The Turing Deception

The outlier, however, for ChatGPT is Appendix F, based on the prompt to generate variants on poetry dedicated to Turing. In this instance, the generated content bypassed Open AI's detector with high confidence as real (99.98%). In their original report [24], the authors found "detection rates of ~95% for detecting 1.5B GPT-2-generated text" and noted that "We believe this is not high enough accuracy for standalone detection and needs to be paired with metadata-based approaches, human judgment, and public education to be more effective." Like the evolution of ever larger language models (>100 billion), refinements also have built-in heuristics or guardrails for model execution. The Instruct-series of GPT-3 demonstrated the ability to answer questions directly without conversational meanderings. The ChatGPT includes longer-term conversational memory, such that the API can track the dialog even with leaps of narration that single API calls could not span. One can test dialogs with impersonal pronouns like "it" carrying forward in the conversation with context to previous API calls in a single session-one easily grasped example for ChatGPT's API memory as both powerful and expensive to encode for more extended conversations. As Turing himself posed the human capacity to list memories [1]: "Actual human computers really remember what they have to do Constructing instruction tables is usually described as'programming.'"