Lecture 4 - Disinformation & Privacy

Potential Criteria for a Deepfake

A deepfake is manipulated media, judged by its data type, creation technology, degree of manipulation, effect on the information, whether it shows a real or fictional person, and how believable it is to viewers.

Legal Problems Caused by Deepfakes

• Disinformation – spreading false information

• False Evidence – can prolong trials as evidence is challenged

• Affronts to Dignity – e.g., nudify apps

• Commercial Exploitation – using people’s likeness for profit

NO FAKES Act (just know it exist)

• NO FAKES Act (draft, USA)

  • beschermt stem, uiterlijk en gelijkenis van personen (you cant replicated it)

• Key idea: It’s about protecting a person’s likeness.

• Everything else — life +10–70, exclusions, takedown procedures, criticisms — is extra detail and not core for exams or slides.

Disinformation

• Intentional, unlike misinformation

• Threatens:

  • Democracy

  • Public health

  • National security

• Creates echo chambers and filter bubbles

  • Echo chamber: Alleen info van mensen met dezelfde mening

  • Filter bubble: Algoritmes laten vooral zien wat je al leuk vindt/mee instemt

Facebook Algorithm

• Facebook’s algoritme toont content die veel reacties/likes krijgt → clickbait en boze posts verspreiden zich snel.

• Troll farms misbruiken dit om fake news te verspreiden voor clicks en reclame-inkomsten.

How to Tackle Disinformation

• Transparency – Algorithms must be public

• Intelligibility – Users must understand them

• Accountability – Platforms should be responsible

Regulatory Landscape – Algorithmic Disinformation

• US: No laws; relies on self-regulation, market-based approach

• EU (some countries, e.g., France): Laws like “Manipulation of Information Law”; transparency rules, but limited accountability

• China: Stricter laws; transparency + platform accountability + more user choice

Regulatory Challenges – Algorithmic Disinformation

• Transparency may be limited for average users

• Users can learn about systems but have little power to change them

• Must consider Freedom of Expression

EU AI Act – Article 50(2)

AI outputs must be marked as synthetic(clearly made artificial) ; solutions should be reliable; exceptions for minor edits, assistive tools, or legal crime detection.

3 Types of Watermarks

1. Visible – Can be seen on the content (e.g., logo on image/video)

2. Invisible – Hidden in the content, not noticeable to the eye or ear

3. Metadata – Embedded information in the file about its origin or copyright

AI-Generated Content Risks

• False positives: Legit content flagged as AI-generated

• Spoofing attacks: Fakes used to damage reputation

EU AI Act – Article 50(4)

• Deployers of AI deepfakes must disclose content as artificially generated or manipulated

• Exception: Authorized by law for crime detection/investigation

• Artistic/creative content: Only need to disclose that manipulation exists, without harming enjoyment

Who is a “Deployer”?

A deployer is everyone and everything using an AI system under their control, except for personal, non-professional use.

Dangers of Watermarking

• Can stifle freedom of speech

• May slow down technology

• Could hinder whistleblowing

• Enables mass surveillance and government abuse

GDPR (TheGeneralDataProtectionRegulation)

• Applies only to personal data

• De-identified, pseudonymised, or encrypted data still counts if it can re-identify

• Lawful processing requires:

  • Consent (not only option)

  • Contract performance

  • Public interest or legitimate interest (must outweigh person’s rights)

• Sensitive data (race, politics, religion, genetics, biometrics, sexual orientation) = higher protection, only processed with explicit consent, public interest, or legal basis (Art. 9)

Deepfakes: Personal Data?

• Not real persons

• Deceased people

• Companies or states

• Composites of people (depends on resemblance to originals)

Article 8 vs Article 10 (EuropeanCourt of Human Rights)

• Article 8 – Right to Privacy: Protects honour; public figures tolerate more intrusion/mockery

• Article 10 – Freedom of Expression: Right to shock, offend, disturb; includes satire

Digital Services Act – Article 34

• Very Large Online Platforms/Search Engines must identify, analyse, and assess systemic risks

• Risks include: illegal content, effects on fundamental rights, civic discourse, public security

Digital Services Act – Article 35

• Mitigation measures must be reasonable, proportionate, effective

• Must make manipulated content (fake news/deepfakes) clearly marked

• Provide easy way for users to flag false information

2022 Code of Practice on Disinformation – Key Measures

• Avoid placing ads next to disinformation

• Increase transparency (e.g., label political ads)

• Reduce manipulative behaviour (fake accounts, bots, deepfakes)

• Enhance tools to recognise, flag, and access authoritative sources

• Support research and fact-checking

2022 Code of Practice – Participation

• Voluntary – 33 companies signed

• Not all measures agreed

• Some major platforms did not sign

What are the problems with deepfake enforcement?

Deepfake enforcement is tricky because it involves privacy, free speech, satire, evolving social norms, and tech limitations.