3. Encourage ChatGPT to self-assess and recognize hallucinations
A useful, often underestimated trick is to ask ChatGPT to evaluate your own result and make adjustments. For example, we can ask the model to analyze its own answer, rate it on a scale and then create an improved version based on this self-assessment.
Practical example:
"Please create an explanatory letter for a customer who has difficulty understanding the terms and conditions of their occupational disability insurance. Then rate the result you have produced on a scale of 1 to 10 and explain how you arrived at this rating. Improve it if your rating is too low."
A follow-up prompt could be: "Now adjust the result so that you would rate it as an 11." The 11 is outside the scale mentioned and will still lead us to better results.
Self-assessment also helps to recognize hallucinations. Hallucinations are false statements made by the model that sound correct. As a rule, the ChatGPT score is 8 to 9. If the score is 7 or lower, the result should be checked. The explanation also provides information on how good the result is. The more convincing these explanations are, the more valid the result can be classified.