There is much being said about new chatbots coming out of the big technical powerhouses. Whether we want them or not, they are being integrated everywhere despite not being fully tested. Anyone that uses a search engine in a legal fashion should pay attention. As the Internet is one, large, unorganized library, this might be a different way to bring some order to chaos. Then again, maybe it is an additional factor of chaos.
At its core, these Generative AIs are an advancement in searching and summarizing large amounts of information. They are far from a final product which should serve as a cautionary flag to everyone. To be worthwhile of use, you will need to apply human intuition and wisdom. An algorithm will not be able to do that for you.
What is generated by the chatbot does seem intelligent and reasonable. The responses do seem to summarize information well and are mostly coherent. Despite being a nice advancement in delivery, are you able to gain any real benefit or will it be a costly path?
No one is an expert at this point. Even the developers and designers that are deep in the underpinnings of the product are just scratching the surface. How can anyone claim that a product can be a master of something that mimics intellectual functionality that took millennia to create and cultivate?
Review Your Goal
One of the first questions anyone must ask is “why.” Why does this exist? Why even do you need to use a chatbot over a search engine? Before using any of them, it is important to define your goal. There are many things that can be acquired without the aid of an advanced chatbot. A simple web engine along with the human ability to link together information can go far on its own. The information could very well be more detailed and accurate than the summarization that is constructed by a generative AI.
Even more, question yourself during this phase. Are you using it for ethical reasons? Ex: cheating on assignments (and the countermeasures1]). If so, maybe you need to consider you are cheating yourself out of growth by being lazy (or far worse, something illegal).
Understand the Tool
If I went into a master carpenter’s shop, there would be a myriad of tools placed in an organized fashion. If I were to then pick up a random tool and ask him to describe it to me, I am certain volumes could be filled with his experience, knowledge, and the resulting wisdom of his works from using the tool.
The same mindset must be applied to anyone who uses these new chatbots. A user might not be able to build the tool itself but should understand what goes into making it work. There has been much effort that went into these bots. Everyone would gain better appreciation by reading into the background of what you are going to use for “everything.”
Just like you do not need to be a highly trained mechanic to operate a motor vehicle, it is not essential for you to have a doctorate in artificial intelligence to use these tools. You should still read and come to know the general concepts. Having respect for what goes into these tools will benefit you in your use of what it does. This might even lead you to dive deeper into the technology. Who knows what can happen. You might become so good and passionate that you develop the next level of AI. Never close a door on what is possible for you unless it gives you pain.
By gaining a deeper understanding into these tools, everyone will also know their limitations. Even the worst “it does everything” tool rides high on the hype curve at release only to be brought down to reality as time takes hold. Come to know what you are using, and you will find its proper niche.
Cost and Impact
There is always a cost to using a tool, let alone the Internet. On the surface, it is true that some chatbots have a subscription cost. Do you know the other costs, such as energy?
It takes a large amount of energy and resources to put these tools in the state they are now. Think of the large data used to train the models. Gathering and cleansing them requires a very large effort. Resources to do this are limited at best. Is all the focus and push into an unproven technology worth losing progress on other advancements?
There are also the intense resources that go into training the models themselves. Even in a simple, small set analytics project, training models is a rather costly process. The process of building these Large Language Models has been described as causing “melting GPUs.”2 Perhaps an extreme, but this should alert anyone concerned about energy efficiency.
I do not doubt that efficiency advancements are considered important by the development groups. We all need to be aware that there is a cost associated to it. That can have downstream consequences. We cannot claim to be working toward a better future for the next generation if we use up all the current resources. Be aware of this as you progress in your experience.
Legal
Legal impact is where the use of AI meets the reality of humans. Is everything going into and out of the tool 100% legal? For example, who owns the source of the training data? If the generated content is copyrighted, are you protected if the owner comes at you?
Even if the sources are legit, who owns anything produced by the chatbot? Will the tech company behind the Generative AI claim they created it and not you?
In my work, some employers have taken the stance that Generative AIs are not to be used. They want to avoid these types of legal issues. No matter how good the algorithm, the risk of litigation is too great to justify its use.
Who Supports You?
There are already many cases where software is developed using Generative AI in a very quick fashion3. This does have a high neat-factor but is it wise?
Thinking of generated logic, processes, and code, if something goes wrong with produced code, who corrects it? Did you remove your developers and replace them with a chatbot only to find out you still need developers? How can bugs be corrected on top of all the other possible potential issues discussed? For the mere fact that you will need to speak to all the nuances of a product, I believe you are still going to need to know how it is built.
