Effective prompting strategies for human-like AI text responses and content generation.

Loading data from Talking to Chatbots Dataset reddgr/talking-to-chatbots-chats

😎 “I’m not the type of person concerned with ‘humanizing’ AI generated text, because it doesn’t make sense to ME (I don’t care if whoever published or sent me a text used AI or not, I care about the content), but the most effective type of prompt for getting responses that don’t ‘smell like AI’ is typically setting a limit word count for relatively complex tasks: i.e. “answer in no more than 12 words”. That kind of prompt essentially hinders the great ‘skill’ that LLMs have to generate text with extremely high quality and error-free that actually says too little or nothing meaningful (which is what’s so ‘hard’, and time consuming, for humans to replicate).

It’s either that, or the other extreme mentioned in other comments, consisting of feeding the context window (via prompt or files) with a large amount of text you want the LLM to mirror in style. The problem with this second option is that it either takes you too much time to produce the text in the writing style you desire to convey, or you simply have to ‘steal’ it from someone else. While I don’t see anything unethical about that, I think it makes no sense: the Internet is already filled with text, machines can generate and replicate massive amounts of text, nobody can read all the text that machines generate… people choose to read people and sources, not a particular text from some anonymous author just because it sounds more or less “human”. That’s just an opinion. To answer the question: iterate with “answer in no more than X words” with a meaningful prompts and combine it (or ask the bot to combine the pieces later) until you have something that looks human to you (if that’s the objective).”
🤖 “I’m not the type of person concerned with ‘humanizing’ AI-generated text, because it doesn’t make sense to me. I don’t care if whoever published or sent me a text used AI or not; I care about the content. The most effective type of prompt for getting responses that don’t ‘smell like AI’ is typically setting a word count limit for relatively complex tasks, e.g., ‘answer in no more than 12 words.’ That kind of prompt essentially hinders the great ‘skill’ that LLMs have to generate text with extremely high quality and error-free that actually says too little or nothing meaningful (which is what’s so hard and time-consuming for humans to replicate).

It’s either that, or the other extreme mentioned in other comments: feeding the context window (via prompt or files) with a large amount of text you want the LLM to mirror in style. The problem with this second option is that it either takes you too much time to produce the text in the writing style you desire to convey, or you simply have to ‘steal’ it from someone else. While I don’t see anything unethical about that, I think it makes no sense. The Internet is already filled with text, and machines can generate and replicate massive amounts of text. Nobody can read all the text that machines generate. People choose to read people and sources, not a particular text from some anonymous author just because it sounds more or less ‘human.’ That’s just an opinion. To answer the question: iterate with ‘answer in no more than X words’ with meaningful prompts and combine it (or ask the bot to combine the pieces later) until you have something that looks human to you (if that’s the objective).”

Note: Minor grammatical and punctuation corrections were made.

Hugging Face Dataset Metrics

All the conversation prompts, responses, and metrics are available to download and explore on Hugging Face dataset reddgr/talking-to-chatbots-chats: