Beyond the Chat: Unpacking ChatGPT's Terms and Conditions

You know the drill: "Don't trust AI," they say. "It's grabbing all your data. Nothing's safe." I've heard it too. And whilst I have wondered about the veracity of people's claims, I've never looked into it myself to confirm - up until now. This article provides my summary of the terms and conditions for using AI, specifically ChatGPT, as an individual in Australia, covering both free and paid (Plus and Pro) accounts.¹ By reading on, this article will offer you insights into the general provisions, as well as a more focused examination of key areas of interest.

It is important to note that this article is not intended to be an exhaustive documentation of each and every term and condition. For a comprehensive understanding, I strongly encourage you to consult the terms and conditions directly. Now let's begin.

Large Language Models - a tech refresh

Let me preface this by saying that I am not a super tech wiz. However, the following aims to provide an understanding of a couple of key concepts, which are important to help understand the terms and conditions of ChatGPT. Specifically, these are Large Language Models (LLMs), and Machine Learning.

As you no doubt know, Artificial Intelligence (AI) is a field focused on creating computer systems with abilities that mimic human intelligence. Large Language Models (LLMs) are a subset of AI, which work by leveraging patterns and statistical relationships between words, to predict the next most probable word in a sequence. ChatGPT is a LLM, which uses this word prediction in the background, to provide you the answers to your question.

How does it learn to predict the words and develop the answer, you ask? Through Machine Learning. Machine Learning is a process that imitates human learning, to improve performance and accuracy over time. For a LLM, this is done by feeding massive amounts of text data to it to automatically process, test itself on, and eventually, apply to refine the underlying algorithms and improve future responses.

Understanding these concepts is important, as this fundamental  design, that combines LLM and Machine Learning, becomes an important determinant in ChatGPT's terms and conditions (specifically around training and information retention). On that note, let's now jump into these T&Cs and see how this is the case.

Overview of Terms and Conditions

As indicated earlier, the analysis below is limited to the terms and conditions for ChatGPT as a service for individuals in Australia,² rather than any enterprise or business offering.³ While the core terms and conditions are established by OpenAI's Terms of Use, which form the agreement between the user and OpenAI, as well as the Service Terms, the full contractual landscape is slightly more intricate, with requirements to comply with all applicable laws, as well as any documentation, guidelines or policies made available by OpenAI. Key additional components of ChatGPT's terms and conditions therefore include OpenAI's:

• Privacy Policy,
• Sharing & Publication Policy, and
• Usage Policies.

Below is a high-level summary of ChatGPT's key terms and conditions. Broadly speaking, these align with what you would typically expect in this kind of agreement. The most interesting terms are those relating to content ownership and usage and accordingly, these are dealt with seperately, in greater detail further below. 

There are of course other terms and conditions which are not captured above, such as the Dispute Resolution terms (omitted from this summary as it is unlikely to be relevant for an individual user in typical circumstances). However, rather than focusing on standard T&Cs, as previouisly mentioned, the terms governing content ownership and storage offer a far more compelling area of focus, as explored below.

In-depth Insights: Content ownership

Under the terms and conditions, users explicitly retain ownership rights to their input, and OpenAI assigns any and all right, title and interest in the generated output to the user. In other words, the user owns the content produced by ChatGPT. From a creator's perspective, this is particularly interesting as ChatGPT can now empower uesrs to create content or conceptualise and refine a wide array of ideas, including developing novels, film scripts, marketing strategies and even detailed methodologies.

Although users are given ownership of content, this comes with a critical caveat: the user also bears sole responsibility for ensuring that any subsequent use or publication using ChatGPT-generated output does not infringe upon existing intellectual property rights. This raises pertinent questions:

• Given that ChatGPT is trained on inputted information, how unique is the output truly likely to be?
• Is there not a risk of it also giving other people the same or substantially similar answers, potentially leading to a 'creative race' (i.e. similar to the race to register domain names, people may start to bulk generate content using AI, and publish this so that they might then claim intellectual property rights if someone later comes up with a similar idea)?
• And considering AI's ability to 'hallucinate' or generate inaccurate information, how trustworthy is an alleged 'unique' answer, and how much should a user be required to confirm the originality of the output before relying on it?

Whilst the terms and conditions content ownership are seemingly straightforward, this area is poised to become increasingly contentious as more individuals rely ont he likes of ChatGPT to create content.

The Sharing & Publication Policy does try to provide some guidance around this tricky issue, by requiring users to clearly disclose the role of AI in formulating content. However, as this relies on humans to acknowledge any use of AI in creating content, it is not clear that this will be enough. Ultimately, the issue of content ownership, intellectual property rights and creativity highlights the ongoing need for critical evaluation and accountability even with AI-assisted content creation. This issue has already hit home in education, for example with students misusing and/or misrepresenting AI content, and will likely expand to other areas and challenge himanity in due course, eventually demanding further consideration and resolution.

In-depth Insights: OpenAI's use of content

Remember how we established that ChatGPT is a LLM that leverages Machine Learning? Well, what better way to gather and feed data then leveraging the content generated and received by content (including inputs) to maintain, develop, improve and train ChatGPT. From a privacy perspective this might be cause for pause, and an individual should always think twice about the information that it is feeding the likes of ChatGPT.

However while this principle of LLMs using content for Machine Learning is generally well known, a less commonly known fact is the availability of an opt out option. By selecting to opt out, OpenAI will stop using a user's subsequent conversations to train its models, offering concerned users a privacy improvement. In addition, ChatGPT's Temporary Chat function appears to provide users another privacy safeguard. Specifically, Temporary Chats and related content do not appear in history, use or create memories, and crucially, are not used to train models. This does not mean go and put all your sensitive information in a Temporary Chat, as OpenAI may hold on to the information for up to 30 days (for safety purposes, not to improve the models). However, it does present a slight mitigation measure again for those people who are worried about the information that OpenAI may be retaining and using for training.

It is also worth noting that, contrary to some apprehensions, OpenAI does not simply collect and retain all raw input data indefinitely. Instead, it indicates that it employs a process to reduce the amount of personal information in training datasets, ahead of using them to improve and train the models.

When you consider this sanitisation of data, coupled with the opt out option and Temporary Chat capability, it really does seem like the focus is on using content to improve the service over time, rather than collecting and building profiles on everyone and everything that interacts with ChatGPT (as I seem to have been told all to frequently).

Conclusion

So, what would I say are the key take aways from this review of the terms and conditions for ChatGPT? Well, there are several and, while they might not be as scary as some people might suggest, they are certainly not without their complexities.

Overall, the terms and conditions are generally standard for a digital product. However, the area of content ownership in particular presents potential future challenges. As AI tools like ChatGPT become increasingly prevalent, the concept of genearting truly original thought and content may become more nuanced, potentially leading the difficulties in intellectual propoerty rights. Privacy and security also remain important considerations although individuals do have options available to somewhat mitigate their concerns.

So, the final message out of this (which I will not tell peopl ewhen they try to share their misinformed beliefs about AI), is go with caution, engage ChatGPT's features (like opt-outs) as you feel appropriate, and remain aware of the evolving capabilities and potential future changes to the terms and conditions.