Question the Generated Output
Given information of any type, prove it is true4. How do you know the information you lifted from an AI is correct? The technology is still very new so make sure you do cross-validation before running into battle with that information. There are examples already of improper use of the technology, such as failing students based off of incorrect output5. Make sure you keep your mind in-line with reality to know this technology is not 100%.
This can be done easily in some cases, such as coding. Copying the code into an IDE, like one does with Stack Overflow, is straightforward to do. That is acceptable provided you are doing so in a contained, development environment. You do have compartmentalized, development environments, right?
Know the sources!
Keeping on with validation, always check the underlying sources for the generated content. Proper and accurate sources are critical to the quality. This should be always at the forefront of a user’s mind. Question where the answers are pulled from. The responses seem very convincing as they have been designed to do so.
The content that is created can sometimes have citations. This is welcome but does not always get produced with the result. I am hoping that improvements are made where all output includes at least basic references.
Bias
When reading through the generated responses, how do you know everything is truthful and impartial? If you ran the same question amongst all chatbots, how identical are they to each other? If they are not the same, which one is right? Are any of them 100% in line with each other even once?
Take this for example from Google Bard6. I ask “show the bias that happens in large language models.” Here is the response.
===
Large language models are trained on massive datasets of text and code. This data can be biased, and as a result, the language models can also be biased. Some examples of bias in large language models include:
- Gender bias: Language models can be biased against certain genders. For example, they may be more likely to associate certain professions with men than with women.
- Race bias: Language models can be biased against certain races. For example, they may be more likely to associate certain negative stereotypes with certain races.
- Political bias: Language models can be biased towards certain political views. For example, they may be more likely to generate text that is supportive of a particular political party or candidate.
- Cultural bias: Language models can be biased towards certain cultures. For example, they may be more likely to generate text that is reflective of the culture in which they were trained.
It is important to be aware of the potential for bias in large language models. This bias can have a negative impact on the way that these models are used. For example, a language model that is biased against women may be less likely to generate text that is supportive of women’s rights.
There are a number of things that can be done to mitigate the bias in large language models. One approach is to train the models on more diverse datasets. This can help to ensure that the models are exposed to a wider range of perspectives and experiences. Another approach is to use techniques such as debiasing algorithms to remove bias from the models.
It is important to remember that large language models are still under development. As we learn more about bias in these models, we can develop better ways to mitigate it.
===
It seems nice that the response recognizes bias as a problem. Does that mean even this response is biased? What can be said about the mass adoption taking place given the origination has known bias? This should raise some concerns in your mind.
Conclusion – Refine Yourself
I might have sounded overly critical in this article. I do believe there are great advances taking place with these tools. We must keep in mind that it is not a replacement for the human brain let alone proper research methods. There should also be encouragement to support and improve these efforts.
Everyone needs to know that turning data into information then into wisdom takes time. Do not just buy whatever gets presented no matter how much money is behind it. Even if takes a long time, always strive to understand the answer to the question you are asking. Remember:
Quid est veritas?7
References
[1] Universities express doubt over tool to detect AI-powered plagiarism. Bethan Staton. Financial Times. April 3, 2023. https://www.ft.com/content/d872d65d-dfd0-40b3-8db9-a17fea20c60c
[2] ChatGPT and generative AI are booming, but the costs can be extraordinary. Kif Leswing and Jonathan Vanian. CNBC. Retrieved from: https://www.cnbc.com/2023/03/13/chatgpt-and-generative-ai-are-booming-but-at-a-very-expensive-price.html
[3] Watch a Blizzard game director code a game using nothing but ChatGPT. Sean Endicott. April 13, 2023. Retrieved from: https://www.windowscentral.com/software-apps/watch-a-blizzard-game-director-code-a-game-using-nothing-but-chatgpt
[4] If it’s on the Internet, it must be true. Retrieved from: https://www.webdevelopersnotes.com/if-its-on-the-internet-it-must-be-true
[5] Professor Flunks All His Students After ChatGPT Falsely Claims It Wrote Their Papers. Miles Klee. Rolling Stone. May 17, 2023. Retrieved from: https://www.rollingstone.com/culture/culture-features/texas-am-chatgpt-ai-professor-flunks-students-false-claims-1234736601/
[6] Google Bard. Google. Retrieved from: https://bard.google.com
[7] John 18:38. Wikipedia. Retrieved from: https://en.wikipedia.org/wiki/John_18:38
Andy’s Neural Works 