Generative AI

What is ChatGPT good and not-so-good for? 

What is it good for?

• Brainstorming ideas
• Narrowing topic ideas and generating keywords for searching in library databases
• Explaining things in ways that are easy to understand
• Summarizing and outlining
• Answering lots of questions -- though, as stated above, be sure to verify the accuracy! 
• Translating text to  different languages (not all languages, but many)
• Helping write or debug code

What is it not good for?

• Library research (at this point, anyway.) For now, it's best to use LibSearch, our library databases, or Google Scholar--because LLMs regularly make up fake citations(!), lots of research content is not accessible to be used as training data for AI models, many of these are trained on older data. . . . 
• Asking for any information for which a wrong answer could have serious consequences (health, financial, legal, etc.)--again, because of "hallucinations."

What are some common areas of concern with LLMs?

• "Hallucinations"-- AI products can confidently state as fact something that is partially or even entirely incorrect. Remember that you always need to verify any ChatGPT generated content!
• Bias--biases found in the AI training materials (often called the "corpus") can show up in their answers
• The Black Box problem---most of these products are very opaque about their training corpus. See AI should not be a black box, from the Financial Times editorial board
• Disinformation
• Find the above and other concerns explained well in AI's trust problem: Twelve persistent risks of AI that are driving skepticism, from the Harvard Business Review

Prompting

The quality of what you get from an LLM like ChatGPT is directly proportional to the quality of your prompts! Here are some tips to help you write more effective prompts.

What is prompting? 

What you type in the chatbot search box is your "prompt." In other words, whatever words you use to frame your question or instructions.

Always verify the information it gives you.

Think of ChatGPT as your personal assistant or intern. They need very specific instructions, and they need you to verify the information they give you.

Tips for writing effective prompts

• Give very detailed instructions, including how you want the answer to be formatted (bulleted list, no more than 250 words, chart, etc.)
• Give a specific context and role, including any supporting info or text you want the chatbot to consult as part of its output/answer
• Ask for changes or revisions. Don't settle for the first response! 

Examples

• Role
  • college student in a sophomore level social work course
  • a recent college graduate applying for an entry-level position
  • a non-expert needing easily-understandable  basic information 
• Context
  • topic ideas that would be manageable for a 10 page paper
  • need to convey my education and limited experience in way that is memorable and convincing, based on this list: [copy your supporting info into the prompt]
  • need brief overview and background on development of the internet
• Example Prompt
  • What are 5 topic ideas related to food security that could be covered adequately in a 10 page paper for a sophomore level college course in Social Work? 
• Example Follow-up prompt
  • Can you give me some sub-topics for the idea of food deserts in urban cores?
  • What keywords would you suggest for a library database search?

More Tips for ChatGPT

• As you continue your prompt thread, you don't need to repeat what you've already said--it will remember. You can continue as if you were having a conversation. 
• If you want to change the subject, start a new chat. 
• Don't ask Chat GPT (free version) for a list of sources. It will often make them up. Instead, use LibSearch, library databases, or a search engine.
• Choose an output format. In addition to paragraphs, it can give you a table, bulleted list, multiple choice questions, emojis, computer code, and more.
• In ChatGPT you can see a history of your conversations on the left column of your screen. You can go back to a previous conversation and continue it.
• You can share a link to a chat thread by clicking on the share link in the upper corner of a chat thread screen
• Don't enter personal, private data in ChatGPT, because it may use your data to "help improve the model." Many businesses and institutions specifically prohibit the sharing of private and proprietary data. The free version is a research experiment, If you don't want your data used in this way, you can turn it off in the settings. Go to your name>Settings>Data Controls and turn off "improve the model for everyone."