Written by: Bianca Lindau

If you’ve been pretty much anywhere on the internet lately, you’ve probably heard the buzz about AI and more particularly about ChatGPT. What exactly is ChatGPT and how does it work? Chat Generative Pre-Trained Transformer (“ChatGPT”) is a large language model (as one of its default responses has informed users endlessly) that is trained on a large body of text data to generate responses that closely mimic human language.¹ At its core, a large language model is a collection of probabilities regarding which words might appear next to others in a response to a given input.²

For instance, given the input phrase “My name’s Alex. What is your ___?” a large language model would assign a high probability to the word “name” as a likely candidate to fill in the blank; given more information about the context, additional responses such as “question” might also be reasonable candidates. You know from experience that the blank word is almost certainly not something like “is” or “quickly,” and a (well-trained) large language model would assign a low probability to those and similar options.

Makes sense, but how is it so good at responding to such a wide range of prompts? One key to ChatGPT’s ability to respond to an enormous range of inputs coherently and accurately is the size of “training data” – what researchers and engineers use to program a machine learning algorithm with the “truth” – that the model is trained on. In ChatGPT’s case, this means providing the model with around 570 GB of text-based training data, or 300 billion words, that are used to fine tune 175 billion parameters controlling the outputs that ChatGPT produces.³ The large amount of training data in essence gives ChatGPT more “experience” with language, meaning it’s better able to provide output that seems intelligent to humans.

Figure 1: Depiction of how ChatGPT was trained using Reinforcement Learning from Human Feedback and fine-tuned using Proximal Policy Optimization (“PPO”), which is is presently considered the state-of-the-art algorithm, which strikes the best balance between performance, cost, and comprehension (OpenAI 2022)⁴

Now that we have a basic understanding of what ChatGPT is and how it works, the next question to consider is, how can ChatGPT be used? The internet is filled with articles, stories and entertaining videos where people have provided the model with a range of prompts and the model provides them with stories, poems, articles, blog posts, speeches, translations, brand name ideas, etc.⁵ So how does this sudden rise in AI prevalence relate to patents and IP ownership? Is ChatGPT on track to have a plaque in the UPSTO Inventor Hall of Fame? It doesn’t look that way. In Thaler v. Vidal, the U.S. Court of Appeals for the Federal Circuit ruled that AI does not qualify as a human and therefore can’t be a named inventor on a patent:

The sole issue on appeal is whether an AI software system can be an ‘inventor’ under the Patent Act. In resolving disputes of statutory interpretation, we ‘begin with the statutory text, and end there as well if the text is unambiguous.’ […] Here, there is no ambiguity: the Patent Act requires that inventors must be natural persons; that is, human beings.

Thaler v. Vidal, 43 F.4th 1207, 1210 (Fed. Cir. 2022).

However, an interesting use case (and one that has potential to upend Google’s search dominance) is using conversational AI to retrieve information about a subject. This has broad implications for traditional search engines and could, for example, provide a much more efficient mechanism to learn about the patent process for inventors (particularly for independent or pro se inventors). However, one of the risks with large language models like these is that while the model might say things that are grammatically correct, they may not be factually accurate.

Actually, can ChatGPT even provide legal advice? This is sure to be debated more deeply in the future, but in the U.S. the answer is generally if you don’t have a license to practice law, you can’t give legal advice. What happens when I provide a prompt about the model’s ability to provide me with legal advice?

To that I say… good bot. ChatGPT is not (yet?) able to supplant legal professionals. OpenAI thankfully doesn’t seem to be claiming that ChatGPT has earned a J.D. from an accredited law school. Not that people haven’t put the model’s ability to pass law school exams to the test.⁶ Even though ChatGPT is not accredited or licensed to provide legal services, the question arises whether, as a practical matter, the model is able to provide a user with information that is both legally and factually correct. Generative AI like ChatGPT and Craiyon may be indicators of big changes to come (including to the world of legal services), but don’t expect Microsoft and Google to replace law firms any time soon.⁷

Due to the amount of interest in ChatGPT and the many legal implications it has, a few of which we have touched upon here, we will publish a few articles centered around ChatGPT. In our next article, we will discuss intellectual property issues that arise when ChatGPT is used, such as ownership issues, infringement of existing rights, and even using the model to help create intellectual property. Further, we will put ChatGPT’s capabilities to practically answer legal questions to the test in another article, by providing ChatGPT with specific prompts that, for example, a budding entrepreneur might want answers to when starting their business.

Caldwell’s Intellectual Property Practice provides advice tailored to our clients’ needs regarding intellectual property ownership and protection. We pride ourselves on our ability to keep up with technological developments and assess how these affect the current legal landscape and our client’s rights.

To learn more about current intellectual property issues when using new technologies, please contact us today.

This publication is distributed with the understanding that the author, publisher, and distributor of this publication and/or any linked publication are not rendering legal, accounting, or other professional advice or opinions on specific facts or matters and, accordingly, assume no liability whatsoever in connection with its use. Pursuant to applicable rules of professional conduct, portions of this publication may constitute Attorney Advertising.

1. Atelier Services, Generative Pre-Trained Transformers, atelier.net, https://atelier.net/ve-tech-radar/tech-radar/generative-pre-trained-transformers (last visited Mar. 7, 2023).

2. Gary Drenik, Large Language Models Will Define Artificial Intelligence, Forbes (Jan. 11, 2023, 10:00 AM), https://www.forbes.com/sites/garydrenik/2023/01/11/large-language-models-will-define-artificial-intelligence.

3. Alex Hughes, ChatGPT: Everything you need to know about Open AI’s GPT-3 tool, Science Focus (Feb. 2, 2023, 11:19 AM), https://www.sciencefocus.com/future-technology/gpt-3.

4. OpenAI, Introducing ChatGPT, OpenAI (Nov. 30, 2022), https://openai.com/blog/chatgpt; OpenAI, Proximal Policy Optimization, OpenAI (July 20, 2017), https://openai.com/research/openai-baselines-ppo.

5. Jack Appleby, Can ChatGPT build an entire brand?, Future Social (Mar. 1, 2023), https://www.marketingbrew.com/future-social/stories/2023/03/01/can-chat-gpt-build-an-entire-brand.

6. AFP, ChatGPT bot passes law school exam, cbs news (Jan. 25, 2023, 7:22 AM), https://www.cbsnews.com/news/chatgpt-bot-passes-law-school-exam.

7. Craiyon LLC, Craiyon v2, Free online AI image generator from text, Craiyon, https://www.craiyon.com (last visited Mar. 7, 2023).