Frequently Asked Questions

AI-generated content refers to text, images, videos, audio, or other media created by artificial intelligence systems rather than humans. This includes:

• Text: Articles, essays, code, and creative writing produced by large language models like ChatGPT, Claude, or GPT-4
• Images: Pictures created by systems like DALL-E, Midjourney, Stable Diffusion, or Adobe Firefly
• Videos: Deepfakes, synthetic videos, or AI-modified footage
• Audio: Voice cloning, synthetic speech, or AI-generated music

Modern AI can create remarkably convincing content that mimics human creativity, making detection increasingly important for maintaining authenticity and trust in digital communications.

Detection accuracy varies significantly based on several factors:

Text Detection: Typically 70-95% accurate for straightforward cases, but accuracy drops with:

• Human editing of AI-generated content
• Newer AI models not in training data
• Short text samples (under 100 words)
• Technical or domain-specific content

Image Detection: Can be highly accurate (85-98%) when checking for specific artifacts, but struggles with:

• Latest generation models with improved realism
• Post-processed or edited AI images
• Hybrid images combining AI and real photography

Video Detection: Varies widely (60-95%) depending on deepfake sophistication and video quality.

No, AI detectors cannot identify all AI-generated content. Several factors limit detection capabilities:

False Negatives (AI marked as human):

• Sophisticated AI models specifically designed to evade detection
• Human editing that masks AI patterns
• Newer AI systems not represented in detector training data
• Content generated with adversarial techniques
• Hybrid content mixing AI and human creation

False Positives (Human marked as AI):

• Highly formulaic human writing
• Technical documentation with consistent style
• Formal business communication
• Content following strict templates or guidelines

Detection provides probability indicators, not certainty. Results should inform but not solely determine authenticity judgments.

Detecting AI-generated content is crucial for multiple reasons:

1. Combating Misinformation: AI can generate fake news, fabricated evidence, and misleading content at scale. Detection helps identify potentially false information before it spreads.

2. Academic Integrity: Educational institutions need to ensure students are developing genuine skills and understanding rather than submitting AI-completed work as their own.

3. Professional Standards: Journalism, research, and professional fields require authentic content and proper attribution when AI assists in creation.

4. Legal and Forensic: Courts and law enforcement need to verify evidence authenticity, especially with deepfake technology making fake evidence increasingly convincing.

5. Fraud Prevention: Criminals use AI for sophisticated scams including fake identity documents, phishing content, and impersonation. Detection helps prevent financial and personal harm.

6. Copyright and Creativity: Artists and content creators need protection from AI systems trained on their work or AI-generated content passed off as human-created art.

7. Informed Consumption: Individuals deserve to know whether content they're consuming is human-created or AI-generated to make informed judgments about its reliability and perspective.

Text AI detectors use multiple sophisticated techniques:

1. Perplexity Analysis: Measures how "surprised" a language model would be by the text. AI-generated text tends to have lower perplexity (more predictable) because AI chooses likely word sequences.

2. Burstiness Measurement: Analyzes variation in sentence length and complexity. Human writing shows natural "burstiness" with mixed short and long sentences, while AI tends toward uniformity.

3. Statistical Patterns: Examines word frequency distributions, n-gram patterns, and syntactic structures that differ between human and AI writing.

4. Stylometric Analysis: Studies writing style markers including punctuation patterns, paragraph structure, and transition word usage.

5. Content Analysis: Checks for AI-characteristic phrases, lack of personal perspective, factual inconsistencies, and "hallucination" patterns.

6. Classifier Models: Machine learning models trained on large datasets of human and AI text to identify distinguishing features.

Advanced detectors combine multiple methods for more reliable results, but no approach achieves perfect accuracy.

While AI text becomes increasingly sophisticated, several indicators may suggest artificial origin:

Content Patterns:

• Repetitive phrases or sentence structures
• Generic statements lacking specific personal details
• Overly balanced arguments without clear stance
• Factual errors or "hallucinated" plausible-sounding information
• Lack of recent or time-specific references

Style Markers:

• Excessively formal or uniform tone
• Overuse of transition words ("furthermore," "moreover," "in conclusion")
• Perfect grammar with no typos or informal language
• Unnaturally consistent paragraph lengths
• Hedging language ("it's important to note," "it can be said that")

Structural Elements:

• Formulaic organization (numbered lists, strict structure)
• Lack of unique metaphors or creative expressions
• Missing personal anecdotes or lived experience
• Surface-level treatment of complex topics

No, ChatGPT and similar AI text cannot always be reliably detected, especially in these scenarios:

Challenging Cases:

• Human Editing: When users heavily edit AI output, detection becomes much harder
• Prompted Style: AI instructed to write in specific, irregular styles may evade detection
• Short Content: Brief texts (under 100 words) lack sufficient patterns for reliable analysis
• Technical Writing: Formal technical documentation may look AI-generated even when human-written
• Hybrid Creation: Combining AI suggestions with human writing creates mixed signals

Newer Models: Each new AI version (GPT-4, GPT-4 Turbo, etc.) introduces patterns not in detector training data, reducing accuracy until detectors update.

