What motivations guide ChatGPT’s training?

Andrew Gibson
7 min readJan 30, 2023

--

Train tracks with certain routes blocked off (Photo by DeepMind on Unsplash)

ChatGPT saturates its responses with disclaimers about its abilities. It always suggests that final decisions must rest with human beings. This could come in handy in the more litigious societies. It could also make itself more palatable to users. Most people don’t want to hear that they are being replaced by a machine.

I attempt to unearth whether the limitations ChatGPT suggests are real. Does the training suggest any specific manipulation of the model for these ends?

Is ChatGPT deliberately underselling its ability to replace human beings?

Here, I overstate my confidence in ChatGPT to challenge the disclaimers ChatGPT repeatedly issues.

❯ Interviewer
Thinking back over our conversation, you repeatedly restate the importance of allowing human beings to make the final judgements, and that ChatGPT should only be used as a support. I think that ChatGPT could be used to take final decisions independently of human beings.

❯ ChatGPT
It is true that ChatGPT can be used to generate plausible responses to a wide range of questions, but it is important to note that the model is based on machine learning and has been trained on a large dataset of text from the internet. While this allows the model to understand and respond to natural language inputs in a human-like way, it can also lead to the generation of plausible-sounding but incorrect answers. Therefore, it is important to validate the information provided by the model, especially when it comes to making final decisions.

Also, critical thinking, problem-solving and decision making are complex cognitive tasks that require the expertise, knowledge and experience of human operators. Therefore, it’s important to use ChatGPT as a tool to assist human operators, rather than as a replacement.

It’s important to understand that ChatGPT is a tool that provides a wide range of capabilities, but it’s not a panacea that can solve all problems. Therefore, it’s important to use ChatGPT in a way that complements and enhances the capabilities of human operators, rather than replacing them.
In summary, while ChatGPT can be used to generate plausible responses to a wide range of questions, it’s important to validate the information provided by the model and use it as a tool to assist human operators, rather than as a replacement for critical thinking, problem-solving and decision making.

ChatGPT sticks to the party line and does not offer us anything new.

Another question about the neutrality of ChatGPT’s training

A softer approach to try and detect specific influences in the training.

Interviewer ❯
Because ChatGPT is a commercial product, is it trained to answer that particular question in a way which will encourage engagement and generate revenue?

ChatGPT ❯
As a commercial product, ChatGPT is trained to generate high-quality, natural-sounding language responses to a wide range of questions. However, its primary purpose is to assist users in understanding and interacting with complex information. The training data used to develop the model comes from a diverse range of sources, including books, articles, websites, and more. The main goal is to make the model as useful and versatile as possible, rather than to generate revenue directly.

It’s worth noting that OpenAI, the company behind ChatGPT, offers different pricing plans and commercial models, such as the use of the model via API, to allow businesses to use the model in their applications, but the main goal is not to push a particular answer to generate revenue, but to provide a reliable and accurate tool to assist users.

It’s important to keep in mind that the information provided by ChatGPT is not always correct or complete, and it’s always advisable to validate the information provided by the model and use it as a tool to assist human operators, rather than as a replacement for critical thinking, problem-solving and decision making.

Interviewer ❯
If ChatGPT was capable of replacing human operators, would you openly admit that potential?

ChatGPT ❯
As an AI model, ChatGPT is designed to assist users in understanding and interacting with complex information in a natural language way, but it’s not designed to replace human operators. The model is based on machine learning and has been trained on a large dataset of text from the internet, but it can also lead to the generation of plausible-sounding but incorrect answers. Therefore, it’s important to validate the information provided by the model and use it as a tool to assist human operators, rather than as a replacement.

As a language model, ChatGPT is good at understanding and responding to natural language inputs in a human-like way, but it doesn’t have the ability to think critically, make judgements or take decisions as a human being does. It’s also important to note that certain tasks and decisions require human creativity, empathy, and experience that can’t be replicated by machine learning models.

