This document discusses regulating social media platforms and algorithms through interoperability requirements. It argues that self-regulation by technology companies is not sufficient and can permit issues like algorithmic discrimination. The document examines options for regulation, including: 1. Requiring dominant platforms to provide interoperability through open APIs to allow competitors access to user data and functionality. 2. Learning from regulations of other industries like telecoms that mandate interconnection. 3. Developing ethical standards and impact assessments for algorithms to address issues like bias, but notes these may not be suitable for all cases. Overall it argues a systematic approach is needed to ensure proper oversight and redress for users affected by automated systems.

1. Regulating social media
platforms for interoperability
SCL Annual Conference
2 October 2019
Professor Chris Marsden
University of Sussex School of Law

2. Prior art....

3. Very basic Machine Learning [ML]
already feared in 1933

4. HAL9000?
50 years ago

5. Largely governed through self-regulation
Technology giants appear set to persuade us that
self-regulation remains the only effective route to
legal accountability for machine learning systems,
jeopardising the sustainable introduction of smart
contracts,
permitting algorithmic discrimination and
compromising the implementation of privacy law.

6. 2017-18 Lowlights from AINow

7. Discriminatory data is likely to lead
to discriminatory results
Discriminatory algorithms
 as well as those not designed to filter out discrimination
can make those results more discriminatory
Justice requires that lawyers study algorithmic outcomes
in order to ascertain such discrimination,
which may be highly inefficient as well as
outrageous to natural justice and fundamental rights.

8. Public administration has generic
solutions
Administrative law
Natural justice –at least ‘reasonableness’
Right to explanation/remedy?
Discrimination law –
applies to corporate decisions
Specialist technology law
Biomedical/nanotech
Railways, roads, telecoms
Data Protection

10. Prosumer Law:

11. Council of Europe: to err is human,
inducing AI complexity does not absolve

12. Caveat: regulation may not be suitable,
appropriate or feasible for many algorithms
But for those that regulators have most
concern about in
sectors that provide the most sensitive
socioeconomic decisions,
it is a remedy that can be explored.

13. Sensitive public facing sectors?
Banking/Credit, Insurance,
Medical Care & Research,
Social Care,
Policing and Security,
Education,
Transport
AI-piloted Airliners &
Autonomous Vehicles,
Social media
Telecommunications.

14. Transparency and replicability are
not the solutions to AI/ML problems
Transparency first requirement of legal recourse
 (though some algorithms can be reverse engineered without
transparency “under the hood” of the machine).
 It is not sufficient, however, for several reasons.
 Claims that the ability to study an algorithm and its operation
 provides a remedy for users who suffer as result of decisions.

15. Things change!
Both the training data and the algorithm itself will change constantly
 e,g. impossible to forecast real time outcomes of Google searches
 vast SEO business attempts approximations without complete accuracy
Remedy that can be achieved is only replicability –
 taking an ‘old’ algorithm and its data at a previous point in time
 to demonstrate whether algorithm and data became discriminatory.
 Estimate just how incomplete a remedy by
 allowing effectively ‘slow motion replays’
 while the game is rushing onwards

16. Bad Place: Real Time Trolley Problem

17. AI regulation and 'ethics washing'
Undertaken by technology companies and their
professional advisors
to persuade policy makers that
self-regulation is the only effective route to legal
accountability for Machine Learning systems,
1. jeopardising the sustainable introduction of
smart contracts,
2. permitting algorithmic discrimination and
3. compromising implementation of data
protection law.

18. Regulators wise to these tricks HT
EU Data Protection Supervisor

19. Ethics washing will fail
Cursory research into
 history of communications regulation and
 Internet law
 demonstrates the falsity of this self-regulation proposition.
See:
 Marsden, C. (2018) “Prosumer Law and Network Platform Regulation: The Long
View Towards Creating Offdata”, 2 Georgetown Tech. L.R. 2, pp.376-398;
 Marsden, C. and T. Meyer (2019) Report for European Parliament: “The effects of
automated content recognition (ACR) technology-based disinformation initiatives on
freedom of expression and media pluralism”

20. Need for systematic redress
by external agency
Ben Wagner (2019) Liable, but Not in Control?
 Ensuring Meaningful Human Agency in Automated Decision-Making
Systems, Policy & Internet, Vol. 11, No. 1, 2019, 104-122 at
https://onlinelibrary.wiley.com/doi/pdf/10.1002/poi3.198
 Self-driving cars,
 police searches using social media/PNR,
 Facebook content moderation

