Keynote presentation at 1st International Workshop on
Disinformation and Toxic Content Analysis
(DiTox 2023) on the problem of onine disinformation and associated technnologies and policies that help against it. This work was co-funded by the EC in the context of the MedDMO project (contract number 101083756)
Integration and Automation in Practice: CI/CD in Mule Integration and Automat...
Disinformation challenges tools and techniques to deal or live with it
1. This project has received funding from the European Union
DIGITAL-2021-TRUST-01. Contract number: 101083756
meddmo.eu
Disinformation
challenges, tools and techniques to deal or live with it
Nikos Sarris
Media analysis, veri
fi
cation and retrieval group (MeVer)
Center for Research and Technology Hellas (CERTH)
2. Body Level One
Body Level Two
Body Level
Three
Body Level
Four
Body
Level
Five
What happens is posted,
but has what’s posted really happened?
huge masses of information are produced
by huge numbers of sources and delivered to huge audiences.
Some is bound to be fake...
the
challenge
How can we help people
distinguish truth from lies?
3. Disinformation
Misinformation
Malinformation
Information that is false and deliberately
created to harm a person, social group,
organization or country.
Information that is false, but not created with
the intention of causing harm.
Information that is based on reality, used to
inflict harm on a person, organization or
country.
INFORMATION DISORDER: Toward an interdisciplinary framework for research and policy making
Claire Wardle, Hossein Derakhshan
Council of Europe report DGI(2017)09
4. …as old as time
A challenge…
high cost in both cases
{disinformation}
{misinformation}
5. much harder today
•the rapid growth of social media and the Internet has made it easier for false
information to spread quickly and widely
•people are often more likely to share or believe false information that confirms
their preexisting beliefs or opinions (confirmation bias)
•the rise of generative AI (e.g. deepfakes, ChatGPT) is making it increasingly
difficult to distinguish between real and fake
•political polarisation and the erosion of trust in traditional media sources has
created an environment in which disinformation can thrive
•disinformation campaigns are often carried out by governments, political
organisations, and other powerful groups with an agenda to manipulate public
opinion and undermine democratic processes
6. the role of social media
•Amplification: algorithms, that actually reward users for sharing content, can
amplify false information, making it more visible and easier to spread.
•Speed of dissemination: false information to spread quickly and widely
•Anonymity: sources can spread false information anonymously, making it
harder to hold them accountable
•Viral Sensationalism: False information that is sensational or attention-
grabbing is more likely to go viral
•Filter bubbles: users are commonly exposed to information that aligns with
their existing beliefs and opinions, making it harder to challenge false
information (debatable)
7. comes in many different shapes
•Misleading headlines attract clicks and generate traffic, even if the content
of the article doesn't support the headline
•Hoaxes and conspiracies: Spread of false information with the intent to
deceive, often involving politically motivated or sensational claims.
•Fabricated stories with the intention of misleading people.
•Satire and parody: although created for satirical purposes can be
misinterpreted
•Political propaganda: to promote a political agenda or discredit opponents.
•Manipulated media: more attractive, persuasive and tempting to share
8. getting harder as we speak
AI synthetic generation of text and visuals
(combined?) is pushing the challenge to a next level
Image synthetically generated by DALLE-2
Most text of this presentation up to this slide was generated by ChatGPT
9. the impact can be very high
•Misleading the public: the spread false or misleading information can
impact individuals' decision-making and beliefs.
•Undermining democracy: false or misleading information can be used to
manipulate elections and undermine trust in democratic institutions
•Exacerbating conflicts: disinformation can fuel social and political
tensions, leading to increased conflict and violence
•Damaging reputations: the distribution of various rumors and claims is
frequently used to harm the reputation of individuals and organizations
•Impairing public health: conspiracy theories and dangerous
misinformation, like non-scientifically based medical advice about
causes, symptoms, and treatments
12. Contributor
Content
Context
What can we find about the source of information?
Does everything contextualise together?
Does the posted content look reliable?
The CCC model
13. Contributor:
Who says?
1
2
3
4
5
Reputation: what do people think of this source?
History: what is the past activity of this source?
Presence: where does this source exist?
Influence: what happens because of this source?
Popularity: who follows this source?
14. Content:
Sounds real?
1
2
3
4
5
Quality: what is the text/visual/audio style like?
Popularity: what is the social interaction with it?
Authenticity: has it been manipulated/
synthetically generated?
History: can it be found in past publications?
Reputation: how is it referenced by others?
15. Context:
Does it stick
together?
1
2
3
4
5
Cross-check: are there any similar reports?