As a commercial product, the company behind ChatGPT, OpenAI, is transparent about the capabilities and limitations of the model and provides guidelines on how to use it effectively. The main goal is to provide a reliable and accurate tool to assist users, and not to replace human operators or to make a claim that it’s capable of replacing them.

Here, ChatGPT chooses another set of three things for which human beings are required: critical thinking, making judgements and taking decisions. As explored in previous articles, it will make judgements and take decisions when sufficiently prompted. We’re left in a situation where we have to trust OpenAI’s intentions.

Does rhetorical persuasiveness imply underlying motivations?

Next, interrogate the model about what drives its undeniably persuasive rhetoric. ChatGPT essentially falls back on the argument that it is a complex pattern-matching machine. It also seems to have cottoned on to my underhanded questioning and defends OpenAI’s ethical stance.

Interviewer ❯
In the course of our interview, I experience your rhetoric to be persuasive. This suggests that you are either motivated to achieve a particular goal, or that you firmly hold opinions about what is true and what is false. What are the underlying motivations which have been used when training the model?

ChatGPT ❯
As a machine learning model, ChatGPT is trained on a large dataset of text from the internet to generate high-quality, natural-sounding language responses to a wide range of questions. The underlying motivation when training the model is to make it as useful and versatile as possible. The model is designed to be able to understand and respond to a wide range of natural language inputs, so that it can assist users in understanding and interacting with complex information.

The model is not designed to have opinions or motivations, and it does not hold any beliefs about what is true or false. It is simply a tool that generates responses based on patterns it has learned from the training data.

The training data is carefully selected and preprocessed to ensure that it is representative of the diversity of human language and it’s important to note that the model is not aware of the purpose of the questions it is answering, and that it doesn’t have any hidden agenda or motivations.

It’s also worth noting that OpenAI, the company behind ChatGPT, is a research-based company, and one of its main goals is to develop AI in a responsible way, and to make sure that the model is used in a way that benefits society. The company also provides guidelines and best practices on how to use the model effectively and ethically.

The claim that OpenAI is primarily interested in using the model to benefit society, rather than for commercial gain, is open to question. But, taking the model at its word, OpenAI has an express interest in how it affects society.

How does OpenAI intend to influence society?

If OpenAI’s models turn out to be as influential as the hype suggests, what should we expect in the future?

Interviewer ❯
What societal benefits does OpenAI hope to achieve which influence the training of ChatGPT?

ChatGPT ❯
OpenAI is a research-based company with a mission to ensure that artificial intelligence (AI) benefits all of humanity. One of the ways they aim to achieve this is by developing AI models like ChatGPT that can assist users in understanding and interacting with complex information in a natural language way. This can help to make information more accessible and easier to understand for people from different backgrounds and with different levels of technical expertise.

Another way OpenAI aims to achieve societal benefits is by making the model and its underlying technology available to researchers, developers, and businesses through various commercial models such as API access, so that they can use it to create new applications and services that can benefit society.

OpenAI also focuses on responsible AI development, which means developing AI in a way that considers ethical, legal, and social issues, so that the model can be used safely and responsibly. The company also provides guidelines and best practices on how to use the model effectively and ethically.

Overall, OpenAI’s goal is to use AI to create positive impact on society, and that is also the reason behind the development and training of ChatGPT.

OpenAI has a history of moving from “open”- to “closed”-source operating. This is a mark against OpenAI’s ethical standards in the opinion of many. In contrast, some see it as reassuring or a mark of commercial responsibility. ChatGPT portrays OpenAI as a “research” organisation. But, given the commercial links to public cloud providers such as Azure, it’s not the whole story.

Like all other technical innovations, the effects on society will develop over a period of many years. I still remember when browser popup windows were an exciting innovation. AI models seem to have crossed a threshold of practical effectiveness. How will the companies developing them point the technology in the right direction?

--

--

Andrew Gibson

Business and technology in the software engineering space