Section 10.1.8 : Disinformation, Deepfakes and Critical Thinking
🎯 Learning Objective
Understand the disinformation risks associated with generative AI (fake news, deepfakes, manipulation) and develop structured critical thinking to detect them. Acquire concrete reflexes to evaluate the reliability of any online content.
The Era of AI-Assisted Disinformation
Generative AI has democratized the creation of fake content at an unprecedented scale. What once required a professional studio (video deepfakes), an army of writers (fake news), or photo editing skills (doctored images) can now be produced by a single person in minutes, without any technical skills.
The danger comes not only from the quality of fake content — it comes from its production speed and volume. A bot farm powered by an open source LLM can generate thousands of disinformation articles per day, personalized for each target audience.
Types of AI Disinformation — Complete Taxonomy
1. Video Deepfakes: doctored videos showing real people saying or doing things they never did.
| Generation | Technology | Quality | Detectability |
|---|---|---|---|
| 2017-2020 | Basic GANs | Visible artifacts, rigid movements | Easy (frame analysis) |
| 2021-2023 | Advanced GANs | Convincing in low resolution | Moderate (specialized tools) |
| 2024-2025 | Diffusion (Sora 2, Veo 2) | Nearly undetectable to the naked eye | Difficult (watermarks needed) |
| 2026+ | Native multi-modal | Perfect in high resolution, real-time | Very difficult |
Examples of real-world impacts:
- →Elections: in March 2024, a deepfake audio imitating Joe Biden called thousands of New Hampshire voters to discourage them from voting in the primaries
- →Finance: in 2024, a Hong Kong company lost $25 million following a deepfake video call where the "CFO" requested an urgent wire transfer
- →Geopolitics: deepfake videos of Zelensky calling for surrender circulated during the Ukrainian conflict
2. Audio Deepfakes (voice cloning): reproducing a person's voice from just a few seconds of audio.
ElevenLabs and other services can clone a voice with only 3 to 10 seconds of sample. The synthesized voice reproduces the timbre, accent, inflections, and even natural hesitations. Massively used for:
- →CEO fraud: phone calls imitating the executive's voice to authorize wire transfers
- →Family scams: "Mom, it's me, I have a problem, send me money"
- →Fake testimony: generated audio recordings presented as evidence
3. Generated texts (large-scale fake news): disinformation articles mass-produced by GPT-5, Claude, or open source models like DeepSeek V3.1 and Llama 4.
What makes AI texts particularly dangerous:
- →Grammatically perfect (no mistakes that betray a foreign bot)
- →Stylistically varied (each article seems written by a different person)
- →Factually plausible (mix of true facts and disinformation)
- →Mass-produced (thousands of unique articles per day)
- →Personalized by audience (variants for each target community)
4. Doctored images: Midjourney v7, Flux, Aurora and DALL-E 3 produce photorealistic images impossible to distinguish from real photos.
Examples of scandals:
- →Grok and deepfake images: in 2025, Grok (xAI) was widely used to generate images of public figures in compromising situations, causing major controversies due to its less strict restrictions
- →Fake war images: AI-generated images of fictitious war scenes circulated as if they were real during several conflicts
- →Fake Pope: the viral image of Pope Francis in a Balenciaga puffer jacket (Midjourney, 2023) was the first deepfake image to fool millions of people
5. Fictitious accounts and astroturfing: creation of fake social media profiles with AI photos, generated biographies, and coherent posting histories. These accounts create the illusion of a grassroots movement (astroturfing) or amplify specific narratives.
