## TL;DR
As AI-generated text, images, audio, and video become indistinguishable from human-created content, detecting what's real becomes a critical societal challenge — from identifying AI-written student essays and fake product reviews to detecting political deepfakes. Detection is an asymmetric arms race: generation improves faster than detection, and each detection method spawns evasion techniques.

## Core Explanation
AI-generated text detection approaches: (1) Statistical detectors — analyze token probability distributions. Human writing has natural variability (some words unexpected); AI-generated text has concentrated probability (model selects most likely tokens). GLTR (2019) visualizes this; DetectGPT (2023) perturbs text and measures probability curvature — AI text tends to be at local probability maxima, slight changes decrease probability; (2) Classifier-based — fine-tuned discriminators (RoBERTa, GPT-2 detector) trained on human vs. AI text pairs. Strong on specific models, poor generalization; (3) Watermarking — during generation, bias token selection toward a "green list" of tokens. Detector checks if text has disproportionately many green-list tokens — provides statistical guarantee. SynthID (Google, 2023-2025) embeds imperceptible watermarks in generated text, images, and audio; (4) Retrieval-based — store all generated text in a database, check if candidate text matches. Only works for known generators.

## Detailed Analysis
Deepfake detection: (A) Image — face artifacts (inconsistent eye reflections, asymmetric earrings, blending boundaries), frequency domain analysis (GAN fingerprints), and physiological signals (heart rate from facial color changes, inconsistent with real humans). SOTA detectors (XceptionNet, EfficientNet-based) achieve 95%+ on known generators, drop to 60-70% on unseen generators; (B) Video — temporal inconsistencies across frames (flickering, unnatural blinking patterns), audio-visual asynchrony (lip movements not matching speech); (C) Audio — vocoder artifacts, unnatural prosody patterns. The arms race: every detection method spawns evasion — paraphrasing defeats statistical text detection; adversarial noise defeats image detection; re-encoding defeats watermarking. C2PA represents a proactive approach — instead of detecting fakes, establish authenticity for real content. Cameras sign images at capture with device attestation; editing software (Photoshop) adds edit claims; social media platforms display "content credentials" showing provenance chain. SynthID embeds watermarks during generation rather than post-hoc. Key open problem: the asymmetry — detection must catch 100% of fakes (one missed deepfake can cause significant harm), while generation needs only one success. Regulatory landscape: EU AI Act requires AI-generated content labeling; China mandates watermarking of deep synthesis; US AI executive orders encourage provenance standards. The consensus: no single detection method suffices — defense-in-depth combining multiple orthogonal approaches is necessary.

## Further Reading
- SynthID: Google DeepMind Watermarking
- C2PA/Content Authenticity Initiative (Adobe)
- GLTR: Visual Forensic Tool for AI Text (Harvard/IBM)