Diversity: are there multiple coherent reports?
Provenance: how has this travelled through time?
Influence: what happens because of this?
Proximity: do source locations relate to events?
16. The ΑΜΙ model
INFORMATION DISORDER: Toward an interdisciplinary framework for research and policy making
Claire Wardle, Hossein Derakhshan
Council of Europe report DGI(2017)09
18. The BLOC model
Nwala, A.C., Flammini, A. & Menczer, F. A language framework for modeling social media account behavior. EPJ Data Sci. 12, 33 (2023).
https://doi.org/10.1140/epjds/s13688-023-00410-9
A language framework for modelling social media account behaviours
consisting of symbols representing user actions and content
Action alphabet
T:: Post message
P:: Reply to friend
p:: Reply to non-friend
π:: Reply to own post
R:: Reshare friend’s post
r:: Reshare non-friend’s post
ρ:: Reshare own post
Content alphabet
t:: Text
H:: Hashtag
M:: Mention of friend
m:: Mention of non-friend
q:: Quote of other’s post
ϕ:: Quote of own post
E:: Media object
U:: link (URL)
19. Modeling
(dis)information
Various models have been proposed
The signals of credibility are many
The most signals are addressed the better the assessment
A wholistic view appears to be more constructive than a unit
21. Assessment of
each credibility
signal
Overall assessment
of post credibility
Overall
assessment of
each pillar of the
model
Assessment of trustworthiness
Rule-based assessment of the credibility signals following the CCC model
Further information
provided for each
pillar of the model
22. Evolution towards transparency
Model evolved to more intuitive pillars:
Activity - Network - Influence
Automatic assessment was removed
A lot more data was added to aid the users
into making their own assessments
23.
24. Visual heatmaps
help users quickly
understand the
results
15 image analysis
algorithms help
uncover possible
forgeries
Heatmap overlay
on image helps
users spot
manipulated areas
Image Verification Assistant
Quickly analyse images to identify possible forgeries
Manual annotation
space helps users
discuss findings
25. Temporal video
segmentation
illustrating the
manipulation
probability of every
segment
Top level analysis
for the entire video
Image/video
player window for
detailed viewing
Deepfake detection
Assess the possibility of deepfake manipulation in image or video
Detailed analysis
per video segment
26. Example images
from the most likely
location illustrating
visual similarity
with the query
image
Map illustrating the
most likely location
for the image
Window for
detailed image
viewing
Location estimation
Estimate the most likely location for an image based on visual cues
27. Objects and actions
are automatically
identified and
added as tags
Automatic image
analysis creates
short descriptions
Assets are
automatically
classified to
specific categories,
as disturbing,
NSFW or memes
Multimedia Archive
A platform where you can easily analyse, find and annotate media
Any tag can be
manually added by
users
28. Results include
images with flags
of visual similarity
to the query
window
Window for
selecting part of
image that search
should be based
on
Search by visual similarity
Find images that are visually and semantically similar
Selecting to search
for similar images
29. Results include
images that are
almost (visually)
identical to the
query image
Search by visual similarity
Find images that are visually nearly identical
Selecting to search
for near duplicate
images
30. Results include
video that includes
scenes that are
visually identical to
scenes from the
query video
Search by visual similarity (video)
Find videos that contain common scenes
Selecting to search
for near duplicate
video
Frames from the
retrieved video
Frames from the
query video
31. Checks Google and
Wolfram
Knowledge Bases
to return relevant
results
Fact-check a claim automatically
Automatic fact-checking by Claimbuster
Returns relevant
fact-checks using
the Google Fact-
Check Explorer
32. Converts audio
from an event into
text and then
searches for
matches among the
previously
published fact-
checks in the
ClaimReview
database
Fact-check a live claim automatically
Live, automated fact-checking during political events by Squash
Relevant matches
are chosen by
human editors
and displayed on
users’ screens
within seconds of
the politician
making the claim
33. Monitors Twitter
accounts and alerts
when a new tweet
is classified as
check-worthy
Spot check-worthy posts
Automatically notifying on check-worthy posts by ClaimHunter
Beltrán, Javier et al.
“ClaimHunter: An
Unattended Tool for
Automated Claim
Detection on
Twitter.”
KnOD@WWW
(2021).
