AI-Generated Content Detection

Why Detection Matters

The proliferation of AI-generated content creates challenges for:

Academic integrity

News media verification

Social media authenticity

Legal evidence authentication

Brand protection

Text Detection Methods

Statistical Approaches

AI text tends to have different statistical properties:

Lower perplexity (more predictable)

Higher burstiness patterns

Different n-gram distributions

python
import transformers
import torch
def calculate_perplexity(text: str, model_name: str = "gpt2") -> float:
    model = transformers.AutoModelForCausalLM.from_pretrained(model_name)
    tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)
    
    inputs = tokenizer(text, return_tensors="pt")
    with torch.no_grad():
        outputs = model(**inputs, labels=inputs.input_ids)
    
    return torch.exp(outputs.loss).item()
Low perplexity might indicate AI generation
score = calculate_perplexity("The quick brown fox...")
print(f"Perplexity: {score}")  # Human text typically higher

Classifier-Based Detection

python
from transformers import pipeline
detector = pipeline(
    "text-classification",
    model="roberta-base-openai-detector"
)
result = detector("Your text here...")
print(f"Label: {result[0]['label']}, Score: {result[0]['score']:.3f}")

Image Deepfake Detection

python
import cv2
import numpy as np
from tensorflow.keras.models import load_model
def detect_deepfake(image_path: str) -> dict:
    model = load_model("deepfake_detector.h5")
    img = cv2.imread(image_path)
    img = cv2.resize(img, (224, 224)) / 255.0
    prediction = model.predict(np.expand_dims(img, 0))[0][0]
    return {
        "is_fake": prediction > 0.5,
        "confidence": float(prediction)
    }

C2PA Content Credentials

The Coalition for Content Provenance and Authenticity (C2PA) standard embeds cryptographic provenance in media files.

Watermarking AI Content

Invisible watermarks can be embedded in AI-generated content for later detection.

Limitations

No detection method is 100% accurate. Treat results as probabilistic signals, not definitive verdicts.