ChatGPT Upgraded to GPT-4 is Here. Can You Trust It for Important Business Tasks?

By now, you have certainly heard of ChatGPT. It’s the generative artificial intelligence (AI) that can be used to produce amazing content, such as a term paper about a book.

OpenAI, which offers ChatGPT, recently made a more powerful version available to those who pay for a subscription, which costs $20 a month. The free version of ChatGPT uses the GPT-3.5 language model, whereas the premium version makes the more powerful GPT-4 available.

Microsoft says its new version of its Bing search engine is already powered by GPT-4. Unfortunately, that upgrade is in beta testing, so it’s not generally available.

Will upgrading to GPT-4 make ChatGPT ready for business prime time? Here is the critical question many businesses are puzzling over: How do I judge when ChatGPT or its AI competitors reach the stage when it’s cost-effective and safe to use in my business?

Here’s how to decide: Generative AI must do five things to be cost-effective for many businesses, and perhaps a sixth thing. Different kinds of businesses have different needs and sensitivities, so some needs may not apply to your industry.

1. Avoid Factual Errors. The AI should be sufficiently trustworthy to not state factually wrong things. The tech community generously calls these failures “hallucinations.” Ideally, the AI would cite its sources for factual claims so you can check them.

2. Avoid Important Omissions. The AI should not omit important information in an answer.

3. Protect Confidentiality and Privacy. You need to be able to put confidential or private information in the prompt without fear of it being disclosed to others or retained by the AI system. In some circumstances, it may be okay if the AI retains the information, provided it maintains confidentiality and is technologically secure.

4. Avoid Infringement. Your use of the output of the AI should not infringe upon or violate the rights of others, especially the copyrights to the material studied by the AI in its dataset. Usually, that means the output shouldn’t be entirely or even substantially a single piece of work by somebody else.

5. Be Current. Your prompt may require the AI to have knowledge running up to the present. Is the AI limited by a dataset ending sometime in the past?

6. Avoid Bias. For some uses, you may want the AI to produce an output that is not biased for or against a particular group (e.g., people of a certain race), such as when using an AI to filter applicants’ resumes in a hiring or admissions decision. This factor gets into politics, namely the debate over whether people should be chosen on individual merit or, instead, whether individual merit must give way to enable proportional representation by race. Thus, whether an output is biased depends upon your viewpoint on what is improper bias.

Where does ChatGPT using GPT-4 stand on those needed things? Overall, it still has significant limitations.

OpenAI says GPT-4 will have a lower hallucination rate, but they will still happen.

By default, ChatGPT doesn’t cite its sources. You can ask it to cite its sources, but the sourcing may be bad. Sometimes ChatGPT cites to a dead web link. In the legal realm, sometimes it cites a case to the wrong court.

Regarding avoiding important omissions, that’s hard to control due to the nature of AI. In theory, an AI can do well when the answer comes from a defined database entirely available to the AI, although there is no guarantee. For example, if you are hunting for case law on a particular issue, the AI might be unlikely to miss an essential case if it has the entire corpus of case law in its database. The AI is likelier to omit important information when the realm of relevant information is not limited to a defined database.

Concerning confidentiality, it appears GPT-4 continues the practice from GPT-3.5 of using inputs from other ChatGPT users to make it better, so there is still a hypothetical risk that your input could show up in someone else’s output.

As for the risk of getting an output that infringes upon someone else’s copyright, OpenAI says ChatGPT is engineered to try to avoid making a single piece of data it studied (such as someone’s blog post on the Internet) a substantial part or the entirety of an output. Still, if your prompt is super narrow, you can’t be sure that your output doesn’t largely capture someone else’s copyright property.

Regarding being current, it appears ChatGPT still does not know about events after September 2021.

We will soon get to the point where working in partnership with an AI will be standard for almost every knowledge worker. Be prepared!

A longer version of this column (more information) is available on John Farmer’s Substack. You can view and subscribe to that Substack here.

Written on March 22, 2023

by John B. Farmer

© 2023 Leading-Edge Law Group, PLC. All rights reserved.