Trust … but Verify

If you are new to using AI large language models (LLMs), you may be struggling with how to harness the power of generative pretrained transformers. GPTs are LLMs that have been trained via deep learning to generate text that appears to be authored by a human, but that have actually been trained on massive data sets of both code and text.

LLMs, like GPT, don't think in terms of topics or concepts. They think in tokens. At a basic level, tokens are groups of characters that sometimes align with words. When you ask GPT a question, it does not really know the meaning of what you've asked it. It simply knows which tokens you fed into your prompt. It runs those tokens through a black-box algorithm that contains millions of bits of math to create an algorithm that spits out tokens related to what you've input.

GPTs can logically respond to human questions or prompts, creating high-quality content. However, there are some risks associated with using GPT that the newcomer should understand to avoid problems.

When you use a GPT to develop a rough draft, always closely review and edit that draft as needed. Never allow the GPT to create the final text you submit. There are a handful of risks you must consider beforehand. Ask yourself:

How Thoroughly Was the GPT Trained?

Consider a GPT that was trained to follow a flow of tokens. If token A connects to token B and only token B, when the GPT is faced with token A, it will next go to token B. Perfect!

But say token A in training is connected to more tokens than token B - for example, tokens C, D, E, and F. In that case, when the GPT is generating text and gets to token A, it will need some help to make sure that it chooses the correct path of interest to you. Without that human direction, it might write about token C when you really wanted it to write about token F.

How do you help the GPT avoid that problem? Give it additional prompts that will let it understand that you care about token F. For example, you could ask the GPT to tell you about PMW 160. PMW 160 is both the name of a commercially available camcorder and the name of a government program office. It may not “know” that it gave you a “wrong” answer; you will need to supply additional prompts to get the GPT back on the right track.

Have You Addressed Possible Hallucinations?

In our case above, suppose that the GPT was only trained as far as token A, and you want to see what it can write about token B. A peculiar thing about GPTs is that they seem to be reluctant to tell you they are unable to do something. Instead, they sometimes “hallucinate,” in which they “make up” an answer that is not grounded in their training and may be nonsensical. And yet, it can sound very much like a correct answer about which you had known nothing.

How can you tell if your GPT is hallucinating? Write additional prompts, such as asking it to cite the references that it used for its answer. Of course, you can also do additional research on the heretofore unheard-of material.

What Materials Were Used to Train Your GPT?

When using a GPT, it’s important to consider any biases in the material it is trained on. If some of the training material was from a 1950s Reader’s Digest, your GPT may write about women belonging in the kitchen rather than the workplace. Dated training material can induce many racial, sexual, and religious biases that you would not expect in scholarly writing in 2026. If the GPT has been trained with internet entries from social media, some of the answers you receive could reflect far-left or far-right ideologues, spouting venom or unsubstantiated theories not supported by fact.

How can you identify questionable responses? Give more prompts. You might ask the GPT to provide a reference for each of its ideas.

How Timely Is Your Training Material?

If you are asking your GPT about very recent developments in technology, be aware that it may have last been trained prior to the emergence of that technology.

There are two things you can do to determine if your GPT has not been trained on the material in question: 1) check to see how recently the GPT’s last training took place, or 2) input additional prompts designed to uncover what the GPT was most recently trained on regarding  the topic you are inquiring about.  

But your best recourse to getting the answer you seek may ultimately be to use another GPT, either alone or in concert -with the GPT you are currently using. (While not difficult, working with multiple GPTs simultaneously requires more than an entry-level understanding of AI.) There are many different GPTs out there (e.g., Microsoft's Copilot, Anthropic's Claude, OpenAI's ChatGPT, Google's Gemini, and Google's Search Experience). It’s wise to choose a GPT that is suited to the project you are working on.

Remember, Your GPT Sees Nothing Wrong with Plagiarism

Of course, plagiarism in the work you submit for academia could get a failing grade or even get you expelled, since some instructors now use products that are designed to detect plagiarism. Not to be discounted is that plagiarism by your GPT could violate copyright laws and make you vulnerable to a lawsuit with potential financial penalties.

How can you be certain that the GPT you are using is not plagiarizing someone’s work? Once again, prompts asking the GPT to identify sources for key pieces of its work can help detect potential plagiarism. It’s fine to have GPT organize your thoughts or develop a first draft, but you must turn that input into a final product that you can submit with pride and without fear that your GPT has set a trap for you. While you want to be able to trust an impressive write-up developed by your GPT, it is crucial to verify what it gives you. When you initially prompt a GPT to address a topic, start broadly, use follow-up prompts to clarify and verify, and scrupulously review and edit the product of your joint effort.