Pangram CEO: People believe AI writing is indistinguishable from ours. It's not (bigthink.com)

Pangram's CEO, Max Spero, emphasizes the growing concerns around AI-generated content and the limitations of current AI detection technologies. Early AI detectors, which often misidentified classic texts as generated by AI, relied on perplexity and burstiness metrics that failed to capture the nuances of human writing. Despite ongoing improvements, such as a claimed false positive rate of just one in 10,000, false positives still pose a challenge, leaving writers vulnerable to accusations of inauthenticity. Spero highlights Pangram's approach of utilizing deep learning to train on a diverse corpus of both human and AI-generated text, creating what he terms "synthetic mirrors" to teach the model the differences in writing styles. This method aims to produce a more reliable detection tool, especially as new generative models are released. He argues for the importance of developing social norms around AI usage, noting a public backlash against AI-generated content in spaces like LinkedIn, where approximately 40% of long-form posts are now deemed AI-produced. As AI continues to evolve, the conversation on maintaining authenticity in writing becomes increasingly pertinent within the AI/ML community.