AI-Generated Content (AIGC)

AIGC exhibits a highly predictable cadence and sentence structure. Unlike human writing, which varies dramatically in length and complexity, machine-generated text tends toward a uniform distribution of sentence lengths. This lack of variance is measured by burstiness scoring, a key metric for detection.

• Mechanism: Models optimize for average log-likelihood, avoiding the erratic structural spikes of human prose.
• Detection: Algorithms flag text where the standard deviation of sentence length falls below a human baseline.
• Risk: Training on low-burstiness text causes tail erosion, where the model forgets how to generate stylistically diverse or complex outputs.

AIGC is statistically 'obvious' to a language model. Perplexity filtering uses the generating model's own probability distribution to see if the text is too predictable. Since models generate tokens they deem highly probable, synthetic text often has lower perplexity than human text.

• Self-Reference: A small model can often detect the output of a larger model by identifying token sequences that are 'too perfect'.
• Limitation: This fails against adversarial generation techniques that deliberately inject low-probability tokens.
• Contamination Vector: Recursively training on low-perplexity data narrows the output distribution, leading to model autophagy.

In a self-consuming loop, AIGC amplifies subtle artifacts and biases present in the teacher model. Without the corrective noise of human-originated data, models begin to repeat specific phrases, syntactic structures, and semantic concepts with increasing frequency.

• Artifact Amplification: Minor quirks in the original model become dominant features in the child model.
• Bias Amplification Loop: Societal biases that were subtle in the original distribution become extreme and caricatured.
• Mitigation: Requires rigorous training corpus sanitization and the injection of high-quality human-originated data to break the loop.

AIGC often contains factually incorrect or nonsensical information known as hallucinations. When this content is scraped and reintroduced into training corpora, it creates a contamination pathway called model hallucination recycling.

• Error Propagation: A false statement generated by Model A becomes 'ground truth' for Model B.
• Verification Gap: Unlike human text, AIGC lacks the implicit link to physical reality or lived experience.
• Solution: Data provenance verification and C2PA standard content credentials are required to distinguish grounded human data from unverified synthetic text.

To combat contamination, modern AIGC systems embed imperceptible signals. AI watermarking like SynthID modifies the generation process to create a detectable pattern that survives screenshots and edits.

• C2PA Standard: Attaches a cryptographically verifiable manifest to content, establishing a chain of custody from camera to cloud.
• Filtering: Downstream data pipelines can automatically exclude watermarked content to prevent data contamination.
• Limitation: Watermarks can be broken by aggressive paraphrasing or regeneration, requiring a defense-in-depth approach.

AIGC tends to represent the 'average' of a distribution, failing to capture the long tail of rare events, minority perspectives, and edge cases. Training on this data causes distribution shift.

• Tail Erosion: The model loses the ability to represent fringe data points, effectively 'forgetting' the edges of the knowledge space.
• Diversity Loss: Over generations, the output collapses to a mean representation devoid of novelty.
• Detection: MinHash deduplication fails to fix this, as the problem is not exact duplicates but a loss of statistical variance.