Responding to Errors and Mistakes in ChatGPT and Other Generative AI Systems

Since Open AI released ChatGPT for public use on November 30, 2022, its rate of adoption by the public has been explosive and appears to be increasing at a near exponential rate.  Within two months after its public release, it had acquired an estimated 100 million users and surpassed one billion page visits by the end of February, 2023 — an increase unprecedented and creating strong incentives for other generative AI systems to join the fray.

In its official announcement of the release to the public of ChatGPT, Open AI stated very clearly that the system would sometimes generate errors of fact and other types of mistakes in a random manner difficult to predict. It also warned potential users of the urgency of checking ChatGPT responses to user queries against other sources for verification. Most other Generative AI systems have issued similar warnings.

The cumulative impact of the public release of ChatGPT, followed by the public release of other generative AI chat systems, is yet to be determined. However, it will be disruptive— at least in the near term — and may require significant changes to the workplace. Some of these changes might very well involve introducing what I would call error management systems to help manage the deluge of items generated by ChatGPT and other systems in a more organized manner.

As the usage rate of queries processed by ChatGPT and other generative AI chat systems increase, the corresponding number of potential errors and mistakes associated with those queries will probably also increase, even as Open AI and other generative AI systems continue to struggle to reduce the rate of such errors and mistakes. Eventually, organizations using generative AI systems will need to consider introducing more formalized error management systems, as mentioned earlier, to help mitigate the possibility of undiscovered errors and mistakes, which could cause undue harm and possible embarrassment to their organizations.

When developing error management systems or other formalized approaches to coping with these issues, we need to consider the potential role of editors, proofreaders, and others experienced with fact checking.  Professional librarians should also be considered.  As information managers they are charged with a variety of related  responsibilities, including the use of technology, to facilitate detailed checking for errors and inconsistencies while cataloguing materials, indexing, and other similar tasks involved in ensuring the accuracy and integrity of the collections they are responsible for maintaining. The involvement of professional librarians is not without precedent.

In her book. Information Hunters, Kathy Peiss describes how at the beginning of World War II, professional librarians were enlisted by the U.S. Government to help organize and bring added coherence to intelligence gathering at that time. Much of the rationale for selecting professional librarians, as described by Peiss, is similar to the qualities mentioned earlier in this posting and should be quite helpful in supporting the type of error management being discussed.

For some additional sources of information regarding errors and mistakes in ChatGPT and other generative AI chat systems, see:    AI deep fakes, mistakes, and biases may be unavoidable, but controllable | Computerworld,   The 5 biggest risks of generative AI, according to an expert | ZDNET,  A Brief History of The Generative Pre-trained Transformer (GPT) Language Models – WordBot,  10 ChatGPT mistakes you’re making and how to fix them | Descript,  That’s Not Right: How to Tell ChatGPT When It’s Wrong | PCMag , How Accurate Is ChatGPT In Providing Information Or Answers? | Botpress Blog. 

Since much of this posting focuses on Open A! and ChatGPT, I visited the ChatGPT site and entered the following query: How can ChatGPT prevent itself from manufacturing false, erroneous information? After an approximate five-minute wait, I received the following response, which I am reproducing here in its unedited form:

How can ChatGPT prevent itself from manufacturing false, erroneous information?

ChatGPT

As an AI language model, ChatGPT is designed to generate responses based on patterns and examples it has learned from a vast amount of training data. While efforts have been made to ensure the model’s accuracy and reliability, it is still possible for it to occasionally produce false or erroneous information. Here are some ways in which ChatGPT and similar AI systems aim to prevent the generation of false information:

1. Data preprocessing: Before training the model, the training data goes through a rigorous preprocessing phase where erroneous or false information is removed or corrected to the best extent possible. However, this process may not catch all errors, and some incorrect information could still make its way into the training data.
2. Training on diverse data: To minimize biases and errors, AI models like ChatGPT are trained on diverse datasets from various sources. This helps expose the model to a wide range of perspectives and reduces the likelihood of consistently generating false information.
3. Quality control during training: Human reviewers work closely with the AI research teams to review and rate potential model outputs. This feedback loop helps the model learn over time and improves its performance. Guidelines provided to reviewers explicitly state not to favor false or unsupported statements.
4. Ongoing evaluation and iteration: AI models are regularly evaluated on their performance and iteratively improved based on user feedback and additional training data. OpenAI, the organization behind ChatGPT, actively seeks feedback from users to identify and rectify errors, biases, and other issues.
5. Disclosure of limitations: It is important to acknowledge that AI models like ChatGPT have limitations. OpenAI explicitly communicates that the model may not always provide accurate or up-to-date information, and it should not be relied upon as a primary source of information.
6. Encouraging user skepticism: OpenAI encourages users to critically evaluate and verify the information provided by ChatGPT and not blindly accept it as fact. Promoting media literacy and critical thinking can help users make informed judgments about the information generated by AI systems.

It’s worth noting that while these measures can help mitigate the risk of false information, they cannot completely eliminate it. Users should exercise caution, verify information from multiple reliable sources, and remain aware of the limitations of AI systems.