Regulating Disinformation
with Artificial Intelligence (AI)
Prof. Chris Marsden (Sussex) &
Dr Trisha Meyer
(IES, Free University of Brussels VUB)
United Nations Internet Governance Forum
29 November 2019

Defining disinformation
“False inaccurate or misleading information
designed, presented and promoted to
intentionally cause public harm or for profit”
In line with European Commission High Level Expert Group
We distinguish disinformation from misinformation,
which refers to unintentionally false or inaccurate information.

Defining Automated Content
Recognition (ACR)
 Within Machine Learning techniques that are advancing towards AI,
 ACR technologies are textual and audio-visual analysis programmes
that are algorithmically trained to identify potential ‘bot’ accounts and
unusual potential disinformation material.
ACR refers to both
 the use of automated techniques in the recognition and
 the moderation of content and accounts to assist human judgement.
Moderating content at scale requires ACR to supplement human
editing

ACR to detect disinformation is
prone to false negatives/positives
 due to the difficulty of parsing multiple, complex, and possibly
conflicting meanings emerging from text.
 Inadequate for natural language processing & audiovisual
 including so-called ‘deep fakes’
 (fraudulent representation of individuals in video),
 ACR has more reported success in identifying ‘bot’ accounts.
 We use ‘AI’ to refer to ACR technologies.

Disinformation a rapidly moving target
 We analysed 250 articles, papers and reports
 strengths and weaknesses of those focussed on AI disinformation
solutions on freedom of expression, media pluralism & democracy
 We agree with other experts: evidence of harm is still inconclusive
 2016 US Presidential election/UK ‘Brexit’ referendum
 Investigated US Department of Justice and UK Parliamentary Committee

Fake news?

International Grand Committee on
Disinformation (“Fake News”)
 Leopoldo Moreau, Chair, Freedom of Expression Commission, Chamber of
Deputies, Argentina,
 Nele Lijnen, member, Committee on Infrastructure, Communications and
Public Enterprises, Parliament of Belgium,
 Alessandro Molon, Member of the Chamber of Deputies, Brazil,
 Bob Zimmer, Chair, and Nathaniel Erskine-Smith and Charlie Angus, Vice-
Chairs, Standing Committee on Access to Information, Privacy and Ethics,
House of Commons, Canada,
 Catherine Morin-Desailly, Chair, Standing Committee on Culture, Education
and Media, French Senate,
 Hildegarde Naughton, Chair, and Eamon Ryan, member, Joint Committee on
Communications, Climate Action and Environment, Parliament of Ireland,
 Dr Inese Lībiņa-Egnere, Deputy Speaker, Parliament of Latvia,
 Pritam Singh, Edwin Tong and Sun Xueling, members, Select Committee on
Deliberate Online Falsehoods, Parliament of Singapore
 Damian Collins, Chair, DCMS Select Committee, House of Commons

Gigantic wave of Brexit ads in last
week (electioneering suspended)

Result

FBK/Cambridge Analytica via N.Ireland

The Fog of War?
Carl von Clausewitz (1832) Vom Kriege
 “War is the realm of uncertainty; three quarters of the factors on which
action in war is based are wrapped in a fog of greater or lesser uncertainty
 A sensitive and discriminating judgment is called for;
a skilled intelligence to scent out the truth.”
 Russian Hybrid Warfare: A Study of Disinformation
 Flemming Splidsboel Hansen, 2017, Danish Institute for International Studies
 Marsden, C. [2014] Hyper-power and private monopoly: the unholy marriage of
(neo) corporatism and the imperial surveillance state, Critical Studies in Media
Communication Vol.31 Issue 2 pp.100-108
http://www.tandfonline.com/doi/full/10.1080/15295036.2014.913805
 Marsden, C. [2004] Hyperglobalized Individuals: the Internet, globalization,
freedom and terrorism 6 Foresight 3 at 128-140

Neil Mohan made Jordan Peterson
& Yaxley-Lennon famous
 https://twitter.com/heyneil
“Google Paid This Man $100 Million: Here's His Story”
https://www.businessinsider.com/neal-mohan-
googles-100-million-man-2013-4?r=US&IR=T
“the visionary who predicted how brand advertising would fund
the Internet, turned this vision into a plan, and then executed it”

Methodology: Literature &
Elite Interviews
Project consists of a literature review, expert interviews and mapping of policy
and technology initiatives on disinformation in the European Union.
 150 regulatory documents and research papers/reports
 10 expert interviews in August-October;
 took part in several expert seminars, including:
 Annenberg-Oxford Media Policy Summer Institute, Jesus College, Oxford, 7 Aug;
 Google-Oxford Internet Leadership Academy, Oxford Internet Institute, 5 Sept;
 Gikii’18 at the University of Vienna, Austria, 13-14 Sept;
 Microsoft Cloud Computing Research Consortium, St John’s Cambridge,17-18 Sept
 We thank all interview respondent and participants for the enlightening
disinformation discussions; all errors remain our own.
 Internet regulatory experts:
 socio-legal scholar with a background in law/economics of mass communications;
 media scholar with a background in Internet policy processes and copyright reform;
 reviewed by a computer scientist with a background in Internet regulation and
fundamental human rights.
 We suggest that this is the bare minimum of interdisciplinary expertise
required to study the regulation of disinformation on social media.