21. Competition or comms/media regulation?

22. What can and should be done?
1. Ethical standards for all AI deployed in ‘wild’ – to public
1. ISO standards being formed, basic privacy/human rights impact
assessment
2. No mandated interoperability for public communications providers
– Instant Messaging/Search/Social Media companies
3. APIs opened to dominant (SMP) operators
 Based on Microsoft remedies in longest most expensive antitrust case in
EC history: case started in 1993 in US, EU 1998-2010
 Google case started 2009 – ongoing a decade later
Commission decision of 27 June 2017 Case AT.39740 - Google Search (shopping)

23. 1. Ethical standards for all AI
deployed in ‘wild’ – to public
ISO standards being formed
1. Can be quite powerful influencers c.f. ISO27001 on cybersecurity
2. Typically technical engineering realm not normative standards
3. Embedded in national laws can become weak coregulatory signal
Basic privacy/human rights impact assessment
1. Proposed by UN Rapporteur Prof. David Kaye
2. Also see ‘Regulating Code’ (Brown/Marsden)
3. AI impact assessment suggested by European Data Protection
Supervisor

24. Standards still important!
 Standards Australia chairing ISO Working Party:
 ISO/IEC JTC 1/SC 42 Artificial intelligence
 https://www.iso.org/committee/6794475.html
 Australian Computer Society AI Ethics Committee:
 https://www.acs.org.au/governance/ai-ethics-committee.html
 Data61 (Australian Commonwealth Scientific and Industrial Research
Organisation (CSIRO):
 Dawson D and Schleiger E*, Horton J, McLaughlin J, Robinson C∞, Quezada
G, Scowcroft J, and Hajkowicz S† (2019) Artificial Intelligence: Australia’s Ethics
Framework. Data61 CSIRO, https://data61.csiro.au/en/Our-Work/AI-Framework
 Greenleaf, Graham and Clarke, Roger and Lindsay, David F., (2019)
 Does AI Need Governance? The Potential Roles of a ‘Responsible Innovation
Organisation’ in Australia; Submission to the Human Rights Commissioner on
the White Paper Artificial Intelligence: Governance and Leadership
http://dx.doi.org/10.2139/ssrn.3346149
 UK Information Commissioner’s Office, Feedback request — profiling and
automated decision-making, 6 April 2017,
 https://ico.org.uk/media/about-the-ico/consultations/2013894/ico-feedback-
request-profiling-and-automated-decisionmaking.pdf

25. Interoperability as an algorithmic
regulatory remedy
Attempt to move beyond glances in the rear view mirror
Silicon Valley mantra is “move fast and break things”
To enforce access to dominant regulated company’s API
 Application Programme Interfaces
Enables brokers, comparator programmes, regulators
to access algorithms in real time & controlled conditions
to observe the algorithm’s behaviour.

26. 2. Interoperability option for
public communications providers
Instant Messaging/Search/Social Media companies
1. Not so radical – required for broadcasters and telcos
1. Electronic Programme Guides
2. Telephone numbering schemes
3. NOT interconnection – up to smaller Ims to decide how to comply
4. Co-regulatory standards
2. Not as utilities but as media providers
1. This is NOT common carrier regulation
2. Not equivalent to energy/postal providers
3. Not as publishers but as printers
1. Arguments on fake news/hate speech for another time
2. Attempts to impose ‘Duty of Care’ fiduciary in UK/US are highly inappropriate

27. MIT Tech Review summarizes

28. US policy consideration?

29. EU Commmissioner Vestager on
interoperability and large platforms
 3 June speech: “Competition and the Digital Economy”
 https://ec.europa.eu/commission/commissioners/2014-
2019/vestager/announcements/competition-and-digital-economy_en
 “Making sure that products made by one company will work
properly with those made by others –
 can be vital to keep markets open for competition.”
Microsoft’s takeover of LinkedIn approval depended on
 agreement to keep Office working properly,
 not just with LinkedIn,
 but also with other professional social networks.
“Commission will need to keep a close eye on strategies that
undermine interoperability”

