Generative Artificial Intelligence

AI Basics Overview

Generally speaking, generative artificial intelligence focuses on applications that use Large Language Models (LLMs) that generate responses to prompts using large pretrained data sets (often the internet but sometimes localized datasets). Example AI tools are Claude, ChatGPT, Perplexity, Co-Pilot, and Gemini. Many AI tools that also create images in a similar process (using diffusion models that were trained on images to produce visuals from text responses).

Source: Grammarly for Education and Higher Ed Dive's studioID survey of Higher Education leadership and Faculty, 2024

Prediction and variability

LLMs generate responses by predicting the most likely next word or phrase based on patterns learned from their training data. For example, if you ask an AI tool to complete the phrase “the cat in the _”, it would be statistically probable that based on the internet (should that be the LLM’s data set) the response should be “hat” as opposed to “spaceship”. LLMs generate responses probabilistically, meaning that the same question may yield slightly different answers. This variation occurs because AI does not retrieve pre-written responses. Instead, it predicts the most probable next words based on the given prompt, its training data, and system parameters (e.g., randomness (temperature)[1] which impacts the variability of responses; prescribed response length).

Understand what training data an LLM uses by either researching or asking the Application Programming Interface (API) tool. Most AI tools rely on pre-trained datasets. You can always ask the AI for the date in which its training data was last updated. Some AI tools will allow the occasional use of real-time access to the internet if needed. This is often limited due to safeguards against misinformation and copyright concerns. For those that use real-time information from the internet, it still uses the LLM to generate a predictive response and is not just copied from the source. Using real-time browsing allows access to the latest information but this can introduce variability and potential misinformation (by using unverified, inaccurate, or even malicious information). 

Source: Digital Education Council Global AI Student Survey, 2024

Unavoidable Bias

You interact with Application Programming Interfaces (API) powered by LLMs. LLMs are trained on human generated content – so they reproduce and amplify bias based on what is/isn’t in the dataset (e.g., if the API pulls from the internet, including Reddit, the responses you get will be predictions based on the content within). Depending on the data used by an AI model for training, it will have tendencies towards cultural, political and ideological perspectives. Inequities and marginalization are promoted through generative AI for multiple reasons: bias in trained data (which is discriminatory, promotes misrepresentations, or is unrepresentative), and through bias among users in a society that perpetuates these issues. Read more about this here. In contrast, if the database is peer-reviewed journals, the predictions will be based on that content (which can be different than direct citations). These are important limitations to always be aware of. 

 

Hallucinations in AI Responses

The types of hallucinations can vary based upon the API. For example, early ChatGPT critics complained that citations often included hallucinations (i.e., made up references). Even AI tools only trained on peer-reviewed research can still produce inaccuracies. This happens because AI does not ‘look up’ facts like a search engine. Instead, it predicts responses based on learned patterns. So, if an AI cannot find a strong match in its dataset, it may generate a plausible but incorrect response, a phenomenon known as hallucinations.

AI tools hallucinate in different ways depending on the AI tool’s design, data source, and safeguards. A model trained on general web content (e.g., Reddit) may generate conversational but unreliable claims, while a research-specific AI may hallucinate plausible but incorrect citations. Understanding an AI tool’s dataset and intended use can help you anticipate and mitigate hallucinations.

Source: Digital Education Council Global AI Student Survey, 2024

It is a good practice to verify the accuracy of responses pulled from AI tools. Here are some possible steps you could take:

Detecting the use of AI

Both humans and AI detection tools struggle to accurately identify whether work is produced entirely by the student, entirely by the AI, or collaboratively. This means that instructors may be unable to determine whether student work was created entirely by them, entirely by an AI tool, or collaboratively.

In addition, some software that students and faculty use have AI features behind the scenes to support work quality. Spell-checkers or grammatical recommendations (e.g., software programs like those in the Microsoft suite and additional AIs within Grammarly) further blur the lines between intentionally or unintentionally using AI. Essential tools like speech-to-text software are often flagged as being created by AI by detection software – even though students are not using AI to create their work.

Source: Digital Education Council Global AI Student Survey, 2024

Considering inaccuracies for false positives and negatives in AI detection software (including Turnitin), they should be used with caution and never the sole source of evidence for plagiarism.

In fact, some institutions have disabled AI detection due to reliability concerns and advise faculty against using AI detection tools like Turnitin's AI-writing detectors. Notable examples include:

• Vanderbilt University: After extensive evaluation, Vanderbilt decided to disable Turnitin's AI detection tool, citing concerns about its reliability and the risk of falsely accusing students of academic misconduct. 
• University of Texas at Austin: The university opted out of using Turnitin's AI detection feature, expressing doubts about its accuracy and the potential for incorrect accusations.
• St. Joseph's University: At the beginning of the fall semester in 2023, St. Joseph's University removed the AI detection feature from Turnitin. This decision was influenced by concerns regarding the reliability of the tool and its potential to produce false positives.
   

