(David Wang | Student Life)

The University has upgraded its secure WashU ChatGPT model to ChatGPT-4 Omni (ChatGPT-4o). The upgrade was brought about by the Digital Intelligence & Innovation Accelerator (DI2), the group responsible for leading WashU’s Digital Transformation initiative. 

According to an email by the Digital Transformation Chair, Dr. Philip Payne, and Chief Information Officer, Jessie Minton, the upgrade boasts double the previous text generation speed and an experience more congruous with current, public large language models.

“[The] ChatGPT-4o ‘Omni’ model [is] one of [OpenAI’s] most advanced artificial intelligence models capable of complex problem solving,” Payne and Minton’s email read.

WashU’s version of ChatGPT, which was launched in December of last year running the ChatGPT 3.5 model, remains free for use by the University community with the upgrade. OpenAI’s free plan provides limited access to ChatGPT-4o, with a fee required for less limited use.

The University’s choice to make its ChatGPT free is part of DI2’s push for technology equity among students. Dr. Albert Lai, the deputy faculty lead for DI2 and a leader on the University ChatGPT project, said that making this tool free “levels the playing field.”

“These AI tools have been made available to all our students free of charge, instead of having some students who are willing to pay for these chatbots…versus other students who may not be able to or are unwilling to pay for that service,” Lai said. 

ChatGPT-4o, notably, increases support for non-English languages, a feature Payne and Minton highlighted in their email. Lai said that this, too, was an important feature for equity.

“Computational computer science and Silicon Valley have been very heavily English language-focused at a high level,” Lai said. “And, you know, the world is not entirely English-speaking. I think having [the AI]…be better at multiple languages is important for equity purposes.”

Dr. Natalie Monzyk, Assistant Director for Educational Technology at the Center for Teaching and Learning (CTL), said that an AI proficient in languages besides English can be a boon for multilingual students.

“I do think that, increasingly, our student body is multilingual,” Monzyk said. “So, if you are having trouble understanding a concept or something, having it explained to you in the language that you feel most confident in for that particular subject is really, really useful.”

Monzyk added that the CTL is helping teachers bring AI into the classroom. One difficulty she highlighted was the wide variety of opinions on whether or not ChatGPT should be in the classroom.

“We don’t tell faculty whether or not they should include AI [in the classroom],” Monzyk said. “We try to facilitate discussions so that faculty feel confident making intentional choices about using AI in the classroom or not.”

However, Monzyk said that she has seen “a lot” of professors come to the CTL for help introducing WashU’s ChatGPT to the classroom. 

To help professors come up with ways to implement AI into their course, the CTL has partnered with the University Libraries to launch an “Open Scholarship Institutional Repository” where professors can submit their ideas for generative AI teaching activities for others to see and use.

Researchers on campus are also using WashU’s ChatGPT thanks to its security. According to DI2, WashU ChatGPT is HIPAA and FERPA protected. The University also has a deal with Microsoft, which prohibits OpenAI and Microsoft from viewing WashU user data or using it for training their large language models.

While this security has been part of WashU’s ChatGPT since its initial launch, the upgrade did add several secure “API endpoints,” which Lai explained were useful for the “high volume/bulk requests” used in research. API endpoints connect AI with other services.

“We have some research happening where we want to extract phenotypes or patient characteristics from clinical records,” Lai said. “It’s one thing to say ‘I’m going to paste in a patient record and ask it some questions,’ but if I want to do this on a population, you would want to use…[a secure] API endpoint to do that.”

Assistant Professor of Computer Science and Engineering, Dr. Umar Iqbal, who is a computer security and privacy expert, expressed concerns about large language models like ChatGPT being used in the medical space. He said that he was hoping for more communication from the University about how WashU’s ChatGPT was being used.

“I don’t know how WashU is using ChatGPT,” Iqbal said. “I can think of some custom cases where it could be used, but I think those custom-use cases seem somewhat sensitive, and in those sensitive environments you need more scrutiny.”

Iqbal warned of AI “hallucinations,” where a large language model produces false information, and encouraged the scrutinizing of specific use cases.

“I don’t know how exactly WashU’s custom ChatGPT is intended to be used and whether the stakeholders who would use [it] know the limitations of such models,” Iqbal said, “There are obvious safety harms because you don’t know [if] what it is generating is actually factually correct.”

Lai and the Institute for Informatics, Data Science, and Biostatistics (I2DB) team have done research into the problem of AI hallucinations to try to ensure the accuracy of WashU’s ChatGPT and have worked to minimize mistakes.

“That’s an active area of research,” Lai said. “What we’ve noticed so far is, in general, you will get fairly consistent responses back with these large language models, which is reassuring. I think there are ways to mitigate some [hallucinations]. In our user interface we have ‘creative,’ ‘balanced,’ and ‘precise’ as options. If you choose ‘precise,’ it is less likely to hallucinate.”

As artificial intelligences like ChatGPT are a burgeoning technology, there is much discussion about their use. Iqbal said that a hundred papers were being written every day on security and privacy issues alone. 

Both the CTL and the DI2 offer routes for community input on WashU’s ChatGPT: the CTL offers consultations on their website or by booking through their email, and DI2 hosts weekly coffee hours where anyone can join to discuss their thoughts with Digital Transformation leaders.