Most of us buy goods on the internet without reading the terms and conditions. We take it as a given that the language in these standardized agreements is non-negotiable, and hope that it is in our best interests. Too often, this does not seem to be the case. From laptop makers to airlines to buy now, pay later companies, there have been endless clashes over whether terms and conditions are fair. Many consumers, for example, have experienced difficulty enforcing their contractual rights or were unaware of what they agreed in the first place to when they bought something.
One argument in response to such consumer fury is that people should read the relevant terms before clicking “I agree,” and the reality is that most of us do not have the time or ability to do so. But there may soon probably be a solution to this. Instead of having to plough through (and make sense of) all of the small print each time we download a new app or make a purchase online, we might soon be able to do it using artificial intelligence.
What Exists Already
Artificial intelligence (“AI”) tools to analyze legal documents have existed in a very basic form for a while. They can flag potential issues, such as rights violations that the consumer might want to investigate further. Not a perfect solution: in many cases, consumers will have to copy and paste terms, sentence by sentence, because the AI is limited in how much text it can handle. These models are designed to serve as a guide to reading a company’s terms and conditions blurb yourself rather than removing the need altogether. There are more sophisticated AI tools that solve the related problem of reading web policy documents. Rather than pasting in text, you upload the relevant URL.
The important – but narrow – focus here is on how web providers use your data; this makes it easier to teach an AI model everything it needs to know – particularly in such a heavily regulated area.
With terms and conditions, the challenge is their often-varied nature. Vendors are relatively free to formulate the Terms in their own words, and such a lack of uniformity makes detecting and understanding these agreements much more difficult for an AI model. There are also significant variations in companies’ Terms between jurisdictions (right down to the actual jargon), which means that an AI trained with U.S. data may mislead consumers from the United Kingdom. This issue is exacerbated by the fact that it often is not clear in the existing tools which jurisdiction they are designed for.
You might be wondering if the alternative might be just to copy and paste terms and conditions into one of the latest AI chatbots like ChatGPT, but that is not perfect a solution either. These general models are not specifically trained on legal texts or legal analysis, which means that any advice they give is just as likely to be accurate, inaccurate or entirely made up. (One need not look further than the recent case in which a New York-based lawyer who submitted a brief produced by ChatGPT that cited six non-existent judicial decisions.)
Fixing the Problem With AI
As far as we are aware, no team of developers is trying to create a dedicated terms and conditions AI for consumers using models such as Open AI’s GPT-4, which underpins ChatGPT. Instead, many AI developers seem to be concentrating on the more lucrative area of creating tools that will automate legal work for law firms and other companies. (This could even lead to terms less favorable to consumers, since the focus will likely be on cutting costs rather than improving service quality.) To change this situation, Jens Krebs and his colleague Ella Haig at the University of Portsmouth have been developing a terms and conditions app for England and Wales.
When fully developed, it will enable people to copy and paste an entire document into the prompt. It will then list any terms that might unexpectedly affect the consumer, for example, by failing to meet legislative standards, such as the Consumer Rights Act 2015. It will also compare all terms to those used by comparable vendors to ensure that nothing unusual has been slipped in. When it spots something unusual, it will advise the consumer to read that part before deciding whether to go ahead.
The project is currently at the stage of testing the app on different AI models to see which is most effective. So far, Google’s Bard is coming out best with 81 percent accuracy, testing it against data where the researchers know what the perfect result should be. Nothing will be launched until accuracy hits 90 to 95 percent, and the aim is that the app will be made available to consumer groups like in 2024 and then go on general release (for free) in 2025.
The key obstacle for such a project is the lack of examples of detrimental terms on which to train the AI – exactly the same problem consumers face if they are brave enough to try to judge terms and conditions. The long-term plan for continuing to increase accuracy in the Portsmouth app is to supplement and replace its training data with real data from consumer organizations, government agencies, and consumers.
The developers are angling to position the model at the forefront of a new generation of AI tools designed to make terms and conditions less opaque. In addition to potentially reducing the number of unhappy consumers, these models might also help people who are already signed up to unreasonable terms to prepare their case – thereby, reducing the need for lawyers. If such services take off, the hope would be that they also discourage vendors from pushing the boundaries of what is acceptable, and if terms and conditions become a bit more favorable to consumers, that would be a huge win for this emerging technology.
Jens Krebs is a Senior Lecturer in Law at the University of Portsmouth.
Enguerrand Boitel is a PhD Candidate in Computing/Research Assistant at the University of Portsmouth.
Paris Bradley is a PhD Candidate in Law and a Research Assistant at the University of Portsmouth.
Share