Currently, AI detectors (like Turnitin’s AI detection tool) struggle with reliably delineating between human-written text vs. AI-generated. Furthermore, AI detection tools cannot reliably determine whether a student used AI for brainstorming, editing, or writing entire sections. As AI-assisted writing tools become more common (e.g., Grammarly, Microsoft Copilot), distinguishing between acceptable and unacceptable AI use will require clear academic policies rather than just detection software. 
              This article discusses the accuracy rates of Gen AI detection tools. Note that the mean accuracy rate to properly label content that was AI generated across tools was only 40% and for Turnitin was 61%.  Any additional manipulation of the content to evade detection further reduces the accuracy rate (Turnitin accuracy decreased to 8%). There were also high false positive rates across AI detection tools (human generated text was falsely labeled as AI generated).  Fortunately, in this study, Turnitin did not produce false detections. Please also note that Turnitin has changed its stance that false positive rates are less than 1% to now state that they are 4% at the sentence level.

The Department of Education Office of Civil Rights: Avoiding the Discriminatory Use of Artificial Intelligence helpfully described (November 2024) scenarios where AI could be problematic, including false accusations of AI use by users of AI detection software.

Source: Grammarly for Education and Higher Ed Dive's studioID survey, 2024

Privacy of your prompts and shared documents with an AI  

When using AI tools, it is essential to always think about the content you share in a prompt or uploaded files. Avoid including sensitive, confidential, or personally identifying information. Not only do some AI systems not securely store their data, but they may use the content of prompts to train models.

If you have any privacy concerns about the content of your prompt and/or documents you are sharing with an AI tool, determine how that tool retains data, whether it uses your responses to train future models and whether it shares the data with third parties. Some AI tools explicitly state they do not train on user inputs, and some have “opt-out” features as a user’s setting.

When engaging AI tools to support UNH work, it is essential to only use USNH approved AI tools with strict data protections. See USNH Resources and Guidelines section below for more details.

Source: Digital Education Council Global AI Student Survey, 2024

Additional Ethical Considerations 

There are many serious ethical considerations for general AI. They include privacy issues, plagiarism, misuse and deepfake, bias and misrepresentation, inequities and marginalization, and high energy usage and pollution concerns.

Plagiarism: Many LLMs (but not all) are trained on datasets that scrape the internet. This means that the LLMs include all available data from broad sources, regardless of copyright status or usage permissions.  Anything scraped, including materials people have not intentionally posted or intended to share have trained the model. Therefore, when you submit your prompt to the AI tool, it responds with an amalgamation of content produced by others without considering permissions. Using content produced by others without their permission and without proper attribution (citation) constitutes plagiarism. Furthermore, when you use many AI tools (but not all), your prompts may continue to provide data to retrain the model regardless of your intent. Therefore, intellectual property infringement is a legitimate ethical concern when using AI tools.

Misuse: People can use AI tools to create scientific articles that are highly convincing but inauthentic. Read about the potential to create AI generated medical research articles here. In addition, deepfake is the creation of fake images, videos or audio seeking to imitate someone. Read more about deepfake here. You can also read about strategies to identify deepfake images here.

Sustainability (environmental costs): AI tools are computationally intensive; queries use more electricity than standard web searches, and this will increase global energy use. In addition, the data centers produce electronic waste (like mercury and lead), use rare minerals (which are mined un-sustainably), and waste a large amount of water (which is globally scarce). You and/or your students may reasonably object to the environmental cost of utilizing these resources which you can read about in more detail here.

Review guidance for Using Artificial Intelligence in a University Setting

Review the USNH Artificial Intelligence Standard

Review the Syllabus Guidance from the Office of the Provost Website

• see page 13 within the Fall 2024 linked documents for 2.6.1 Artificial Intelligence

AI Detection Software: 

In addition, here is a resource on AI detection Software Best Practices developed by Teaching and Learning Technologies.
 

USNH currently has two tools that include data protections:

1. Copilot is recommended for faculty, staff, and students. There is a free version and a Pro version of Copilot. The free version is available to all USNH employees and students. You may access this here, and you will be prompted to login to your USNH account. When you login, you will notice that Copilot has conveniently saved your recent chats for you to access. Your data will leave the USNH network, but it is not monitored, shared or provided to third parties. You will have to paste your queries into the prompt (the free version does not have upload capabilities).
   ***The Pro version also provides integration with Microsoft word, excel and outlook, as well as PowerBI reports. However, Pro currently costs $30/month. You can request licensing here.
2. DeepThought AI is an excellent choice if you have sensitive data, a high volume of queries, or require integration with other applications. The LLM runs on a high-performance computing cluster (data does not leave USNH).Chatbots in this platform are highly customizable. You can find out more about using DeepThought AI here.

