AI in IP: Everybody’s talking about ChatGPT
We may well look back on 2023 as the year Artificial Intelligence (AI) came of age. Until recently, the idea of an intelligent computer that generates human-like responses to questions was largely the preserve of sci-fi novels and Hollywood blockbusters. However, the launch of Open AI’s large language model ChatGPT has brought the potential of AI home to everyone (and there are millions of us) who has tried it out in the past couple of months.
There’s no doubt that ChatGPT is impressive. Originally developed as a customer service chatbot, it has been trained on a vast dataset that enables it to effectively act as a search engine on any topic you care to quiz it on. In response to prompts, it analyses its dataset and processes its response using Natural Language Processing (NLP) to deliver an output in the form of grammatically correct, contextually appropriate text. If you haven’t already tried it, we recommend you have a go.
ChatGPT is extremely flexible and can tailor its response to fit the expected parameters of all sorts of communications. Blog posts, academic essays, songs, poems, letters – you name it, and it will appear within seconds of you inputting your prompt.
As such, it is already raising concerns among teachers, who are seeing students submitting assignments generated by ChatGPT, and there is a range of AI-detector tools on offer aiming to nip this academic shortcut in the bud.
Across knowledge-based industries, including law, there is a mix of excitement and concern as we analyse what the advent of this technology means in terms of opportunities and threats. How trustworthy is AI-generated output? How can it be used as a productivity tool? And what are the intellectual property considerations around the content produced by AI?
While this is an area that will undoubtedly evolve rapidly, we have considered some of the early issues and opinions on the topic.
Despite its confidence, accuracy and consistency can be a concern
As Google discovered during the launch of its own chatbot, ‘Bard’, a large language model trained on human-generated content can be prone to making human-like factual errors. In this case, Bard confidently stated that the James Webb Space Telescope was the first to take pictures of a planet outside the Earth’s solar system, which experts quickly pointed out was not the case.
The problem is that the chatbot is designed to deliver confident responses. These are convincing on a first reading – especially if you are not an expert in the subject. However, as with most assertions, just because they are convincing and confidently made, it doesn’t mean they are correct (as anyone who has ever done a pub quiz with my partner will attest!).
At WebTMS we put ChatGPT through its paces on the issue of Colin the Caterpillar, asking it to write an article about the outcome of the case. It confidently told us that Marks and Spencer had won the case, when it was actually settled out of court and – given that Cuthbert is resolutely back on the shelves in our local ALDI – it seems a stretch to say that M&S was the winner.
So, we advise ChatGPT users to proceed with caution and check all assertions for accuracy in the same way that you would verify any search engine results – however confident the response seems!
Similarly, Clarivate’s Robert Reading found an issue of consistency when he asked the AI to draft a trade mark application for ChatGPT covering “chatbots” and including the appropriate NICE classifications. While he was impressed with the initial results, the same question asked six hours later generated a different, less detailed response. In the subsequent discussion, Robert notes that it is “intriguing that you can ask a “machine” a question twice and get two different answers.”
You could say it is natural that a “learning” entity that is constantly updating and evolving will deliver varying responses – although you might assume that each response would become more refined, rather than simply being different from the previous answer, as in this case. Certainly, if nothing else, we need to be careful about the assumptions we make around responses from AI chatbots.
Copyright issues and AI as author
Ultimately, AIs are trained on existing data, which makes it likely that there will be copyright issues somewhere along the line. This is clearer in some cases than others. For example, Getty images is currently suing Stability AI, creators of open source AI art generator Stable Diffusion, citing “brazen infringement of Getty Images’ intellectual property on a staggering scale”. Getty Images claims that Stability AI copied 12 million of its images to train its AI. Among other effects, this has resulted in some of the AI-generated images containing Getty’s watermark, which amounts to trademark dilution and/or blurring, according to the company.
On the other side of the coin, there is also a considerable debate around the protectability of AI-generated works – whether they are text or images. Who is the author when a work has been created by algorithms in response to a human prompt? Can a work be said to be original when it is devised entirely as a combination of existing works – albeit brought together in a novel way? And, as we’ve seen from Robert Reading’s example above, given that the response to an identical prompt can vary, does that add more or less weight to an argument for originality?
These questions have been exercising the legal system for some time already and the debate will become more intense as AI – now it is more accessible – is used more frequently in artistic expression and for daily tasks.
The issue of whether AI can be recognised as an inventor is a hot topic at the time of writing. Technologist Dr Stephen Thaler is seeing his bid to have the DABUS AI listed as an inventor on a patent application heard in the UK Supreme Court. He is contesting the UK IPO’s finding that the UK Patents Act 1977 requires a natural person, rather than a machine, to be listed as an inventor. Pinsent Masons has more on the case here. The ruling – which is not expected for some time – is likely to have considerable impact on IP across the whole spectrum of patents, copyrights and trade marks.
For more on AI protectability and how IP law is evolving to cope, Dentons has published a useful blog covering the salient points.
AI as an IP productivity tool
Above are some of the challenges around AI in IP, but what about the opportunities? Undoubtedly, ChatGPT and its fellows are a new and interesting way to search for information and gain inspiration – a welcome anti-dote to writer’s block when the words just won’t come.
Ashley G. Kessler, Trade Mark Attorney at Cozen O’Connor, decided to go a stage further and ask ChatGPT to help with the task of identifying potential brand names in her specialist area of cannabis and CBD companies. She asked, “What are some available names for trademarks for a cannabis brand” and was impressed by both the creativity of its responses, which leaned heavily on alliteration with the likes of “Pot Palace”, “Weed Wishes” and “Herb Haven”, and also with the fact that only two of the ten proposed marks was already registered with USPTO.
Despite this success, Ashley noted that further investigation saw several of the marks already in common law use, and that some would not be registrable on the basis of descriptiveness and genericism – so there is little likelihood of an AI replacing the counsel of an experienced trade mark attorney just yet. Nevertheless, she recommends adding a little AI assistance into future trade mark availability searches.
When considering the use of AI as a content generating tool for brands, there are interesting possibilities. We’ve noted that much of what ChatGPT delivers in response to prompts can be quite bland and lacking in the more human, anecdotal touches. However, an AI could be trained to respond in a brand voice – Aldi’s cheeky customer service team immediately springs to mind – creating content that chimes with the company’s tone and humour.
Certainly, the fascinating discussion around AI, intellectual property and its practical use in day-to-day IP management is set to continue and develop. It is very much a case of “watch this space” … and see if it fills with AI-generated content!