Evasion Techniques: Users can employ various methods to make AI text less detectable:

• Paraphrasing tools applied to AI output
• Mixing content from multiple AI sources
• Intentionally introducing errors and informalities
• Using AI to "humanize" AI-generated text

OpenAI discontinued its own AI text classifier in 2023 due to low accuracy (26% true positive rate), highlighting the difficulty of reliable detection.

Detection accuracy improves significantly with longer text samples:

Minimum Recommendations:

• 50-100 words: Minimum for basic analysis, but high error rates (below 50% accuracy)
• 200-300 words: Sufficient for moderate confidence detection (60-75% accuracy)
• 500+ words: Ideal for reliable analysis (75-90% accuracy with good detectors)
• 1000+ words: Best accuracy possible with current technology (80-95% accuracy)

Why Length Matters:

• More text provides more patterns to analyze
• Statistical methods require sufficient data points
• Longer passages reveal consistency patterns characteristic of AI
• Short texts may not contain distinctive markers

Quality Over Quantity: Complete paragraphs with varied content provide better detection signals than fragmented text, even if longer. Context-rich passages (narratives, arguments, explanations) analyze better than lists or sparse notes.

Identifying AI-generated images requires examining multiple visual and technical elements:

Visual Inspection:

• Hands and fingers: Count digits carefully; AI often generates extra, missing, or malformed fingers
• Eyes and facial symmetry: Check if both eyes match, have realistic reflections, and proper iris detail
• Text and logos: Any readable text should make sense; AI typically produces gibberish letters
• Background details: Look for objects that blend together unnaturally or have impossible geometries
• Lighting consistency: Shadows should align with light sources; check for multiple conflicting shadows
• Hair strands: Individual hairs should look natural, not melted or clumped strangely
• Teeth and jewelry: Count teeth, check earring pairs, look for accessories phasing through clothing

Technical Analysis:

• EXIF metadata: Check for camera information; AI images often lack detailed EXIF data
• File properties: Examine creation software, modification dates, and file size patterns
• Image dimensions: AI generators often use specific sizes (512×512, 1024×1024, 768×768)
• Compression patterns: Analyze artifacts that may indicate AI generation versus camera capture

Context Clues:

• Is the image too perfect or idealized?
• Does it match other photos of the same subject/location?
• Can you find the source or photographer?
• Does the setting match the claimed time/place?

The "AI hand problem" stems from several technical and training challenges:

Complexity of Hands:

• Articulation: Hands have 27 bones and numerous joints allowing countless positions and configurations
• Occlusion: Fingers overlap and obscure each other differently in every pose
• Perspective: Hand size and finger length change dramatically based on viewing angle
• Context-dependency: Proper hand appearance depends on what the hand is doing (grasping, pointing, gesturing)

Training Data Issues:

• Hands appear in infinite variations making it hard for AI to learn consistent patterns
• Training images often show hands in motion (blurred) or partially visible
• Many training photos focus on faces, with hands as secondary elements
• Labeled data specifically identifying hand structures may be limited

AI Architecture Limitations:

• Diffusion models generate images holistically, struggling with localized precision
• Models learn statistical patterns, not anatomical rules
• Finger count isn't explicitly constrained in current architectures

Newer models (2024 onwards) are improving hand generation, but it remains a telltale sign worth checking in suspected AI images.

