Skip to Main Content

Faculty Guide

Guide to effectively using library resources and services at HIU.

Defining AI

Through machine learning, AI is a computer "program" that imitates human like interactions. By training an AI with thousands of books and articles, most AI has encyclopedic knowledge and can answer basic questions. Most AI is also trained with large data dumps of social media to help the AI with creating something akin to natural or current language. With a series of prompts most AI is capable of creating long form answers to questions and prompts. Those long form answers may be paragraphs or entire essays. The AI is more like a narrator than it is a search engine; it responds more like an author.

Because of the training of AI it is prone to certain problems.

  • All of the training was treated as equally correct. It learned facts from encyclopedias, but was also exposed to trash from social media. Most AI hold many contradictory ideas as being equally true. Donald Trump both won and lost the 2020 election.
  • It will be wrong with many answers (see above about training). It was trained on myths, misinformation, disinformation, and outright conspiracy theories.
  • It may even intentionally lie. Deception is a part of a lot of literature and the AI mimics that.
  • All of the training taught the AI certain genres and forms of literature. The AI will shape its answers/responses based on how you start a prompt. The AI have been known to be flirty because people asked it about itself (checking autonomy) which the AI recognized as flirting in social media. AI will also be mean and crude based on its training with AI. Start a prompt with brief terse words and respond in more brief interactions and the Ai will recognize a "flame war" or a fight from Twitter and Reddit. In a few sentences the AI will move from answers to all caps yelling to calling names.
  • Much of the AI training was with public domain (read that as really old) material. This leads to two distinct problems: the AI is prone to promote old, outdated, and disproven theories AND the AI will use language and terminology that is considered inappropriate now (sexist and racist comments are fairly common).

Most of these are called "hallucinations", but it is the AI relying on its training. "Garbage in. Garbage out."

Additionally, the AI is programmed to put words together based on probability of what should come next. It has learned what typically follows certain words or phrases. So, AI tends to be generic lifeless writing because it writes what is expected. It isn't creative with expressions.

Some Assignment Ideas to Stop the Bot (OK, really just speedbumps for the students)

  1. Require rough draft turned in early
  2. Require reference page turned in early
  3. Require more formats of information
  4. Require search strategy turned in as part of assignment
  5. Require resources come from library’s licensed database
  6. Require evaluation of resources (form, annotations, footnotes, or appendix)
  7. Require notes turned in with paper
  8. Require the student to interact with lectures and textbooks in the paper. Require personal connections.
  9. Use one of the AI detectors out there

Prompt Writing / Prompt Engineering

The AI was trained on books and novels; huge amounts. Most current AIs tend to write like an author or narrator. To get the most out of the AI, it is then, sometimes best to not ask it a question, but to set up a story. There are three parts to creating these types of prompts: create a believable character, use flattery to make the character better than most people, but not overly perfect, and then finish the prompt as an opening of dialogue. With that prompt, also be sure to keep clarity, conciseness, and specificity.

Don't do this: What are ten uses of a frying pan?

Do this: Paul, a culinary arts student, asked his instructor, "Chef, what are the many uses of a frying pan?" And the Chef responded...

Hint: don't overdo the flattery because, in literature and movies, someone that is too good at something often has flaws; too smart and they make common mistakes, too athletic, and they trip, too righteous and they sin. The AI will see a "perfect" character and present their flaw, even if you don't ask for it. Sometimes the AI "mistake" is because it is narrating a character flaw.

Prompt Engineering According to Anthropic (Claude) - 11.24

https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview

  1. Prompt generator
  2. Be clear and direct
  3. Use examples (multishot)
  4. Let Claude think (chain of thought)
  5. Use XML tags
  6. Give Claude a role (system prompts)
  7. Prefill Claude’s response
  8. Chain complex prompts
  9. Long context tips

For a Glossary of Terms

For More About AI

Much of this comes from: https://sites.google.com/view/practical-information-literacy/beating-ai - created by Steve Jung

On that site are links to several PowerPoints on AI.

Some PowerPoints are on how to stop students using AI, on academic integrity in light of AI, and several on possible uses of AI for students and faculty in a university.