How to Detect AI Content — Practical Guide
Warning signs for AI texts:
- →Uniformly neutral and "perfect" tone — lacking human roughness
- →Absence of personal perspective, anecdotes, emotional nuances
- →Repetitive structures (lists, symmetrical paragraphs, predictable transitions)
- →Vague or unsourced information delivered with an air of authority
- →Subtle contradictions in factual details
- →Sentences that are "too well written" — almost too polished, too balanced
- →Tendency to cover "all aspects" rather than taking a clear position
Warning signs for AI images (2025-2026):
| Area to Check | What to Look For | Test Reliability |
|---|---|---|
| Hands | Extra fingers, impossible joints, inconsistent nails | Moderate (recent models correct this) |
| Text in image | Illegible letters, invented words, inconsistent characters | Still fairly reliable |
| Backgrounds | Objects merging, impossible perspectives, repetitive patterns | Moderate |
| Facial symmetry | Too perfect (real faces are asymmetric) | Low (corrected by recent models) |
| Jewelry/accessories | Different earrings, asymmetric frames | Moderate |
| Eye reflections | Inconsistent light reflections between both eyes | Good (hard to correct) |
| Clothing/skin transition | Fusion between collar and skin, impossible seams | Good |
Warning signs for deepfake videos:
- →Lip movements desynchronized with sound (even slightly)
- →Frozen facial expression or absent micro-expressions
- →Inconsistent light transitions when the person moves
- →Face border that "glitches" or flickers
- →Eye blinking too regular or too rare
Detection tools:
| Tool | Type | Usage | Reliability |
|---|---|---|---|
| GPTZero | AI text | Detection of LLM-generated text | 80-90% on English text |
| Originality.ai | AI text | Web content analysis, AI paraphrase detection | 85-95% |
| Deepware Scanner | Video deepfake | Video analysis to detect manipulations | 75-85% |
| Sensity.ai | Multi-modal | Enterprise deepfake detection platform | 80-90% |
| Google Reverse Image | Image | Image origin verification | Good for stolen images |
| Hive Moderation | Multi-modal | AI detection for images, text, video | 85-90% |
| C2PA / Content Credentials | Metadata | Authenticity certification (Adobe, Microsoft, Google) | Very high if supported |
The Arms Race: Detection vs Generation
AI content detection is a permanent arms race. Every improvement in detectors pushes generators to improve, and vice versa. The long-term solution isn't detection (which will always lag behind) but authenticity certification: the C2PA standard (Coalition for Content Provenance and Authenticity), supported by Adobe, Microsoft, Google, and the BBC, embeds a cryptographic signature in content guaranteeing its origin. It's like a certificate of authenticity for images and videos.
The SIFT Reflex: Your Anti-Disinformation Shield
When facing questionable content, systematically apply the SIFT method (developed by Mike Caulfield, digital literacy researcher):
S — Stop: don't share immediately. Resist the emotional reaction. The more a piece of content provokes strong emotion (outrage, fear, euphoria), the more likely it is to be manipulative. Emotion is the primary vector for disinformation virality.
I — Investigate the source: who published it? What credibility? What track record? An account created 3 days ago with an AI profile picture doesn't have the same credibility as the New York Times. Check the account's age, editorial consistency, and reputation.
F — Find better coverage: do other reliable sources report the same thing? If explosive information is only relayed by an obscure blog and anonymous accounts, that's a major red flag. Real important events are covered by multiple independent media outlets.
T — Trace the original: go back to the original source. A screenshot is not proof. A tweet cited in an article is not proof. Find the original document, the complete video (not the 10-second clip), the official statement.
Critical Thinking About AI Itself
Disinformation doesn't only come from "bad actors" — AI itself can be a source of unintentional misinformation:
Hallucinations: LLMs sometimes invent facts, quotes, scientific studies that don't exist. An AI text can cite a "2024 McKinsey report" that was never published, with precise figures and detailed methodology — all fabricated.
Confidence bias: the confident, expert tone of AI responses creates an illusion of reliability. A false but well-formulated answer is more misleading than an approximate but honest one.
Amplification effect: if false information is present in the AI's training data, it will be reproduced and amplified by the model, which presents it as established fact.
The 5 Reflexes of the Critical AI Practitioner
- →Verify key figures: ask for sources, cross-check with Google (30 seconds)
- →Beware of perfection: an answer that's "too perfect" without nuance is suspect
- →Look for contradictions: ask the same question from different angles
- →Ask for limitations: "What are the limits of your answer? What might be inaccurate?"
- →The "too good" test: if it's too good, too shocking, or too perfect — verify twice
Societal Impact: Beyond Fake News
AI disinformation threatens pillars of our societies:
Democracy: election manipulation through automated disinformation campaigns. In 2024, deepfakes and AI bots were detected in elections in over 40 countries.
Justice: fake audio/video evidence presented in courts. The legal system is not yet equipped to systematically authenticate digital content.
Social trust: the "liar's dividend" phenomenon — when anything can be fake, real compromising content can be denied as deepfakes. "That's not me in that video, it's a deepfake."
Education: students increasingly struggle to distinguish reliable sources from AI content. Entire essays generated by GPT are submitted as original work.
Practical Exercise: Disinformation Detection
Duration: 20 minutes
- →Find 3 recent viral images on social media
- →For each one, apply the complete SIFT method
- →Verify with Google Reverse Image Search whether the image is original or modified
- →Run an image generated by Midjourney through a detector like Hive Moderation
- →Compare your results: how many fake contents did you detect?
According to the World Economic Forum 2024, what is the world's #1 short-term risk?
What is the best long-term strategy against deepfakes, according to experts?
Learn AI, from prompts to agents