The GenAI Chatbot
What the public has learned from the use of chatbots is that, with minimum information input by the user, the application produces a remarkable, authentic-sounding, and sometimes seemingly comprehensive response. Directives or questions to chatbots range from whimsical to substantive, such as, “Write a poem about the Library of Congress”; “Write a small claims complaint against xyz repair shop that failed to fix my car correctly”; “Write a congratulatory note [or speech] to celebrate my nephew’s college graduation [or marriage, or first home sale as a real estate agent, or other accomplishment]”; “Write a paragraph for a legal brief that compares plausibility pleading under Twombly and Iqbal in U.S. Federal Courts to the notice pleading standard articulated in Conley that has been adopted in most state courts. Provide citations to appropriate legal authority in Bluebook format”; “Write an opening paragraph to a judicial opinion about cats suing dogs”; “Write an essay on [choose your topic]”; “What is my illness if my symptoms are sneezing, runny nose, sore throat, and headache?”; etc. The question or request a person might submit to a chatbot is limited only by the person’s imagination. The most well-known chatbots are ChatGPT, Bing, and Bard.
How Chatbot Creators Train Their Applications to Give Coherent and Human-Sounding Responses
General-purpose chatbots such as ChatGPT, Bing, and Bard are trained on massive datasets, typically called large language models (LLMs). The training data come from almost every source imaginable, including the internet and other public and nonpublic material sources. Based on the training data, chatbots are trained to recognize language and generate a response by predicting the next most likely word or phrase as it delivers its response. As an example, if someone asked a chatbot to predict the next word in the phrase, “buckle my . . . ,” it might choose the word “shoe” based on the nursery rhyme that begins “one, two, buckle my shoe.” However, depending on the dataset on which the model was trained, there may be instances after the phrase “buckle my” in which the word “belt” occurs most often, in which case the chatbot would finish the phrase with the word “belt.”
As an example of the vast amount of data on which a chatbot is trained, the Congressional Research Service reports that the LLM on which ChatGPT was trained contained over 45 terabytes of text data “scraped” from the internet, including Wikipedia and digitized books. Further, the ChatGPT-3 models had over 175 billion parameters—variables used by the model to process and understand the training data. The parameters guide the model to provide coherent (hopefully accurate) and human-like responses. However, when I asked ChatGPT about its training data and parameters, it responded as if it were guarding proprietary information, stating:
As an AI language model, I don’t have access to specific details about the training data or the exact number of parameters used in ChatGPT-4. However, I can tell you that models like ChatGPT-4 are typically trained on large-scale datasets consisting of diverse text sources from the internet. The training process involves millions or even billions of text samples to help the model learn patterns and generate responses. The number of parameters in a model can vary, but it is typically in the range of hundreds of millions to billions.
Well said, ChatGPT—spoken like a well-trained company spokesperson.
Copyright Protection Concerns
Numerous authors, writers, journalists, and playwrights contend that the datasets of LLM models are trained on massive amounts of their writings without the consent or compensation of the content creator or rightful owner. More specifically, the claimants allege that AI developers have misused their writings by copying their works without permission to train the artificial intelligence–powered chatbots to respond with human-like responses. Several content creators have filed lawsuits seeking an injunction against future use and compensation for the unauthorized use of their works.
Among those alleging improper use of their data are the Washington Post, the New York Times, and other prominent news publications, some of which have installed a blocker that prevents their content from being collected and used to train ChatGPT. To resolve what they claim are improper uses of their data, several publication entities and news outlets are negotiating with chatbot creators to license the use of their materials. They claim the chatbots are becoming direct competitors with the newspapers by creating responses based on the original reporting of the newspapers’ journalists and writers. If these negotiations are unsuccessful, it is widely assumed that the next step will be more litigation alleging improper use of copyrighted materials. Most experts agree that judicial determinations will resolve whether AI platforms are committing some form of copyright infringement or whether the use of those materials falls within the fair use doctrine, which would not be a violation of the copyright laws.
Data Privacy Concerns
When considering the term “data privacy,” think broadly to include the concepts of client confidentiality. Not only is it possible that a chatbot may compromise private individual information revealed by an individual in a blog, on social media, or in a chatbot inquiry, but the chatbot may also acquire confidential information about an individual from input or postings by a lawyer or other professional who has a duty of confidentiality to that individual. As mentioned, LLM training datasets are created by acquiring information from numerous sources. That information might include names, addresses, phone numbers, financial and medical information, and more, which the chatbot might reveal in response to another person’s inquiry. Currently, the major chatbots are trained to respond negatively when queried about a person’s private information. However, it is only a matter of time until a bad actor, a less sophisticated chatbot, or a “glitch” causes the sharing of private or personal information. A paper produced following research by academic and industry experts reports that sensitive data and personal information from datasets can be extracted from or revealed by the datasets on which some LLMs have been trained.
Claims of improper use of personal data are not confined to the United States. For example, the Italian Data Protection Authority briefly banned OpenAI from operating in Italy in response to complaints that OpenAI unlawfully collected users’ data and failed to prevent underage users from accessing inappropriate material. OpenAI resumed its services for users in Italy after it “addressed or clarified” the issues the Italian Data Protection Authority raised. In addition, OpenAI is facing more challenges based on reports that Spain, Canada, and other countries have opened or considered opening investigations into its practices, including how it collects training data for its large language model and what information that model produces for users.
As this column was being finalized and submitted for printing, a federal lawsuit seeking class action certification was filed in the Northern District of California against OpenAI, the creators of ChatGPT. The lawsuit alleges that the AI company violated the copyrights and privacy of millions of internet users to train their AI platform when it used data scraped from the internet to train its software, including social media comments, blog posts, Wikipedia articles, and family recipes. This is another instance of litigation to watch that may provide some answers to the hotly debated issues arising from the creation and use of chatbots.
The extraordinary developments in AI capabilities have been transformative. As with other transformative inventions, issues have arisen regarding their effects on society. Because the training data on which AI chatbots are trained come from vast sources, AI developers should ensure that they are in compliance with the law regarding their acquisition of data used to train their models. Concurrently, users should be concerned not only about the accuracy of responses they receive from AI chatbots, but also about the confidentiality of the information they input to AI platforms and whether the responses they receive are subject to copyright or privacy restrictions. No less than with other transformational technological advancements throughout history, AI is also suffering growing pains. Fear not; I am confident the future will present even more challenges with each new transformational technological advancement.