Detection becomes significantly more challenging with edited images, but some analysis remains possible:

Types of Editing and Detection Impact:

Minor Edits (High Detection Possible):

• Color correction and brightness adjustments
• Cropping or resizing
• Adding text overlays or watermarks
• Basic filters applied uniformly

Moderate Edits (Medium Detection):

• Selective adjustments to specific areas
• Removal of obvious AI artifacts
• Combining AI-generated elements with real photos
• Artistic filters and effects

Extensive Edits (Low Detection):

• Professional retouching removing AI telltale signs
• Compositing multiple sources (AI and real)
• Complete style transfer or heavy manipulation
• Regeneration through multiple AI passes

What Still Works:

• Metadata may reveal editing software and timeline
• Some statistical patterns persist through editing
• Unedited regions may still show AI characteristics
• Forensic analysis can detect composition of multiple sources

Image format significantly affects detection capabilities:

Best Formats for Detection:

1. PNG (Portable Network Graphics):

• Lossless compression preserves all original data
• Maintains subtle artifacts useful for detection
• Retains full color depth and detail
• Common output format for AI generators

2. Original/Uncompressed Formats:

• TIFF, BMP, or raw AI output files
• Maximum preservation of generation artifacts
• Best for forensic analysis
• Rarely available in practice

Acceptable Formats:

3. High-Quality JPEG:

• Minimal compression (quality 90+)
• Some artifacts remain detectable
• Very common in practice
• Metadata often preserved

Problematic Formats:

4. Heavily Compressed JPEG:

• Lossy compression destroys subtle patterns
• Multiple save cycles add noise
• Compression artifacts obscure AI artifacts
• Reduces detection accuracy significantly

5. WebP, Social Media Re-encodes:

• Platforms often re-compress uploaded images
• Metadata frequently stripped
• Makes technical analysis much harder
• Visual inspection becomes primary method

Best Practice: Always analyze the highest quality, least compressed version available. If possible, request original files rather than screenshots or re-shared versions.

Deepfakes are synthetic media where AI replaces or synthesizes video/audio content to create realistic but fake recordings:

Types of Deepfakes:

• Face Swapping: Replacing one person's face with another's in video
• Face Reenactment: Puppeteering someone's facial expressions from another video
• Voice Cloning: Synthesizing someone's voice saying different words
• Lip Syncing: Modifying lip movements to match different audio
• Full Body Synthesis: Creating entirely synthetic people or actions

Creation Technology:

1. Deep Learning Networks:

• GANs (Generative Adversarial Networks) pit two AI models against each other
• Autoencoders learn to compress and reconstruct facial features
• Training on thousands of images of target and source faces

2. Face Mapping:

• Facial landmarks identified and tracked frame-by-frame
• 3D face models created from 2D images
• Expressions transferred between face models

3. Rendering and Compositing:

• Synthetic face blended into original video
• Color and lighting matched to surrounding pixels
• Temporal smoothing ensures consistency across frames

Accessibility Concerns: Deepfake technology has become increasingly accessible with free apps and services, making creation possible for non-experts. This democratization raises significant concerns about misuse for fraud, harassment, and disinformation.

Detecting deepfakes requires careful attention to multiple visual and audio elements:

Facial Analysis:

• Blinking: Unnatural patterns - too frequent, infrequent, or asymmetric
• Lip sync: Mouth movements not perfectly matching spoken words
• Facial boundaries: Blurring or discoloration around face edges, especially near hair
• Skin texture: Overly smooth, waxy, or plastic-like appearance
• Facial expressions: Limited micro-expressions or emotional nuance
• Eye contact: Unnatural gaze direction or both eyes not focusing together

Technical Inconsistencies:

• Lighting mismatch: Face lighting doesn't match environment
• Color grading: Face color different from neck, hands, or surroundings
• Focus depth: Face sharpness inconsistent with background focus
• Motion blur: Face remains sharp when head moves quickly
• Reflections: Face not properly reflected in glasses, mirrors, or shiny surfaces

Frame-by-Frame Analysis:

• Watch at slow speed or pause frequently
• Look for momentary glitches, warping, or artifacts
• Check consistency of facial features across frames
• Watch for background elements flickering or changing

Audio-Visual Sync:

• Audio quality matching video quality
• Room acoustics matching visible environment
• Breathing visible when voice sounds breathless
• Background sounds matching visible activities

Contextual Red Flags:

• Video appears only from single source with no corroboration
• Content seems out of character for the person
• Unusual circumstances or implausible scenarios
• Video quality lower than expected for the situation

Yes, several automated deepfake detection tools exist, though all have limitations:

Commercial Tools:

• Sensity: Enterprise platform for detecting and monitoring synthetic media
• Reality Defender: Real-time detection API and browser extension
• Microsoft Video Authenticator: Analyzes videos for manipulation indicators
• Intel FakeCatcher: Real-time deepfake detector analyzing blood flow patterns

Research Tools:

