Verification Methodology
How Aiscern verifies AI-generated content — the models, signals, accuracy benchmarks, and known limitations explained openly.
Accuracy Benchmarks
All figures are measured on held-out test sets from publicly available benchmark datasets. These are conservative, honest numbers — not cherry-picked accuracy on easy samples.
HuggingFace roberta-base-openai-detector + Gemini 2.0 Flash ensemble
EfficientNet-B4 fine-tuned on Midjourney/DALL-E/SD datasets
Wav2Vec2 + spectral fingerprint classifier
Frame-sampled image verification + temporal consistency analysis
* Accuracy is the percentage of samples correctly classified on public benchmark datasets. Individual results may vary based on content type, compression, and AI generator used.
Ensemble Pipeline
Aiscern runs every examination through multiple independent forensic signals. Rather than relying on a single model, results are combined using a weighted voting mechanism — models with higher validated accuracy on that modality receive more weight in the final verdict. This ensemble approach reduces false positives and improves robustness against adversarial inputs.
Text Forensic Signals
Measures how statistically predictable each word choice is. AI text scores low; human writing scores high.
Variation in sentence length and complexity. Human writing has high burstiness; AI tends toward uniformity.
Ratio of unique words to total words. AI frequently reuses high-frequency vocabulary.
AI text tends toward balanced paragraph lengths and consistent heading hierarchies uncommon in natural writing.
Specific token-choice patterns associated with known LLMs, verified via trained classifier.
Image Forensic Signals
Fourier-domain analysis reveals the periodic artifacts left by diffusion model upsampling steps.
Geometric consistency of landmarks — eye spacing, ear symmetry, catchlight positions.
Shadows, reflections, and perspective consistency between foreground subjects and background.
AI images lack camera EXIF data. Absence of shutter speed, ISO, and GPS is a strong signal.
JPEG blocking artifacts appear in atypical locations in AI images vs. real photography.
How to Interpret Integrity Ratings
Strong signals of human authorship. Low probability of AI generation.
Mixed signals. Use additional context and human judgment before acting.
Strong signals of AI generation. High-confidence ensemble verdict.
Known Limitations
AI verification is a probabilistic task. We publish our limitations openly so you can make informed decisions about when to trust — and when to double-check — these results.
Short text (under 150 words) has insufficient signal for reliable classification
Non-native English speakers may trigger false positives due to constrained vocabulary patterns
Heavily compressed images (< 50 KB) lose frequency artifacts detectors rely on
AI content edited by humans after generation reduces detectability significantly
Hybrid content (AI inpainting on real photos) is currently below 70% accuracy
Very short audio clips (< 5 seconds) provide insufficient spectral data
Novel AI generators released after our last model update may evade verification until the next fine-tune