HokieAI vs. OpenAI guardrails

Virginia Tech introduced HokieAI on August 23rd, 2026, bringing an exciting new advancement into the Fall 2026 semester. The Large Language Model (LLM)’s purpose is to provide accessible tools to all students, staff, and faculty. “The News Feed,” which is a local online news page covering the New River Valley region, explains that HokieAI was developed to build an understanding for users about ethical usage and implementing security. In the artificial intelligence world, security defines the term guardrails.

What are guardrails, you might ask? They act as security layers where instructional models check each input and output based on the policies developers add during the LLM pre-training process. The first process guardrails go through is called input checking, which effectively scans user prompts, ensuring no malicious content appears before the LLM provides a response. On the other hand, output checking scans the LLM’s solution before revealing it to the user to avoid unwanted bias or hate speech. Despite HokieAI’s six constructed models originating from OpenAI, the models were strictly developed to ensure stronger protections from dangerous prompts and to allow user interaction with artificial intelligence to be efficient for personal learning. 

Instances at other universities display a harmful effect of when guardrails are not properly implemented. A specific reason may include developer error when fine tuning the model during the model’s input and output checking. Vanderbilt University’s model, Amplify, created by the Computer Science department demonstrated a struggle with the integration connection of the plagiarism tracker software, Turnitin, which led to its permanent shutdown due to high false-positive error rates of assignment accuracy which contained various misinformation of differentiating between copy paste of external sources and student context of paraphrasing the original source. Vanderbilt University currently has Amplify available for faculty, staff, and students to establish personal learning of utilizing artificial intelligence however a disconnection of the model to a third party software differs from nationwide public concern where guardrails can extremely affect civilian lives or privacy. 

47% of domestic civilians were concerned and raised questions about controversies that involve public concern of artificial intelligence advancements, mainly about user safety from harmful topics. The April 17, 2025 Florida State University mass shooting resulted in the deaths of two faculty and seven injured students. The suspect, Phoenix Ikner, utilized OpenAI during the day of the shooting, as revealed by the State of Florida Circuit Court. Ikner typed prompts such as “how the media portrays a mass shooting”, “what firearm/ammunition to use”, and “best time to arrive on campus”. The model provided a sensitive response which seemed beneficial for the perpetrator with the number of national media coverage of mass shooting victims. This resulted with Florida being the first state to sue OpenAI’s CEO Sam Altman, with reasons that it helped the perpetrator with the shooting and its effect in harming children. OpenAI’s response was brief, stating that ChatGPT was not responsible for the attack and that the model focuses on providing external sources from the Internet in providing solutions. OpenAI continues to be in demand, with half of the United States adult population reported utilizing ChatGPT for daily tasks. In multiple legal battles that involved OpenAI, the basic reply from the company is to “implement stronger safeguards”. 

To establish a degree of independence from OpenAI models, Virginia Tech modified HokieAI’s GPT submodels by avoiding rewriting the base model parameters and including an architectural security layer for data privacy/management. This allows “no model retention,” which analyzes each prompt via output checking, ensuring all input and output are contained inside the university’s cloud service: ECAP (Enterprise Cloud and AI Platforms). Administrators or students are not allowed to check what users have typed previously, while the model logs are disabled and are only accessible to university legal auditors.

HokieAI models have a set time limit for high-intensity workflow due to a continuous traffic from 70,000 Virginia Tech affiliates, with a circular usage tracker for model utilization (similar to a financial budget planner) that aids with the university’s data demand, costs, and stability among all faculty, staff, and students. Another implementation of HokieAI guardrails is that it does not allow users to customize a model using an application programming interface (API), even after logging in with their Hokie PID. This allows more security from unwanted server addresses into the main interface.

With the growth of artificial intelligence since its origins, HokieAI and its levels of guardrails implemented should be seen as a robust model for internal Virginia Tech use and other institutions to follow, which would ultimately provide accessible information about the usage and security of artificial intelligence models.

By introducing transparent AI tools, universities, including Virginia Tech, are creating an academic environment where students can safely explore emerging technologies without risk of privacy leakage or harm to university policies. 

HokieAI and various AI tools in universities also increase digital literacy across majors, helping students understand how LLM systems operate, how guardrails protect users, and why secure design matters in real-life applications. With the emergence and high success rate of HokieAI, it displays innovation and safety can coexist, offering as a final prototype to institutions seeking to adopt artificial intelligence in a way that empowers learning while maintaining user trust.