It’s a question I’m being asked a fair bit at the moment and for the legal-minded folk I’m working with a great deal at the moment, this is how I respond…
Machine Learning is really interesting. It’s being applied to all sorts of problems where traditional models of programming have fallen short. It’s mostly based on types of statistical analysis of very large data sets. It’s also not infallible – it tends to rely on spotting correlations as being more important than understanding causation. If that leads to useful things, that’s not a problem. If you are looking for something absolute, then it can be.
Put it this way: Google Translate, using methods of Statistical Machine Translation, has managed to move automated language translation between dozens of languages forward quite dramatically in recent years. You can quickly get the gist of text written in a plethora of unknown languages, and all for free. It will even translate text in photographs, and the spoken word too (for some languages).
But it’s not an absolute translation. You get the general idea of what the original text said, but particularly when taken out of original context it can become gibberish.
By way of example, here’s that last sentence translated in Google from English to Portuguese to Welsh and then back to English:
“You get the general idea of what the original text, but especially when taken out of context can become meaningless original.“
Not bad, but not quite good enough for, say, inclusion into a legal contract. And that for me is the benchmark today for the use of Machine Learning. It’s good, but it’s probably not good enough to be used to create legally binding documents.
So where is it useful? Well, in a legal context, Machine Learning technologies could be quite useful for adding metadata to an archive of old contracts. You could use them to mine through text to spot particular types of information – dates, contract clauses, addresses and so on – and then pull that data out to provide a structured database which in turn could be used to query and discover the original contracts. But you would be brave to allow the machine to then make actual decisions based on the information that it had created.
The more specific the task, the more accurate it is that machine learning algorithms will get things right more of the time. But there again the more specific the task, the more likely it is that more traditional computing approaches would achieve the same outcome.
We are obviously at a phase of particular hype for machine learning technologies. There are no doubts that they will provide greater and greater benefit as time goes on, and they are providing useful things already. But, like any tech hype, understanding the technology strengths and weaknesses are crucial to understanding where they are likely to provide benefit in the future.