CALL US NOW ON 01494 422626
Some of you may know that a few years ago, I built a race car in my garage, using a well-known kit. I raced that car for a few years, wrote it off in an unfortunate accident at circuit, and rebuilt it onto a fresh chassis. I still race that rebuilt car, from time to time.
Building the car was a huge learning experience; there were a lot of new tools to buy and mechanic skills to learn. Whilst a lot of the tools are straightforward – a spanner is pretty self-explanatory – some required a little knowledge. For example, did you know that a torque wrench needs to be wound back to zero (or its lowest setting) after use, else it starts to lose its calibration?
Now, it’s obvious when using tools like these that you need to learn how to operate them, and understand how they work. It can be less obvious in relation to the tools that you use in your business, and here I’m looking very firmly in the direction of AI. Some of these AI tools that have recently appeared do seem very useful… but do you really understand how they work? There are plenty of scare stories about lawyers using ChatGPT to write their legal submissions, and being caught out when the AI “hallucinates”, i.e. creates plausible-sounding but entirely fictitious case law to support the argument. That results from a basic misunderstanding of how a large language model (or “LLM”) works. ChatGPT is not intelligent, it is merely good at guessing the next word in the sentence in order to generate text that appears to be an answer. Ask it a question that is widely discussed on the Internet, and it will come back with a surprisingly good answer that approximates to the general consensus – thus giving the impression of being intelligent. Ask it to prepare a submission for a specific case, however, and there is a risk that it will simply produce a string of words that are in the style of the legal submissions that it has seen elsewhere on the Internet.
There is another more subtle risk that I want to flag up, though. Most people are aware that the free versions of these tools will take the information you give them and use it as training data to try and improve the language model. So your text input will go into the training pool and be kept by the LLM owner – it is no longer exclusively yours. Clearly, that carries a risk of disclosure, so for any sensitive work such as an unfiled patent application, don’t use a free version. The paid versions usually promise to ring-fence your data to just your use, but do check.
However, have you checked out the mechanisms behind all the features of an AI before using it? For example, there was a recent investigation which discovered that the “share” button on ChatGPT actually created a publicly-accessible page with the full chat history on view in plain text. The impression of many users was they were creating a private share with a chosen third party, but actually all that was being shared was a link enabling them to access the open chat. Worse, the pages with the shared cats were being indexed by Google search, so a few judicious search queries revealed all sorts of unfortunate admissions. Now, since that news broke, ChatGPT has removed the shared chats from Google’s indexing. However, that just means that the pages with the chats on them are (much) harder to find; they are still there, though.
Now, patent law has a very simple approach to disclosure – if material has been “made available to the public” then for the purposes of the European Patent Office (and most of the other Patent Offices of Europe…) the material is disclosed and is no longer novel. Famously, a book that is on an accessible shelf of a library is disclosed, even if no individual has ever lifted the book from the shelf. Taking away the library’s index does not change that. So if a discussion about a new invention is shared, that could well be a disclosure of the invention, enough to invalidate any later-filed patent application. Meanwhile, other AIs are still sharing, for example Grok shared chats can be searched via https://www.google.com/search?q=site%3Agrok.com%2Fshare; they are (apparently) short-lived but still open and indexed.
So… what do we take from this? I’m not saying you should stay away from AI – I’m not a luddite, and would be in the wrong profession if I was. But tread carefully, and maybe take the time to understand both what it is doing, and what it perhaps cannot do for you.
Until next time...
Best Wishes
Michael