34. Collaborative verification
• Provides ways to easily discover
content from various sources
• Organises information in real-time
shareable views
• Provides access to internal and
external archives
• Allows use of many 3rd party tools
• Easily exports findings to share or
publish stories
• Used in EDMO and hubs to easily
share their findings
39. Detecting
(dis)information
Various tools are available but none is the silver bullet
Many more can be built, but still not the silver bullet
Combining the findings from many tools is important
Collaboration can help, but is hard to achieve
41. EU action plan
•Increase transparency of online political advertising
•Improve resilience of citizens against disinformation
•Enforce accountability of online platforms and other actors for
the spread of disinformation
•Enhance cooperation between EU Member States, online
platforms, and other stakeholders
•Protect the confidentiality of journalists' sources and the
freedom of the press
https://digital-strategy.ec.europa.eu/en/library/action-plan-against-disinformation
46. Why don’t we just ask ChatGPT?
Matthew R. DeVerna†, Harry Yaojun Yan, Kai-Cheng Yang, Filippo Menczer, Artificial intelligence is ineffective and potentially harmful for fact checking
Observatory on Social Media, Indiana University, https://doi.org/10.48550/arXiv.2308.10800
47. Why don’t we just ask ChatGPT?
Matthew R. DeVerna†, Harry Yaojun Yan, Kai-Cheng Yang, Filippo Menczer, Artificial intelligence is ineffective and potentially harmful for fact checking
Observatory on Social Media, Indiana University, https://doi.org/10.48550/arXiv.2308.10800
does not significantly
affect participants'
ability to discern
headline accuracy or
share accurate news
decreases beliefs in true
headlines that it
mislabels as false and
increases beliefs for
false headlines that it is
unsure about
48. Why not let it self-regulate?
Johnson T, Kromka SM. Psychological, Communicative, and Relationship Characteristics That Relate to Social Media Users' Willingness to Denounce
Fake News. Cyberpsychol Behav Soc Netw. 2023 Jul;26(7):563-571. doi: 10.1089/cyber.2022.0204. Epub 2023 May 30. PMID: 37253156.
68% of social media users believe people should respond with a
correction when they witness the sharing of misinformation
73% of users elect to ignore misinformation posted online
self-esteem, argumentativeness, conflict style, and interpersonal
relationships relate to users’ willingness to denounce (or ignore)
disinformation.
avoiding arguments on social media is easier than confrontation and this
avoidance may take precedence if confrontation does not incentivise
social capital benefits.
users who are media literate and trained in argumentation can make a
difference - but are there enough of such users?
49. Can platforms ban disinformation?
Jahn, Laura & Kræmmer Rendsvig, Rasmus & Flammini, Alessandro & Menczer, Filippo & Hendricks, Vincent. (2023). Friction Interventions to Curb the
Spread of Misinformation on Social Media, https://arxiv.org/abs/2307.11498
Do they want to?
✓ information (esp. false) is driving traffic (i.e. profit) into platforms
Can they identify all false information with certainty?
✓ impossible as in most cases subjective views are involved (see next slide)
Do we want them to?
✓ debatable as we also need to preserve freedom of speech (see next slide)
Can they somehow help?
✓ Yes, by introducing transparency in their way of operation
✓ Yes, by introducing at least some fact-checking on clear cases
✓ Yes, by introducing friction, i.e. making it less easy for users to
propagate questionable information*
50. Should we ban d/misinformation?
Would it be ideal for the society to ban all false information?
✓ This would mean allowing only the truth
How do we define truth?
✓ The correspondence theory: A belief is true if it corresponds to the
way things actually are – to the facts
✓ The coherence theory: A belief is true if it is part of a coherent
system of beliefs
✓ Pragmatist theories: Truth is satisfactory to believe. True beliefs will
remain settled at the end of prolonged inquiry.
Can we handle the truth?
✓ Imagine a society where everyone knows the truth about everything
51. Dealing with
(dis)information
All available tools should be used and more should built
Fact-checking should be intensified in a coordinated way
Media literacy should be organised centrally
Transparency should be enforced on the platforms
Laws and regulations should hold platforms, individuals and
organisations accountable
Strong coordination and collaboration is necessary
52. Thank you for your attention
Nikos Sarris
Senior Researcher
Media analysis, veri
fi
cation and retrieval team
Information Technologies Institute
Center for Research and Technology Hellas
mever.iti.gr - www.iti.gr - www.certh.gr
Advisor on media technologies
Athens Technology Centre
http://ilab.atc.gr
@nikossarris - https://www.linkedin.com/in/nsarris/
53. CONTACT
us
This project has received funding from the European Union
DIGITAL-2021-TRUST-01. Contract number: 101083756
meddmo.eu
Fact-Check by MedDMO
@MEDDMOhub
info@meddmo.eu