Interdisciplinary study analyses implications
of AI disinformation initiatives
Policy options based on literature, 10 expert interviews & mapping
We warn against technocentric optimism as a solution to
disinformation,
 that proposes use of automated detection, (de)prioritisation, blocking
and removal by online intermediaries without human intervention.
 More independent, transparent and effective appeal and oversight
mechanisms are necessary in order to minimise inevitable
inaccuracies

What can AI do to stop disinformation?
Bot accounts identified:
 Facebook removed 1.3billion in 6 months
Facebook’s AI “ultimately removes 38% of hate speech-related posts”
 it doesn’t have enough training data to be effective except English-Portuguese
Trained algorithmic detection of fact verification may never be as effective
as human intervention:
 each has accuracy of 76%
 Future work might want to explore how hybrid decision models consisting
of fact verification and data-driven machine learning can be integrated
 Koebler, J., and Cox, J. (23 Aug 2018) ‘The Impossible Job: Inside Facebook’s
Struggle to Moderate Two Billion People’, Motherboard,
https://motherboard.vice.com/en_us/article/xwk9zd/how-facebook-content-
moderation-works

Zuckerberg on AI & disinformation
Some categories of harmful content are easier for AI to identify
and in others it takes more time to train our systems.
Visual problems, like identifying nudity, are often easier than
Nuanced linguistic challenges, like hate speech
 Zuckerberg, M. (15 Nov 2018) ‘A Blueprint for Content Governance and Enforcement’,
Facebook Notes, https://www.facebook.com/notes/mark-zuckerberg/a-blueprint-for-
content-governance-and-enforcement/10156443129621634/

Legislators should not push this
difficult judgment exercise onto
online intermediaries
 Restrictions to freedom of expression must be
 provided by law, legitimate and
 proven necessary and as
 the least restrictive means to pursue the aim.
 The illegality of disinformation should be proven before
filtering or blocking is deemed suitable.
 Human rights laws paramount in maintaining freedom online

AI is not a silver bullet
 Automated technologies are limited in their accuracy,
 especially for expression where cultural or contextual cues necessary
 Imperative that legislators consider which measures
 may provide a bulwark against disinformation
 without introducing AI-generated censorship of European citizens

Different aspects of disinformation
merit different types of regulation
All proposed policy solutions stress the
importance of
literacy and
cybersecurity

Five recommendations
1. Media literacy and user choice
2. Strong human review and appeal processes
where AI is used
3. Independent appeal and audit of platforms
4. Standardizing notice and appeal procedures
Creating a multistakeholder body for appeals
5. Transparency in AI disinformation techniques

1. Disinformation is best tackled
through media pluralism/literacy
These allow diversity of expression and choice.
Source transparency indicators are preferable over
(de)prioritisation of disinformation,
Users need the opportunity to understand
how search results or social media feeds are built
and make changes where desirable.

2. Advise against regulatory action
to encourage increased use of AI
for content moderation (CoMo) purposes,
without strong independent
human review
and
appeal processes.

3. Recommend independent appeal
and audit of platforms’ regulation
Introduced as soon as feasible.
Technical intermediaries moderation of
content & accounts
1. detailed and transparent policies,
2. notice and appeal procedures, and
3. regular reports are crucial.
Valid for automated removals as well.

4. Standardizing notice and appeal
procedures and reporting
creating self- or co-regulatory multistakeholder body
UN Special Rapporteur’s suggested “social media council”
Multi-stakeholder body could have competence
to deal with industry-wide appeals
better understanding & minimisation of effects of AI
on freedom of expression and media pluralism.

5. Lack of independent evidence or
detailed research in this policy area
means the risk of harm remains far too high
for any degree of policy or regulatory certainty.
Greater transparency must be introduced
into AI and disinformation reduction techniques
used by online platforms and content providers.

Specific Policy Options:
The Co-Regulatory Triangle

Option and
form of
regulation
Typology of regulation Implications/Notes
0 Status quo Corporate Social
Responsibility, single-
company initiatives
Note that enforcement of the new General Data Protection
Regulation and the proposed revised ePrivacy Regulation, plus
agreed text for new AVMS Directive, would all continue and likely
expand
1 Non-audited
self-
regulation
Industry code of practice,
transparency reports, self-
reporting
Corporate agreement on principles for common technical
solutions and Santa Clara Principles
2 Audited self-
regulation
European Code of Practice of
Sept 2018; Global Network
Initiative published audit
reports
Open interoperable publicly available standard e.g. commonly
engineered/designed standard for content removal to which
platforms could certify compliance
3 Formal self-
regulator
Powers to expel non-
performing members,
Dispute Resolution
ruling/arbitration on cases
Commonly engineered standard for content filtering or
algorithmic moderation. Requirement for members of self-
regulatory body to conform to standard or prove equivalence.
Particular focus on content ‘Put Back’ metrics and
efficiency/effectiveness of appeal process
4 Co-
regulation
Industry code approved by
Parliament or regulator(s)
with statutory powers to
supplant
Government-approved technical standard – for filtering or other
forms of moderation. Examples from broadcast and advertising
regulation
5 Statutory
regulation
Formal regulation - tribunal
with judicial review
National Regulatory Agencies – though note many overlapping
powers between agencies on e.g. freedom of expression,

Prior art....

Questions?