1 The Genie

GPT (Generative Pre-trained Transformer) is a technology that relies on ML (Machine Learning) to build a model. The model, in return, is used by a transformer to generate content.

Extremely simplified, a trained model is a stochastic action suggester. Based on the current state of an application, the transformer mechanism probabilistically predict what should occur next. One of the possible action is chosen based on the probabilities.

GPT also utilizes a mechanism called “attention” to maintain a context. This is partially why ChatGPT can seemingly generate text that is self-coherent.

The model utilized by GPT is trained using a vast amount of text content. This includes all the text that is freely available two years ago on the Internet. To give us a sense of scale, the entire web site of Wikipedia contributed about 3% of the content used to trained GPT.

While the learning mechanism of GPT is sophisticated, the real power GPT has to do with brute force. The sheer amount of text data used to train GPT is one aspect. The second is the number of parameters that ML part of GPT utilized for modeling purposes.

Consequently, the performance of future GPTs can benefit from additional sources of training data as well as additional processing power.

2 Is GPT artificial intelligence?

The popular view is that GPT is the pinnacle of artificial intelligence as of 2023.

However, in a sense, GPT is neither artificial, nor intelligent.

GPT is not artificial because it is a high-function mimic. If GPT is trained on faulty text data, GPT will act just as faulty. Unlike “weak AI” approaches, GPT has no intrinsic algorithms or knowledge for solving problems. It solves problems by mimicking textual description of problem-solving. Due to the probabilistic nature, GPT also lacks rigor in its “application of logic.”

GPT is also not intelligent, in a way. Having 175 billion parameters and 96 layers of neural networks. The ML part of GPTv3 is capable of abstraction (extracting patterns) and maintaining the correlation of patterns. GPT learns patterns of patterns, as well (due to the number of layers of neural networks). However, the lack of rigor in the application of the knowledge and patterns acquired makes GPT not exactly intelligent from the perspective of problem-solving.

3 The short term local scope impact of GPT on higher ed

If the time frame and scope is down to classes, the most obvious impact of GPT on higher ed is cheating. Work that is supposed to be completed by students can be completed by GPT. Despite the limitations of GPT, depending on how assessments are designed, GPT is quite capable of satisfactorily complete higher ed assessments.

This issue compounds on the lack of effective proctoring for online assessments. Using a KVM (keyboard video mouse) switch, a second computer, and a PIP (picture-in-picture) capable monitor, a student can quite easily forward a question to GPT and manually copy the answer.

Note that GPT is not only helpful with essay type assessments, it is also quite capable in terms of solving computational, mathematical, as well as logical problems. However, a GPT solution often contains flaws that are not obvious.

An interesting case are multiple choice questions. A student can phrase a question to GPT based on the question and the choices given. If the multiple-choice question is within the knowledge range of GPT, GPT is likely to be able to produce a correct answer.

If we maintain the current methods, standards, and meaning of assessment, GPT essentially means assessments that are not on-paper without electronics are meaningless for a wide variety of disciplines.

4 Can’t we turn a blind eye to the possibility of cheating using GPT?

The answer depends on many factors.

Philosophically speaking, if we assume that we should trust all students are honest, then the answer to the question is definitely “yes” because GPT is just a possible way of cheating.

However, many assessments are high-stake. This has little to do with how an individual instructor sets up assessment policies, but it has to do with the education goals of students. In the context of a community college, the majority of students have a goal that is career or income related. As such, the education goals themselves are intrinsically high-stake from the perspective of most students.

Even if an instructor sets up assessment repeatability and use a competency-based model for scoring, we all have a tendency to save power.