• Deepware Scanner: Free online tool for video analysis
• WeVerify: Browser plugin for video verification
• FaceForensics++: Academic dataset and detection models

How They Work:

• Biological signals: Detecting absence of natural pulse in facial blood flow
• Temporal inconsistency: Identifying frame-to-frame artifacts
• Generative model fingerprints: Recognizing specific AI model characteristics
• Compression artifacts: Analyzing re-encoding patterns from face swapping
• Frequency analysis: Examining image data in frequency domain

Limitations:

• Accuracy drops with newer deepfake methods
• High-quality deepfakes can fool automated systems
• Compressed or low-resolution videos harder to analyze
• Some tools require original, unedited footage
• False positive rates can be significant

The deepfake threat is significant and growing, affecting individuals, organizations, and society:

Current Real-World Impacts:

1. Political Manipulation:

• Fake videos of politicians saying inflammatory things
• Election interference through synthetic media
• Undermining trust in authentic evidence and recordings

2. Financial Fraud:

• Voice-cloning scams impersonating executives to authorize payments
• Video calls with synthetic participants for identity fraud
• Stock manipulation through fake CEO statements
• Cases involving millions of dollars in losses

3. Personal Harm:

• Non-consensual intimate deepfakes (primarily targeting women)
• Reputation damage through fake videos
• Harassment and bullying using synthetic media
• Identity theft for various malicious purposes

4. Erosion of Trust:

• "Liar's dividend" - dismissing authentic evidence as fake
• Decreased confidence in video evidence generally
• Difficulty distinguishing truth in information ecosystem

Scale of the Problem:

• Deepfake detection services report 900% increase in deepfake content (2019-2023)
• Free and user-friendly tools make creation accessible to anyone
• Real-time deepfake technology enables live video manipulation
• Detection capabilities lag behind generation advances

Mitigating Factors:

• Increasing public awareness and media literacy
• Development of authentication technologies
• Legal frameworks and regulations emerging
• Platform policies against non-consensual deepfakes
• Technical improvements in detection methods

The threat is serious enough that governments, tech companies, and researchers are prioritizing solutions, but the technology continues to advance faster than countermeasures.

Using AI-generated content is generally legal, but subject to important context-specific requirements and restrictions:

Legal Considerations by Context:

1. Academic Use:

• Most institutions require disclosure of AI assistance
• Some prohibit AI use for certain assignments entirely
• Undisclosed AI use may constitute academic dishonesty
• Policies vary widely between institutions and departments

2. Professional/Commercial:

• Generally legal but may require disclosure
• Copyright of AI-generated content is legally uncertain
• Some jurisdictions don't recognize AI authorship for copyright
• Clients may require original human-created work in contracts

3. Journalism and Media:

• Ethical standards typically require disclosure
• Using AI images without attribution may violate policies
• Passing AI content as human-sourced reporting is deceptive

4. Political Advertising:

• Some jurisdictions require disclosure of AI-generated political ads
• Federal regulations emerging in multiple countries
• Deepfakes of candidates may face specific restrictions

5. Deepfakes and Impersonation:

• Non-consensual deepfakes increasingly illegal (especially intimate content)
• Impersonation for fraud purposes is criminal
• Defamatory deepfakes subject to existing defamation law
• Some US states have specific anti-deepfake laws

Copyright and Training Data:

• Ongoing legal questions about AI training on copyrighted material
• Generated content may inadvertently reproduce copyrighted elements
• Legal landscape evolving with ongoing litigation

AI-generated content raises numerous ethical questions spanning multiple domains:

Authenticity and Trust:

• Deception: Is it ethical to present AI content as human-created without disclosure?
• Trust erosion: Does widespread AI content undermine trust in all digital media?
• Attribution: Should AI contributions always be acknowledged?

Creative and Labor Ethics:

• Artist rights: Is training AI on artists' work without permission ethical?
• Labor displacement: What obligations exist when AI replaces human creative work?
• Skill development: Does AI use prevent development of important human capabilities?
• Creative authenticity: Does AI-generated art have the same value as human creation?

Information Integrity:

• Misinformation risk: Who bears responsibility for AI-generated false content?
• Verification burden: Is it fair to shift verification responsibility entirely to consumers?
• Platform responsibilities: Should platforms be required to detect and label AI content?

Consent and Privacy:

• Likeness rights: Is using someone's appearance in AI-generated content without permission ethical?
• Voice cloning: What consent is needed to create synthetic versions of someone's voice?
• Data privacy: How should personal data used in AI training be protected?

