UD Logo
School of Education

Learning with AI

Module 1. What Are AI Chatbots?

AI chatbots are intelligent conversational agents powered by large language models pre-trained on massive databases (such as the entire Wikipedia or the Common Crawl) that are used in generating responses by transforming what the user types. The company OpenAI calls this process a Generative Pre-trained Transformer (GPT). The first version of GPT appeared in 2018. Subsequent versions are numbered as in GPT-2, GPT-3, and GPT 3.5, which powers ChatGPT. A subsequent version called GPT-4 accepts both text and images as input. Microsoft invests heavily in ChatGPT, which Microsoft is integrating across its product lines including Bing search, Office 365, GitHub, and Visual Studio.

At the same time as OpenAI was working on ChatGPT, Google was developing an AI Chatbot called Bard. When Microsoft released ChatGPT and Google saw how it went viral by amassing more than a million users in its first week, Google felt some pressure to release its own chatbot. At first, Google's chatbot was named Bard. In 2024, Google rebranded its name from Bard to Gemini. Later in this course you will come to a module about Gemini, which is powered by Google’s Pathway Language Model (PaLM).

More AI is underway. Amazon is working on Bedrock. Elon Musk's xAI company has Grok. Zuckerberg's Meta AI uses a Large Language Model called Llama. OpenAI has a DALL-E deep learning model that can generate AI images in response to natural language descriptions, thereby providing a text-to-image model. Here are some recommendations from ZDNet about which tools work best:

To understand how these tools work from a theoretical standpoint, there are a few foundational readings:

  1. The "Transformer" Paper (The Architectural Foundation): This landmark paper introduced the Transformer architecture, the engine behind every modern LLM. It replaced older, slower methods with an "attention" mechanism that allowed AI to process massive amounts of data simultaneously. This paper, published in June 2017, remains the "Big Bang" for modern generative AI.
  2. The "InstructGPT" Paper (The Conversational Breakthrough): This research explains how models are "aligned" to follow instructions using human feedback. It is why modern AI can converse, adopt specific personas, and stay helpful rather than just outputting random text. Released in March 2022, this research detailed Reinforcement Learning from Human Feedback (RLHF), which is the method used to make AI models like ChatGPT helpful and safe.
    • Ouyang, L., et al. (2022). Training language models to follow instructions with human feedback. Proceedings of the 36th International Conference on Neural Information Processing Systems (NIPS '22). Curran Associates Inc., Red Hook, NY, USA, Article 2011, 27730–27744. https://arxiv.org/pdf/2203.02155
  3. The "Stochastic Parrots" Paper (The Foundational Critique): This paper warns against the dangers of building ever-larger models. It highlights how LLMs lack true understanding, instead "parroting" patterns found in their massive training data, which can lead to the reproduction of social biases and significant environmental costs. Published in March 2021, this paper is crucial for understanding the ethical risks and environmental impacts of scaling up AI.
    • Bender, E. M., Gebru, T., McMillan-Major, A., & Mitchell, M. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT '21). https://dl.acm.org/doi/10.1145/3442188.3445922

While these three papers are still the foundational classics, the field has shifted toward newer research that addresses their limitations. In 2026, researchers are moving beyond basic "Transformers" and "Stochastic Parrots" toward more efficient and "reasoning" models. Here are the modern successors to those three categories:

  1. Beyond the Transformer: Sub-Quadratic Architectures. The original Transformer is being challenged because it is computationally "heavy" (quadratic). New papers focus on architectures that can process massive contexts (entire books or codebases) much faster. The following successor paper introduced a "State Space Model" (SSM) that achieves better-than-Transformer performance while being much faster. In 2026, many experts believe Transformers will be largely replaced by these sub-quadratic models.
  2. Beyond InstructGPT: Direct Preference Optimization (DPO). InstructGPT used a complex process called RLHF (Reinforcement Learning from Human Feedback). Newer research has made this much simpler and more stable. The following paper showed that you can "align" a model to be helpful without the massive engineering overhead of the original InstructGPT method. Most modern open-source models (like Llama 3 or DeepSeek) now use DPO or its variations.
  3. Beyond Stochastic Parrots: The "Reasoning" Revolution. The Stochastic Parrots critique argued that AI just predicts the next word without "thinking." Recent breakthroughs in 2025 and 2026 focus on "Chain-of-Thought" and "Inference-time Scaling," where the model actually "thinks" before it speaks. These papers move the conversation from "parroting" to "systematic reasoning". By allowing models to generate internal "thoughts" or reasoning chains, LLMs can theoretically solve complex math and logic problems that the original "parrots" never could.

The readings listed above are admittedly advanced. For a less technical introduction to AI, view the following series of video tutorials from code.org: