Artificial Intelligence (AI) has popped out of the box in the last couple of years from being a relatively esoteric field of computer science research to a tool that many of us are using, often unwittingly. It scares some of us spitless as we contemplate some of the possible paths in the future. But, it is becoming evident that we are living with AI, even in our discussions on Sabbath School Net, and it is time to talk about it
Artificial Intelligence is a rapidly expanding field and raises many issues about its future use, ethical boundaries, and our perceptions of ourselves. It would take far too long to discuss all these important issues, so I have chosen to be practical and talk about the use of the Large Language Model (LLM) most ordinary people encounter now.
For most of us, our initial interaction with AI involves LLM chatbots such as Gemini and ChatGPT, or versions of them embedded in search engines such as Google or applications (Copilot in Word). We are often surprised at the apparent depth of knowledge they traverse and the eloquence they use in expressing their responses. I know students who have asked an LLM chatbot to write an essay on a typical science or history topic and have received a response they could hand in. For educators, that is scary and raises several questions about how we best tackle the issue in education
The Large Language Model has two main components:
- A huge knowledge base. Essentially, that means most of the publicly available information on the Internet. This is not just the data itself, but relationships and data patterns.
- A statistical writing engine. It really understands language and knows from its statistical analysis the most likely word patterns used to express information and ideas. This is not just sentence structures, but can include paragraphs and even pages of writing
There are two aspects we need to consider in the Sabbath School Net discussion forum:
- How LLMs source their data, and
- How they report the results.
Gathering Information
An LLM is essentially a statistically based search engine. Given the task of finding information on a particular topic, it finds as much information on the topic, given the constraints of time (we are impatient and don’t want to wait forever). It assigns a weight to each piece of information and constructs its answer based on that weighting. For straightforward information that is generally not an issue. For example:
- Write a timeline of the major events of the French Revolution.
- Summarise the contribution of Earnest Rutherford’s work to our understanding of the atom.
Typically, where there is a body of generally accepted data, you get a reasonably good summary.
However, where concepts are less clear or where there are biased views depending on religion, politics and so on, there are arguments about the validity of the data. Then the results can be poor or even misleading. For example, try asking questions about “climate change.”
LLMs are known to be lazy and fill in gaps by making things up based on probability. These are called hallucinations. If you have a feel for the topic under discussion, you can probably pick those and ask for it to be corrected. In the early days, they would often provide another answer that was just as silly. (This is not LLM, but it illustrates the problem. I sometimes use generative AI to create images. I asked Firefly, Adobe’s AI generative engine for Photoshop, to create an image of an Airbus A380 landing in front of a beautiful sunset. It got the sunset right, but the plane had 2 engines, and everyone knows that the 380 has 4 engines. I pointed out the error and explained it had 4 engines, and asked it to recreate it. Firefly put two more engines on top of the wings!)
Cognitive bias is also a problem with LLMS. They are trained on data, and if the training data is biased, then the statistics are weighted towards that bias. This can result in stereotyping of gender roles and racial characteristics. The fact that most LLMs are trained on English text means that they are biased towards Western culture. Frequency of occurrence is also an issue. If most of the data on a particular topic presents a particular view, other views are often ignored, even though they may be quite valid. An LLM is a probability machine and does not test validity.
There is also the human-machine interaction cognitive problem. We tend to view machines as objective and accept their responses on that basis. This is exacerbated when a machine comes up with an answer that supports our own ideas.
LLMs also remember their interaction with users and use that to help shape their answers. In other words, it tries to agree with you.
Bible Study Example
The question is: Can we use an LLM such as Chatbot or Gemini in our Bible Study? To help you understand their use and some of the issues, I will illustrate with an example related to our recent lesson study. I will give you the conclusion first: LLMs can be useful, provided we use them intelligently and use a strategy that allows us to be reasonably confident of their answers.
When we were discussing “type and “antitype” in the lesson a few weeks ago, I asked Gemini for a summary of the idea and received a very Seventh-day Adventist answer. I sensed this was a cognitive bias. Either most of the information it had used as source material was of Adventist origin and/or in its previous interactions with me it had worked out I was a Seventh-day Adventist and would like an Adventist answer. Rather than just accepting the answer given, I started a conversation asking for more information and giving different perspectives. I was aware of the large body of Adventist writing on the topic and thought it would bias the response.
If you want a better understanding of the topic you need to have an intelligent conversation with the LLM (you supply the intelligence). You should not be satisfied with one question and one answer. (That is being dumb!) You need to have a questioning strategy.
Here are some of the questions I asked Gemini on the type/antitype topic:
- List the sources used. Most LLMs are good at this, although some of the earlier models lied through their back teeth and made sources up.
- What other churches use type/antitype idea?
- How do they use the idea?
- Why is the idea so important to Seventh-day Adventists?
- What are the criticisms of type/antitype idea?
- What words are used in the original biblical languages for “type” and “antitype”?
- Where and how are “typos” and “antitypos” used in the Bible?
- How do non-Seventh-day Adventist churches use the type/anttype metaphor?
- Provide a comparison between the non-Adventist and Adventist understanding of type and antitype.
The list is not exhaustive, but I hope you can see that by the end of the discussion, we should have a better picture of the type/antitype idea, other than just a catechistic Seventh-day Adventist response. You can try this for yourselves.
Writing with AI.
Using an AL LLM to write something for you is a huge temptation. They typically get the grammar and syntax right, and they use expressive language. And if you are the sort of person who struggles with writing, it appears to be the ultimate solution. If we have used the LLM to gather the information using a good strategy to avoid bias, artefacts, and hallucinations, it seems only a small step to get it to write it up for you as well.
But the big issue with using a LLM to write something is its methodology of using weighted probablity to assemble the sentences for you. In other words, an idea expressed using an LLM is the result of a weighted average. And, you are not “average”!
I live near Lake Macquarie, and it has an average depth of 8 metres, but, as any of the local boaties will tell you, there is an awful lot of the lake that is less than 1 metre deep. So, if you use the average depth as a guide for your boat navigation, you are sometimes going to find yourself going aground.
LLMs write a good average. Usually, its sentence structure and syntax are spot on. But, none of us is average. We are passionate about some things and indifferent about others, often in the same paragraph. We have idiosyncrasies of expression that add character to our writing and speech. LLM writing may be good, but it is not us. It is an average person without a name. It overwrites your personality.
LLM writing is characterised by using popular words and phrases. It tends to use the same sentence patterns and lengths. And it pads out ideas more descriptively than necessary. It is usually bland and somewhat emotionless.
If you ask it to write on a spiritual topic, apart from the source issues discussed earlier, you tend to get a catechistic answer rather than an idea that challenges your thinking.
Should we use Artificial Intelligence?
The short answer is that we should never use it dumbly. Use it intelligently. The real danger is that we let AI take away our own intelligence and critical thinking, and we substitute blind acceptance of the answers it provides. Interestingly, when I was teaching Information Technology more than a dozen years ago, I realised that search engines such as Google were good at providing answers, but you had to be aware that they had a commercial bias. You were getting Google answers free, but in exchange, those answers were biased, pushing you towards products promoted by sponsors. You had to be smart and know how to filter the answers so that they were less biased.
That has not changed. Artificial Intelligence entities have the potential to be biased by populism, social attitudes, consumerism, and so on. They are based on frequency without ethical considerations or moral filtration, regardless of how authoritatively they express it.
Having said that, AI can provide useful information when you use your “garbage detection” tools to filter the results.
Secondly, if you want to sound authentic when you write, avoid the use of LLMs for writing. This is particularly important in conversation contexts such as online forums (eg Sabbath School Net). We are having discussions with real people and not just exchanging ideas generated by averaging machines. (I should add that exchanging quotes mindlessly is also not having a useful discussion.) At a personal level, I would rather deal with poor syntax and expression knowing that it was “you” on the other end of the conversation, rather than a statistical engine generating perfect sentences and paragraphs.
Do not fall into the trap of thinking that an LLM can speak for you. They might write better language than you, but they are not you. We are much more than a statistical machine.
Personal Note
Do I use Artificial Intelligence? Yes, I do, but, I have a very long history of dealing with computers and AI. I studied Natural Language Processing in the early 1990s, before the Internet became generally available. I used some of the early search engines such as Lycos, Infoseek, Yahoo, and AltaVista. I have learned the importance of using my intelligence and “garbage detection” skills by asking multiple questions to remove bias as much as possible. Modern LLMs are an extension of these search engines, but the addition of AI has not removed the importance of garbage detection.
I find it useful in Bible Study, particularly when I want to find the extent of an idea. Earlier, I used the type/antitype idea as an example. That wasn’t a theoretical example. I used it when commenting on the idea when it came up in the lesson study. I quickly found the words used for type and antitype in the original languages, and it provided a framework for seeing how the words were used. I also discovered how other Christians viewed the type/antitype idea. It gathered essential facts and relationships for me that would have taken a long time to obtain using conventional means. I had a conversation that asked questions from a variety of viewpoints to ensure that I was not getting an “average” answer
I do not use AI for writing per se. I admit that on the odd occasion when a particular expression escapes me, I have asked for a rewrite. It usually reveals some grammatical errors, which I can correct, and I will make alterations based on that. But that is the extent of my use of AI for writing. I want readers to know that it is Maurice Ashton is writing.
Conclusion
If you can use AI without compromising your own intelligence and identity, by all means use it to gather information. You must understand the importance of a multi-pronged approach that identifies hallucinations and cognitive bias. Don’t just accept answers from an AI entity as an oracle.
If you value your own personality, don’t use AI for writing, especially in the conversational context of a discussion forum. An AI entity may write well, but you are not average. You are real, passionate, and have personality. Never let AI take that away from you. If you let AI think or write for you, you are dumber than you think!
(Disclaimer: In this article, I used AI to find the average depth of Lake Macquarie! And I checked the answer from several local sources.)