30. Economist
Opinion
6th June

31. 3. Dominant (SMP) operators
API opened
If dominant –competition and consumer remedy
1. ACCC find dominance by Facebook & Google
2. Only applies to platform aspects of their business
1. i.e. iTunes not Apple phones
Microsoft remedies in longest most expensive
antitrust case in EC history - $5billion fines
1. Case started in 1993 in US, EU 1998-2014
1. Google case started 2009 – ongoing a decade later

32. Note this is not about the
advertising market (only a proxy)

33. Three models – proposed by
Brown/Marsden 2008, 2013
Model 1: Must-carry obligations
broadcasters & Electronic Programme Guides
Model 2: API disclosure requirements
Microsoft from EC rulings
 Case T-201/04, Microsoft v Commission, EU:T:2007:289, 1088
 Decision 24 May 2004 Case C-3/37792 Microsoft; Decision of
16 December 2009 in Case 39530 Microsoft (Tying)
Model 3: Interconnect requirements
Applied to telcos, especially with SMP

34. Interoperability? 3 Types
Protocol interoperability
ability of services/products to interconnect technically
usual interoperability in competition policy
Data interoperability
Recalling Mayer-Schonberger/Cukier
Slice of data to competitors
Full protocol interoperability
What telecoms often thinks of as full
interconnection

35. Why interoperate?
It’s the economics!
Mechanism for achieving any-to-any connectivity –
promotes innovation
There is nothing less valuable than a network with one user!
Interoperability results in increased value of networks
promotes efficient investment in/use of infrastructure
Essential for new entrants to compete with existing
operators on non-discriminatory basis promotes entry

36. Is this remedy more broadly applicable?
Banking/insurance/medical algorithmic ‘AI’?
Self-driving vehicles?
Depends on a variety of socio-economic factors
Many sectors have regulators working on
‘regulatory sandpit’ solutions
Interoperability extensively used in sectors with
which we are most familiar

37. Consumer Data Right?
Oz CDR to deliver open banking, open energy and open telecoms?
 Many Europeans – well, we few –very excited about CDR model
 UK Furman Review of Digital Markets: ‘data mobility’
 Competition and Markets Authority: Data, Technology & Analytics unit
 Innovation and Intelligence team: audit algorithms & research tech markets

38. Questions?

39.  Christopher Kuner, Fred H. Cate, Orla Lynskey, Christopher Millard, Nora Ni Loideain,
and Dan Jerker B. Svantesson, ‘Expanding the artificial intelligence-data protection
debate’ (2018) 8 (4) International Data Privacy Law, 289
 Sandra Wachter, Brent Mittelstadt and Luciano Floridi, ‘Why a Right to Explanation of
Automated Decision-Making Does Not Exist in the General Data Protection Regulation’
(2017) 7 (2) International Data Privacy Law 76;
 Sandra Wachter, Brent Mittelstadt, Chris Russell, ‘Counterfactual Explanations without
Opening the Black Box: Automated Decisions and the GDPR’ (2018) HarvardJL&Tech 1
 Andrew D. Selbst and Julia Powles, ‘Meaningful information and the right to
explanation’ (2017) 7 (4) International Data Privacy Law 233.
 Lilian Edwards, Michael Veale, ‘Slave to the algorithm? Why a ’right to an explanation’
is probably not the remedy you are looking for’ (2017) 16 (1) Duke Law & Technology
Review 18;
 Lilian Edwards, Michael Veale, ‘Enslaving the Algorithm: From a "Right to an
Explanation" to a "Right to Better Decisions”?’ (2018) 16 (3) IEEE Security & Privacy 46

 Lilian Edwards, Michael Veale, ‘Clarity, surprises, and further questions in the Article 29
Working Party draft guidance on automated decision-making and profiling’ (2018) 34 (2)
Computer Law & Security Review 398

40. 10 Steps towards Ethical AI
1. Transparency
 Geeks love this, it’s almost meaningless to average user
2. Explainability
 See above –more useful is replicability
3. Consent
 See GDPR on meaningful & ‘course of business’
4. Discrimination
 Garbage in/Garbage out
5. Accountability to Stakeholders
6. Portability
 Australia’s Consumer Data Right!
7. Redress and Appeal
8. Algorithmic Literacy
 See ‘how to programme your VCR’
9. Independent oversight
10. Governance
 Hosanagar advocates for the creation of an independent Algorithmic
Safety Board, modeled on the Federal Reserve Board
 https://www.vox.com/the-highlight/2019/5/22/18273284/ai-algorithmic-bill-of-
rights-accountability-transparency-consent-bias