactions to Simon and the tutors when they taught you the basics of ethics, to how much you laugh at Marc’s internet memes in the modules, etc…
The final mark predictor is so advanced, it has been audited by NASA, Google, and by 10 winners! However, since some of the tensor- based algorithms used are proprietary to NVidia, who sponsored the array of Geforce RTX3090s used for deep learning, they can’t be revealed in public. Also, the decision made by the predictor is final.”
Suggested further discussion: Duty ethics as a teacher? Virtue ethics?
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Reflection.
Process: who governs the selection of ML models/training data? What vendors are given preference – e.g. why Google (not AWS)? e.g why Tensorflow and Nvidia?
Any conflicts of interest? Any feedback loops?
Why is the whole system shrouded in secrecy?
Who audited it? Can I see the source code/design schematics/rationales? Did I even sign up for this?
Decisions: can we challenge them? Can I take this to court? Who sanctioned this to be official?
Is this another Cambridge Analytica?
Image source: HowToGeek / Imagentle/Shutterstock
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 13
Automated decision-making (ADM) systems: Transparency at every stage?
Image courtesy Unsplash / @WilhelmGunkel
Disclaimer: I am not a lawyer
The information provided in this mini-lecture is summarized from various sources to explain how transparency is a requirement not only for the algorithms, but also the contexts surrounding their implementation.
This lecture won’t make you an expert in administrative decision-making ☺
Image source: CAPCOM / Cinemablend. – COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Source: Administrative Review Council, Automated Assistance in Administrative Decision Making: Report to the Attorney General (Report No 46, November 2004) (‘2004 Report’)
https://www.ag.gov.au/legal-system/publications/report-46-automated-assistance-administrative-decision-making-2004
The 2004 Report was ahead of its time (emphases below are mine)
P27: “Safeguards built into the system are only asking relevant questions, telling customers why questions are being asked (which makes the decision-making process more transparent) and recording and explaining to a customer the reason for a decision”
P43: “Expert systems’ ability to provide an audit trail of the administrative decision-making processes they are involved in is important to the administrative law values of transparency, fairness and efficiency. “
P45: “A good system of internal review is one which is transparent in process and affords a quick, inexpensive and independent review of decisions. Such a system is beneficial both to applicants and
agencies”. (citing Administrative Review Council 2000)
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 16
, ‘Algorithms: transparency and accountability’, Phil. Trans. R. Soc. A.3762017035120170351.(2018). https://royalsocietypublishing.org/doi/10.1098/rsta.2017.0351
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 17
High-level view of AI systems (OECD, 2019)
Source: OECD (2019), Artificial Intelligence in Society, OECD Publishing, Paris, https://doi.org/10.1787/eedfee77-en.
Figure: https://www.oecd-ilibrary.org/sites/8b303b6f-en/index.html?itemId=/content/component/8b303b6f-en#figure-d1e976
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 18
Elements for ML systems
(Google Inc, from Scully et al 2015)
Source: https://cloud.google.com/solutions/machine-learning/mlops-continuous-delivery-and-automation-pipelines-in-machine-learning – adapted from Scully et al (2015) https://papers.nips.cc/paper/2015/file/86df7dcfd896fcaf2674f757a2463eba-Paper.pdf
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Audience activity III [5-10 mins]
Facilitator: Head Tutor, Maddie.
Online: please use Canvas Chat to share your ideas.
In-person: chat with your neighbour, then share your views with the class. Incentive:
Image source: Cadbury
All other parts of a typical ML system can be made more transparent. Can you comment on how?
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Reflection.
From our examples, the code (‘algorithm’) is only a small part of it!
Transparency is required in planning, implementation, auditing…
… concerns the data, design,
… also consider legal aspects & philosophical concepts in the broader sense: incl. fairness, recourse…
Image source: HowToGeek / Imagentle/Shutterstock
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Current issues in transparency:
Social media advertising systems
Image courtesy Unsplash / @WilhelmGunkel
Consider Facebook’s group of products
If we use Instagram, Facebook, WhatsApp, how many of us actually read the TOS?
Source: Scherker (2017) https://www.huffpost.co m/entry/facebook- terms- condition_n_5551965
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 23
☕️ Break time! See you in 5
Image courtesy Unsplash / @WilhelmGunkel
Audience activity IV [2-5 mins]
Facilitator: Head Tutor, Maddie.
Online: please use Canvas Chat to share your ideas.
In-person: chat with your neighbour, then share your views with the class. Incentive:
Image source: Cadbury
When printed out, can you guess how long Facebook or Twitter’s policies and terms of service are.
Let’s say default font, on A4 paper? —-
The big question: do you read them?
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Consider Facebook’s group of products
Roughly 9-10 pages in A4 printed (as of 2021).
Summarised by TOSDR (https://tosdr.org/en/service/182 ) – even the key points take up ~2 A4 pages;
refer image→
This doesn’t even include the additional policies: e.g. ‘Commercial Standards’, ‘Advertising Policies, … (~12 more links)
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 26
Let’s try Twitter?
Roughly 39 pages in A4 printed!!! (as of
2021). https://cdn.cms-twdigitalassets.com/content/dam/legal-
twitter/site-assets/tos-oct-14th- 2020/Twitter_User_Agreement_EN.pdf
Summarised by TOSDR (https://tosdr.org/en/service/195 ) – even the key points take up ~2 A4 pages;
refer image→
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 27
Consider ads: is this ‘transparent’?
