Dealing with ChatGPT - but how?

We asked

"When should / can I use you and when is it better not to?"

The answer from ChatGPT

"As a language model, my main function is to answer questions and provide information [...]
However, there are some limitations to my use. I am able to have human-like conversations, but I am not a complete replacement for human interactions. When it comes to complex topics that require specialized knowledge, it may be more useful to consult an expert in the field.
Also, I can sometimes provide inaccurate or inappropriate answers, especially if the question is unclear or ambiguous. [...]"

From the second paragraph onwards, ChatGPT itself points out important limitations (see next section for additions), such as:

• lack of specialist knowledge on complex topics or
• inaccuracies in answers (which may be somewhat "understated")
  ... and finally logically calls for "critical thinking and consulting other sources" in order to check the accuracy of the generated output.

When using ChatGPT, there should be an awareness of the tool's limitations:

• Not correct, but probable
  ChatGPT is not trained to output an answer that is as correct as possible, but one that has a high probability of matching. It therefore calculates how likely a word is to follow the last one (based on its training corpus).
  This means that output generated with ChatGPT potentially looks very convincingly factually correct, but does not have to be correct in the sense of the above-mentioned factual knowledge.
• Questionable source information
  If you ask ChatGPT for sources, source information is generated according to stochastic principles. These sources look very realistic and often also refer to existing journals, books or even pharagraphs, but these sources usually do not exist in the way indicated.
  (the following External link opens in a new window:2-minute video) gives an insight)
• ChatGPT is a text AI and not a computational AI.
  Only what has been trained and then also output based on probabilities can be "calculated".
• No current knowledge
  The current version is only trained with data up to 2021.
  A connection in the Bing search engine in the Edge browser or the use of add-ons such as WebChatGPT can reduce these restrictions. When installing any unchecked browser add-ons, remember that a browser add-on has access to all data available in the browser.
  While ChatGPT displays the ZLL of the HFU either as "Center for Logistics and Supply Chain Management", as "Center for Teacher Training and Teacher Qualification" and only after several inquiries correctly as "Center for Teaching and Learning", ChatGPT finds the correct solution at Bing right away.

Click here for our Insight entries...

• Internal link opens in the same window:Teaching in the age of AI (ChatGPT)
• Internal link opens in the same window:Studying and learning in the age of AI
• Internal link opens in the same window:ChatGPT and academic writing

Prompts are work orders that are given to an AI.
In prompt engineering, you try to improve the prompts so that the result is as accurate as possible.
Our top tips for good prompts:

1. As much context and info as possible, for example:
   1. Desired action (explain, summarize, make it easier to understand, formulate headings / outlines, etc.)
   2. In the desired style (polite, collegial, for a superior, as marketing text, for social media, etc.)
   3. Desired target group (for a layperson, for a primary school child, for a colleague, etc.)
   4. Taking into account the context of the content (e.g. previous knowledge, previous experience, key words, etc.)
   5. Taking into account the examples given
       
2. Clear definition of the task/output, for example
   1. Length of the text
   2. Format (mind map, table, text, keywords)

Examples

Not:
Ask me exam questions on the topic of "Using videos in teaching"

But rather:
Ask me 5 exam questions on the topic "Use of videos in teaching". I am studying XYZ. The topic is "Using videos in teaching". For example, we went through the SAMR model, different video formats, different research on the topic of video.
My professor likes to ask exam questions in which we have to explain things in a differentiated way. For example: explain why the SAMR model is or is not a good model in terms of the use of videos in teaching.
Discuss two current research findings.

Not:
Write me an email apologizing for an appointment that I can't make.

But rather:
Write me an email apologizing for an appointment that I can't make. Write 3 sentences to my colleague. Unfortunately, I have a doctor's appointment. We are on a first-name basis and rather informal.

And last but not least: Ask ChatGPT what info he needs and chat with him:
Have new responses generated via "Regenerate response" or chat with ChatGPT by providing more info following an unsatisfactory response until you are satisfied with the output.

Sources: ChatGPT, https://shribe.de/chatgpt-prompt-engineering/