Miroslava Čech

Current watermarking standards cannot reliably detect AI-generated text after editing.

Detector scores alone cannot determine whether a human or AI wrote a given text.

Perplexity and burstiness measure different writing patterns that detectors use to flag AI output.

Detecting voice drift requires baseline datasets and architectural fixes, not better prompts.

Different architectures solve different revision problems, not better and worse approaches.

Using multiple models together produces better writing than any single model alone.

Multiple models reviewing each other's work beats single-model generation on complex tasks.