Equity and Access:

• Bias amplification: Does AI perpetuate or amplify existing biases in training data?
• Representation: Who gets to be represented in AI models and how?
• Access inequality: Does AI advantage those with resources to use it effectively?

Emerging Ethical Frameworks:

• Professional organizations developing AI ethics guidelines
• Academic institutions establishing usage policies
• Industry self-regulation efforts (varying effectiveness)
• Cross-cultural differences in ethical perspectives

These ethical questions don't have universal answers but require ongoing discussion, policy development, and individual moral reasoning.

LooksFake AI is designed with privacy as a fundamental principle:

Local Processing:

• All analysis occurs entirely in your browser
• No files or text are uploaded to our servers
• No data leaves your device during analysis
• You maintain complete control of your content

No Data Collection:

• We don't store the content you analyze
• We don't track what you submit for detection
• No accounts or registration required
• No analysis history maintained on our servers

Anonymous Usage:

• Basic analytics use privacy-respecting tools
• No personally identifiable information collected
• No cross-site tracking
• Minimal necessary cookies only

Open and Transparent:

• Our privacy policy clearly explains all data handling
• Client-side code can be inspected in browser
• No hidden third-party services accessing your content

LooksFake AI can be used professionally, but with important limitations to understand:

Appropriate Professional Uses:

• Initial Screening: Quick preliminary check of suspicious content
• Education and Training: Teaching others about AI content detection
• Content Triage: Identifying which items warrant deeper investigation
• Awareness: Understanding what signals to look for in content
• Supplementary Analysis: One tool among multiple verification methods

Important Limitations:

• Not Forensic-Grade: Our tools use heuristic analysis, not comprehensive forensics
• No Legal Certainty: Results aren't admissible as definitive evidence
• False Positive/Negative Risk: Errors occur with all detection methods
• Educational Purpose: Tool designed primarily for awareness and learning

When Professional Tools Are Needed:

• Legal proceedings requiring expert testimony
• High-stakes decisions affecting reputation or careers
• Forensic investigation of sophisticated deepfakes
• Publication decisions for major news organizations
• Corporate security incident investigation

Professional Alternatives:

• Specialized forensic services (Sensity, Reality Defender)
• Expert consultants with credentials in digital forensics
• University research labs with advanced detection systems
• Law enforcement digital evidence units

Best Practice: Use LooksFake AI as a first-pass tool to identify potential issues, then engage appropriate professional resources for critical decisions. Document your detection process and the multiple methods used.

Developing strong AI detection skills requires practice, education, and staying current:

Practice Methods:

• Challenge Datasets: Test yourself with labeled AI and human content
• Daily Practice: Analyze a mix of real and AI content regularly
• Compare and Learn: After guessing, check results and understand why you were right or wrong
• Focus Training: Spend time specifically on your weak areas (e.g., hands in images)

Educational Resources:

• Our Guides: Read our comprehensive detection guide thoroughly
• Research Papers: Study academic research on detection methods
• Online Courses: Take courses on digital forensics and media literacy
• Webinars: Attend talks by deepfake researchers and fact-checkers

Community Engagement:

• Join fact-checking communities and forums
• Participate in deepfake detection challenges
• Follow researchers and practitioners on social media
• Share your findings and learn from others' analyses

Technical Skill Building:

• Tool Familiarity: Learn multiple detection tools and their strengths
• Metadata Analysis: Understand how to examine image/video metadata
• Forensic Techniques: Study basic digital forensics principles
• Statistics: Learn about perplexity, burstiness, and statistical analysis

Staying Current:

• Follow AI research news (new models often require new detection approaches)
• Subscribe to newsletters covering AI and misinformation
• Regularly test your skills against latest AI generation models
• Update your knowledge as detection techniques improve

Develop Critical Thinking:

• Always consider context and source credibility
• Question initial impressions and biases
• Use multiple verification methods before concluding
• Understand the limits of any single detection approach

La précision varie selon les outils et les types de contenu. Pour la détection de texte IA, les outils leaders atteignent 85-95 % de précision sur du contenu entièrement généré par l'IA. Cependant, les contenus hybrides (partiellement édités) et les textes courts réduisent considérablement cette précision. Pour la détection d'images IA, la précision varie de 75-90 % selon le modèle générateur. Aucun outil n'est infaillible — utilisez-les toujours en complément de votre jugement critique.