Let’s say I’ve been shown an ad (sponsored post) on Facebook. As a consumer, I want to know WHY.
(Image sources: Facebook).
I go to the TOS, then find ‘Ads’, then find ‘Data Policy’…
Which takes me to another page… Which scares me.
Let’s try another method – go to the ad, and click on the menu.
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 28
Consider ads: is this ‘transparent’?
Let’s say I’ve been shown an ad (sponsored post) on Facebook. As a consumer, I want to know WHY. (Image sources: Facebook)
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 29
Consider ads: is this ‘transparent’?
Case study thanks to ,
( Institute for Human Development, Germany)
Let’s say I’ve been shown an ad (sponsored post) on Facebook. As a consumer, I want to know WHY.
FB gives me some reasons.
… but also “more factors not listed”.
… and directs me to a huge page with many explanations, but all “mays” (we may, advertisers may…)
… and has a long TOS explaining the many ways.
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 30
Consider ads: is this ‘transparent’?
Case study thanks to ,
( Institute for Human Development, Germany)
https://psyarxiv.com/ea28z/ – Lorenz-Spreen et al (2021)
“At present, the platforms’ transparency measures offer “nominal transparency”, with no real regard for whether people actually can easily access, read and gain insight into the information held about them and whether this transparency in name foster users’ autonomy.
“Aiming for effective transparency—which demonstrably enables users to understand what platforms do with their data and what users’ choices imply, and to then translate this knowledge into measurable behaviour—is an important step towards more acceptable business practices and towards regaining some of the lost autonomy for users (e.g., by prompting people to adjust their privacy settings; Parra-Arnau et al., 2017)
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Audience activity V [2-5 mins]
Facilitator: Head Tutor, Maddie.
Online: please use Canvas Chat to share your ideas.
In-person: chat with your neighbour, then share your views with the class. Incentive:
Image source: Cadbury
What are the reasons tech companies are reluctant to promote effective (as opposed to nominal) transparency?
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Current issues in transparency:
Criminal justice AI systems (and COMPAS revisited)
Image courtesy Unsplash / @WilhelmGunkel
Criminal Justice and AI systems
You have encountered the following examples in your exploration of AI ethics. (Images are from Wikipedia)
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Reading: Rudin (2019) on COMPAS
The focus of this case study is not about the technical explainability for the underlying algorithms, etc.
… but about the transparency of the processes and decisions involved.
Take COMPAS – and its decision making assumptions, in practice (Rudin 2019)
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 35
Reading: Rudin (2019) on COMPAS
Now, take COMPAS – and its choice of models, in practice (Rudin 2019)
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 36
Recall Audience Activity V we just had?
Asaro (2019) on Pred the same vein we consider COMPAS; cf PredPol in Asaro (2019)
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 37
Ferguson (2017) on Pred the same vein we consider COMPAS; cf. PredPol – Ferguson (2017) provides a legal perspective.
Issues at all stages: crime stats; data dossiers; personal/cultural bias; data entry/analysis; tech complexity; financial/IP interests; auditing; metrics…
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Reflection.
AI systems in justice/policing causes serious effects on people’s freedom and status under the law.
The legal perspective of AI ethics gives us another perspective on the need for transparency.
“Transparency is difficult, but it matters to a functioning predictive system that deals with individuals’ lives and liberty” (Ferguson, 2017).
How do we start fixing the issues?
E.g. for PredPol – auditing, public release of metrics, training (Ferguson, 2017).
…or not use PredPol to begin with?
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022 39
Big Data Research & Social Media: From Elections to Pandemics
Image courtesy Unsplash / @WilhelmGunkel
Reading: Walsh (2019)
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Reading: Walsh (2019)
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Not just Cambridge Analytica – which was for election targeting etc.
Walsh (2019) found that studies by academics “to improve voter participation” in fact “increased turnout by about 340,000 additional votes” (citing Bond et al, 2012).
Questions:
1. CA was bad, I’m sure you agree… 2. But for the 2nd experiment – isn’t this a good thing? Increasing voter participation = healthy democracy? 42
Reading: Walsh (2019)
Reflections for transparency in social media and big data research, by Walsh (2019):
1. “The first recommendation is that we may need to take into account not just the impact on the individual under study but the broader impact any experiment might have on society…”
2. “The second recommendation is that ethics approval may be needed…”
3. “The third recommendation is that subjects of any experiment may need to be informed directly after the study about the results and their participation…”
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Audience activity VI [5-10 mins]
Facilitator: Head Tutor, Maddie.
Online: please use Canvas Chat to share your ideas.
In-person: chat with your neighbour, then share your views with the class. Incentive:
Image source: Cadbury
Reflections for transparency in social media and big data research, by Walsh (2019):
1. “The first recommendation is that we may need to take into account not just the impact on the individual under study but the broader impact any experiment might have on society…”
2. “The second recommendation is that ethics approval may be needed…”
3. “The third recommendation is that subjects of any experiment may need to be informed directly after the study about the results and their participation…”
#3 – is this even feasible?
Is the trade-off between ‘not being able to inform everyone’ versus ‘let’s run the experiment anyway’ ethically okay?
– COMP90087 – Semester 1, 2022 – © University of Melbourne 2022
Reading: Walsh (2019)
Reflections for transparency in social media and big dat
程序代写 CS代考 加微信: powcoder QQ: 1823890830 Email: powcoder@163.com