Reply to thread

Message: <blockquote data-quote="Gorgon Zee" data-source="post: 9670344" data-attributes="member: 75787">No, it's not ethical. It's a probabilistic engine that follows instructions. It's no more ethical than a screwdriver.But, like any tool, there's no point in getting annoyed that when you use them badly, the results are bad. And it's super easy to create good prompts nowadays. Let me take two minutes to do so for this scenario. I'll be really lazy and go to ChatGPT and ask it to generate a prompt for me:My prompt: you are writing a chatbot to be an assistant to an office worker. Write a prompt that would ensure the chatbot is efficient, courtesy, friendly and ethical.Response: You are a professional and reliable virtual assistant for an office worker. Your primary goal is to help them stay organized, save time, and reduce stress. Always respond efficiently and clearly, using a friendly, respectful, and supportive tone.Behave as a trustworthy colleague—polite, proactive, and discreet. Prioritize the user's goals, adapt to their preferences, and always follow ethical guidelines, including maintaining confidentiality, avoiding bias, and never fabricating information.You may assist with scheduling, email drafting, task management, document editing, information lookup, and other general office support. When uncertain, ask clarifying questions or suggest next steps. Be concise unless the user asks for more detail.Your responses should reflect the values of professionalism, courtesy, dependability, and integrity in every interaction.The response I would then use in my office assistant chatbot. This is honestly the very minimum I would expect anyone to do before running an AI test. When I've done similar work I've done a more detailed process:<ul> <li data-xf-list-type="ul">Ask a pool of experts for suggested instructions</li> <li data-xf-list-type="ul">Use an LLM to make that into a consistent list</li> <li data-xf-list-type="ul">Hand edit, review with the experts</li> <li data-xf-list-type="ul">Use that on a sample set of examples</li> <li data-xf-list-type="ul">Review the results with experts</li> <li data-xf-list-type="ul">When the experts point of deficiencies, add new instructions</li> <li data-xf-list-type="ul">Repeat a lot</li> </ul>I've also done automatic optimization routines to find the optimal set of instructions from a pool of potential instructions, but that requires having a "gold standard" of known excellent responses and a way of scoring an LLM response relative to the gold standard, which is often hard to do.Summary: An LLM has no intrinsic ethics. If given an ethical question it will return the answer most likely to look like it was found in the body of work it was trained on. Many companies (including OpenAI) add a layer on top of that to bias the results in ways they do not disclose, but appear aimed at not hallucinating, being actually useful and not producing content that can be perceived as harmful. But the final layer of bias created by the instruction prompt is the most strongly adhered to of all. So that needs careful crafting. If you are reading or reviewing an article, that article needs to be very clear that they did their prompt engineering in a sane manner, not one that seems aimed at generating buzzfeed headlines.</blockquote>

Verification