Through various techniques, artificial intelligence (AI) detectors examine written works and attempt to determine whether it was generated by a human writer. For example, it may evaluate each sentence based on its structure and word choice to determine if they match what an AI would generate in the same context. Then, it looks at how those sentences rank within the overall context of the full paragraph. Then it looks at how those paragraphs score in the entirety of the written document. From those evaluations, it will calculate a probability score that the content is human.
To this point, there are some important caveats:
- AI Detectors are not 100% accurate.
GPTs write content based on rules and patterns that they have discerned from their training materials. GPT output is a representative of a basic subset of human writing as the AI will choose the words, sentence structures, and phrasing patterns that are the most common or likely in the human writing and literature that it has been trained on. AI writing choices share the same vocabulary and writing patterns as human writing. To that end, it is reasonable and statistically probable that a human writer and a GPT can be given the same instruction and respond with extremely similar answers.
- The presence of AI content does not establish the purpose or intent.
Word processing programs (e.g. Word, Google Docs) use machine learning techniques to assist with grammar and spelling within a document. In providing suggestions for correcting grammar mistakes, it is offering the most common solution that the AI has determined through its training. In addition, there are online services that offer to fix or even help rephrase text; those services use the same kind of machine learning to make those suggestion. To that end, an AI detector cannot determine whether the AI detected content is the result of these aforementioned writing tools or the use of a GPT to write an essay or paper on behalf of the student.
With that in mind, the current research indicates that a positive result for AI content is unlikely be sufficient to prove academic dishonesty without further investigation or corroborating evidence.
Note: this is a sampling of AI detectors
