ChatGPTOpenAI’s AI-powered chatbot, is an almost daily tool for many of us. And like all tools it has specific strengths and specific limitations. We use it to do homework, write emails or translate texts, but often we use improperly. We trust his answers without verifying them or we give him commands that are too vague, being disappointed when the result is not what we expected.
To learn how to use it better it is important to remember that ChatGPT it is not a search engine – like Google for example – and cannot be treated as such. Language models such as ChatGPT, Claude or Gemini generate plausible answers based on training data, but these are not always accurate. To obtain reliable or at least verifiable information, it is best to use tools such as Perplexitywhich cites sources, or ChatGPT Plus’s Search function.
Furthermore, to get useful answers, it is essential to write clear, detailed and contextualized instructions. General questions produce vague answers. Clarify the goaldefine the context and, if possible, provide exampleshelps ChatGPT better understand the request and generate more accurate and useful responses. When we interact with ChatGPT let’s remember to treat it like an assistant, not an oracle.
ChatGPT is not Google, let’s not use it (yet) as a search engine
A mistake What we commonly do when using ChatGPT is to treat it as if it were Google.
We ask him questions on the most diverse topics and expect the answers to be accurate. Unfortunately, that doesn’t necessarily mean they are. ChatGPT, like Claude, Gemini and Copilot, generates plausible sentences based on the data it was trained on, repeating sequences of probable words, a bit like a parot. This does not mean, however, that the sentences created are necessarily true and that they are not hallucinations.
To have reliable answers to our questions, therefore, we have a couple of options (in addition to using Google, of course):
- The best is to use Perplexityanother language model, optimized precisely to provide information by retrieving and citing sources from the web, so as to always allow us to check where they come from;
- If we have ChatGPT plus available, we can use the Search function by clicking on the web icon. Search can do that web searches using information from Bing, Microsoft’s search engine, but it’s not yet at Perplexity’s level. This feature is not currently available with ChatGPT in the free version, but can already be used if you have a subscription;
- If we have a file with the information we need available (for example: a book, an instruction manual, a scientific article) we can upload the file to ChatGPT or NotebookLM, another AI tool, and use it as the main source for answers to our questions. In this way we minimize the risk of invented information.
If we write the instructions in a generic manner, the response will be unsatisfactory
We often refer to ChatGPT with very generic and context-free instructionsobtaining accordingly answers likewise generic which may be unsatisfactory. For example, if we try to write: “Explain to me how the universe was born in simple words”, we will not get the answer we need. If, instead, we ask: “Explain to me how the universe was born, in particular the Big Bang theory. I am a 10 year old child and I am studying it at school. I have a test tomorrow and I need you to help me study” the answer will be more useful and accurate.
There is a lot of research on how to write an instruction wellin technical terms “prompt”, for a language model. Sometimes it’s more of an art than a science, but these guidelines can be a good starting point:
- Let’s clarify the objective: the more we can explain clearly what we need, the better. In the previous example, having added that we are interested in the Big Bang theory and that we need support in studying helps to narrow the field.
- Let’s define the context: if we delve into the context, we help ChatGPT generate sentences by choosing the correct words. Saying the age of the interlocutor and explaining that we need the answer to study will allow the model to choose the most suitable words. It might also be useful to ask ChatGPT to speak like Neil deGrass Tyson, a well-known astrophysicist and popularizer.
- We provide examples: Language models are very good at copying someone else’s style. If we can, let’s add an example from which ChatGPT can take inspiration. If we have to write an email to our boss and we have another one ready, we remove all the sensitive information and then put it as an example at the bottom of our prompt telling us to write it in the same style.
And if we really don’t know where to start, Anthropic, the company that created Claude, has made an entire library of prompts available to take inspiration from.
If we upload sensitive data, he may use it for future training
If we use ChatGPT to write our resume, analyze our finances or even just plan our next vacation, we have to be very careful about the data we enterWhy OpenAI collects data from users to refine the model’s performance. Luckily, you can control how your data is used by: clicking on your icon profile on ChatGpt, secting “Settings” and clicking on “Data Control” we see the “Improve model for all” option. By deactivating it, we can prevent the use of our conversations for model training, thus protecting any sensitive data.