Non. Les détecteurs d'IA ne peuvent pas identifier tout le contenu généré par l'IA avec une fiabilité absolue. Les principales limitations incluent : les mises à jour fréquentes des modèles IA qui modifient leurs signatures, l'édition humaine qui masque les patterns IA, les textes courts manquant de signaux statistiques suffisants, et les techniques de contournement délibérées. Traitez les résultats de détection comme des indicateurs probabilistes, pas comme des preuves définitives.

La détection du contenu IA est importante pour : préserver l'intégrité académique (plagiat IA), combattre la désinformation et les deepfakes, maintenir la confiance dans les communications numériques, protéger les droits d'auteur, et garantir l'authenticité dans le journalisme, la recherche et la prise de décision. À mesure que l'IA devient plus sophistiquée, la capacité à distinguer le contenu authentique du contenu synthétique devient cruciale.

Les détecteurs de texte IA analysent des patterns statistiques qui caractérisent les textes générés par machine. Les méthodes incluent : analyse de la perplexité (mesure l'imprévisibilité du texte — le texte IA tend à être plus « prévisible »), analyse de la burstiness (les humains alternent entre phrases courtes et longues, les IA sont plus uniformes), détection de watermarks intégrés dans certains modèles IA, et classification par apprentissage automatique entraînée sur du texte humain et IA.

Signes courants du texte IA : structure très uniforme et prévisible, absence de vrais exemples personnels, transitions lisses mais génériques, surutilisation de certaines formulations (« En conclusion », « Il convient de noter »), absence d'opinions tranchées ou controversées, aucune référence à des événements très récents, et une longueur de phrase plus uniforme que dans l'écriture humaine.

Non — le texte ChatGPT ne peut pas toujours être détecté avec fiabilité. OpenAI a lui-même retiré son propre détecteur en raison d'un taux élevé de faux positifs. Le texte ChatGPT peut contourner la détection par : édition humaine approfondie, utilisation de techniques de « humanisation » du texte, ou simplement en générant beaucoup de texte et en sélectionnant les parties les moins détectables. Les améliorations continues des modèles rendent la détection de plus en plus difficile.

Indices visuels des images IA : mains avec un mauvais nombre de doigts ou des proportions incorrectes, texte illisible ou incompréhensible dans l'image, arrière-plans flous ou incohérents, bijoux et accessoires impossibles physiquement, symétrie faciale trop parfaite, transitions de couleur anormalement douces, et absences de détails cohérents dans les bords et transitions.

Les mains sont structurellement complexes avec de nombreuses variations possibles. Les modèles d'IA apprennent à partir d'exemples statistiques et les mains humaines dans les photos montrent d'importantes variations selon les angles, les poses et les occlusions. Cette complexité structurelle est difficile à capturer avec la précision nécessaire, ce qui entraîne souvent des représentations erronées — doigts supplémentaires, proportions incorrectes, ou fusion de doigts.

Les deepfakes sont des vidéos ou audios manipulés par l'IA pour faire paraître une personne en train de dire ou faire quelque chose qu'elle n'a pas fait. Ils sont créés en utilisant des réseaux de neurones génératifs adversariaux (GAN) ou des modèles de diffusion entraînés sur des images/vidéos de la cible. La qualité des deepfakes a considérablement augmenté — certains nécessitent maintenant seulement quelques secondes d'audio ou quelques photos pour créer des imitations convaincantes.

Indices d'une vidéo deepfake : clignement des yeux irrégulier ou absent, synchronisation labiale imparfaite avec l'audio, flou aux bords du visage, inconsistances de l'éclairage entre le visage et l'arrière-plan, expressions faciales légèrement « figées » ou exagérées, qualité d'image soudainement différente, et mouvements de tête non naturels. Les outils automatiques comme FaceForensics++ atteignent une précision de détection allant jusqu'à 99 % dans des conditions contrôlées.

La légalité varie selon l'utilisation et la juridiction. Le contenu IA est généralement légal pour un usage personnel ou créatif. Cependant, il peut être problématique dans : les travaux académiques sans déclaration, le journalisme sans transparence, les deepfakes non consensuels (illégaux dans de nombreux États américains et pays de l'UE), les deepfakes pornographiques (illégaux dans de nombreuses juridictions), la fraude ou l'escroquerie, et la violation des droits d'auteur des données d'entraînement.

Oui. LooksFake AI analyse le contenu localement dans votre navigateur dans la mesure du possible. Les images et textes soumis ne sont pas stockés de manière permanente et ne sont pas utilisés pour entraîner nos modèles. Consultez notre politique de confidentialité pour les détails complets sur le traitement des données.