ICTworks

In the rapidly evolving landscape of Generative AI, tools like ChatGPT have emerged as powerful resources for humanitarians. They can perform a wide range of tasks, from writing creative content to providing detailed explanations on complex topics.

However, understanding how to effectively utilize these tools in international development requires a nuanced perspective.

3 reasons you must think of ChatGPT as an overeager graduate assistant:

• ChatGPT will always give you an answer, even when it’s wrong.
• ChatGPT needs clear guidance, aka prompt engineering.
• ChatGPT responses must be reviewed and fact-checked by a human.

ChatGPT: An Overeager Assistant

GenAI is designed to be highly responsive and versatile. It can create agendas, draft speeches, write reports, and even engage in persuasive discussions. This adaptability is one of ChatGPT’s greatest strengths – and greatest weaknesses.

Like a new graduate assistant, ChatGPT is brimming with enthusiasm to prove its worth. It can rush to complete a task without fully understanding the nuances, leading to errors and hallucinations.

This is especially common in international development topic areas where there isn’t much factual evidence but plenty of organizations claiming success with minimal data. For example, I’ve had ChatGPT make up entire USAID programs that sound plausible but don’t exist when I’ve asked generic questions and did not demand references.

Subscribe Now for more digital development insights!

Give ChatGPT Clear Directions

The quality of ChatGPT output is directly proportional to the quality of your directions, or “prompt”.

Imagine asking a graduate assistant to “write a paper on quantum computing in humanitarian aid” without any further guidance. The assistant will produce a paper, but it could be too broad, too narrow, or miss key points. And who knows if it will be 3, 13, or 30 pages long.

Better is a clear prompt that tells ChatGPT how to analyze the problem, a clear problem to analyze, and the result you expect. For example, here is the result of this prompt:

“You are a journalist. You are writing a 500 word article for a major newspaper with college educated readers. Explain what quantum computing is and how it can improve development outcomes for humanitarian organizations working in developing countries.”

There is a whole body of knowledge around prompt engineering too, which can lead you to over-engineering your prompts – spending more time searching for the right question than just editing what ChatGPT gives you.

Always Verify ChatGPT Responses

Just as you would rigorously review a graduate assistant’s work before passing it on to your boss, you must scrutinize every ChatGPT response. This involves cross-referencing facts, checking for logical consistency, and ensuring the information is current and relevant.

ChatGPT can produce responses that seem plausible at first glance but are factually incorrect or lack the necessary depth. They do not possess true understanding or consciousness; they are sophisticated pattern recognizers that excel at mimicking human-like text based on their training data.

For example, if you ask ChatGPT questions about social development programs using ICT4D in Francophone Africa, you will get wildly vague or inaccurate answers. Even Perplexity, the gold standard in quality GenAI responses, is unable to give good answers to questions about the gender digital divide in small Pacific Island nations.

Like a good graduate assistant paper, GenAI outputs are best used as starting points or supplementary resources for your own investigations, rather than definitive answers suitable to verbatim usage.

Please, please do not ever submit unedited GenAI output as your own.

For example, this is the original output of my starting prompt for this post, which is way too generic for the ICTworks readership. I used ChatGPT as a first draft, and then edited and refined this post to make it relevant to you, my respected reader. Please do the same when you use GenAI tools like ChatGPT.