Did You Know? Guide to Using AI in Research

• Since widespread generative artificial intelligence writing tools such as Chat GPT arrived, the research enterprise has been assessing their use with a variety of responses. The UNH Responsible Conduct of Research and Scholarly Activity (RCR) Committee has published a  Simple Guide to Using Generative Artificial Intelligence Writing Tools in Research and Scholarship at UNH to help researchers understand the fundamental issues with these technologies regarding research integrity.

The Connors Writing Center provides two resources:

• Writing and AI 
• Statement on AI Usage

Overview of Pedagogical Considerations for AI in Higher Education:

• Recording of FITSI Keynote Speaker José Antonio Bowen on Teaching and Thinking with A.I.
• 10 best Practices for AI Assignments in Higher Ed
• Revised Blooms Taxonomy that includes generative AI created by Oregon State University

AI policies and Syllabi:

• Resources and AI policies in your course created at Stanford University (includes  syllabus sample language)
• AI assessment scale for courses (syllabi and assignments)

AI Prompts and Activities for Courses:

• AI pedagogy Project (metaLab at Harvard) includes AI activities & implementation details
• NCFDD generated prompts for various higher education purposes (including teaching prompts)
• University of Central Florida: open source book with AI activity prompts (which you could also implement in Copilot)

​​​AAC&U Free Student Resource

• A Student Guide to Navigating College in the Artificial Intelligence Era 

Book Club: Teaching with AI (click here to register):
The Teaching and Learning Book Club is a 3-session series in which participants read “Teaching with AI: A Practical Guide to a New Era of Human Learning.” Each hour-long discussion will explore different themes from the book and how you might apply what you have learned in your courses. Wednesdays Feb 12, Feb 26, March 12, 11:00am - 12:00pm, (virtual). Please feel free to join us for any of the sessions that you are available to attend!

AI Basics for Instructors (click here to register): 

This session will discuss important details about AI for the classroom setting including AI literacy (e.g., how generative AI works; data and information privacy; unavoidable biases).  
Thursday, March 6^th 11:00 am – 12:00 pm (hybrid)

  You may review the slides for this presentation here.

Integrating AI Intentionally and Realistically in Small Bites (click here to register):

This session will review simple options for embedding AI into course activities. UNH faculty will share specific examples of their experiences developing and implementing AI activities in their courses. Thursday, March 13^th 11:00 am – 12:00 pm (hybrid)

  You may watch Lisa Owen's portion of the session here, where she describes her use of AI in her Occupational Therapy classes.

Continuing community discussions around AI (click here to register):
Many efforts are underway to raise awareness, understanding, and effective use of AI within the UNH classroom. This will be an opportunity to provide some updates to one another and to be a place for the community to share with one another what’s working and what is still challenging. This is an E3 think tank, proving an opportunity for open discussion among participants. Tuesday April 8^th at 11:00am – 12:00pm (hybrid)

Academic Honesty and AI Policy: Transparency, Consistency and Student Conversations (click here to register):

This session will discuss being transparent and consistent in your approach to academic honesty in your syllabus and throughout the semester. In addition, we will talk about fruitful discussions with your students around AI policies and academic honesty in general. Wednesday, April 9^th at 2:00pm – 3:00pm (hybrid)

  You may watch the video here and you may access a summary document here.

• Educators failed students with social media. We can do better with AI - March 13, 2025

• Teaching Tip: Navigating AI in the Classroom | Inside Higher Ed - February 6, 2024
• Embracing Artificial Intelligence in the Classroom | Faculty Focus - December 6, 2023
• Artificial Intelligence: The Rise of ChatGPT and Its Implications | Faculty Focus - August 25, 2023
• AI Eroding AI? A New Era for Artificial Intelligence and Academic Integrity | Faculty Focus - July 19, 2023
• Level Up Higher Education Assessments with ChatGPT | Faculty Focus - May 3, 2023
• Article - Turnitin - AI Detection (Fa... (unh.edu) - April 18, 2023
• Episode 57: Friend or Foe? Faculty Focus Live Podcast | Faculty Focus - April 13, 2023
• How to cite ChatGPT (apa.org) - April 7, 2023
• Will ChatGPT Change How Professors Assess Learning? (chronicle.com) - April 5, 2023
• Write Free or Die Vol. 9, Issue 2 - March 30, 2023