Imaginary Friends 3: ″Capital Punishment?

Search form

Monday, 17 March, 2025 - 10:52

″Artificial Intelligence″ and ″Machine Learning″ depend on, amongst other things, the language used by humans. Four imaginary friends discuss the uncertainties that creates.

Location: an imaginary fireside

A. It’s like I’m, like, likin’ it back here in our comfy lodge. Let’s make a start by thinking about clarity in language.

B. Are we speaking about written or spoken language? I think we should stick with one because there are so many differences between the written and the spoken that to talk about both creates unnecessary complexity.

C. Maybe we should look at that first i.e. how the differences written and spoken cause that complexity.

D. Yes, because that is central to whether software can be produced to accurately interpret the information that is presented to it.

A. I have a particular fascination with emphasis and intonation. As someone who communicates on the page and orally, I am always looking for ways to make what appears on the page carry my voice.

Can you hear what I wrote?

B. That’s at odds with legal drafting, or at least legal drafting as it used to be. If you look at English legal documents from before the 1970s, aside from the annoying heretofore and similar expressions one of the most noticeable factors is the lack of commas, semicolons, colons or anything else that might be used to guide how the reader ″hears″ the document.

A. Yes, I think that’s where I first became fascinated with trying to get my voice onto the page. As a litigator, I wanted the judge who was reading my pleadings, affidavits and supporting documents to get my point. I wanted to remove his ability to interpret. So I developed a style of drafting that was very similar to how I would speak in Court. The plan was simple: my actual speaking would influence how he ″heard″ what he was reading either before the case or when he retired to make his decision. The written and the spoken word reinforced each other. It’s why the so-called Oxford Comma should be banned: it breaks up the flow of a sentence while adding nothing to its comprehension – except in some cases: it’s needed in a list but even that depends on how the list is structured.

B. If I look at a conveyance of title, a property transfer, from the late 18^th century, it’s full of the wherefores and wherebys and so on but if we strip them out the operative part of the document is very short. Basically it’s X sells to Y for Z pounds and Y agrees for himself and future owners to use the land for a specific purpose.

A. That’s a perfect example of how anyone programming software needs to understand far more than the words on the page. It’s not even a question of context. It’s a question of fully understanding the subject area and being able to tell a computer what you know and what to do with it. As you phrased that, the agreement as to future use is not enforceable against anyone other than the first purchaser i.e. Y. That’s because as you phrased it, it’s a ″positive covenant.″ Positive covenants are personal to the parties. But a negative covenant ″runs with the land.″ So if you’d said ″Y agrees for himself and future owners not to use the land for anything other than a specific purpose″ that is enforceable against him and any and all future owners.

B. So any software review of that document needs to be programmed to differentiate between those two?

A. Yes.

C. We’ve moved away from the comparison of written and spoken language.

A. Yes and we will do it with almost every point we consider because we are demonstrating that every complexity hides another and another and another. How many layers of complexity does the programmer have to consider before he’s looked far enough?

B. Please tell me that’s a rhetorical question and we’re not going to talk about golf balls again.

A. Yes and no: we should bear the golf balls in mind in relation to every aspect of what we are looking at because they help us understand the parameters we must work within – or set or ignore and the inherent uncertainty that we have to deal with.

Putting a ″voice″ on the page.

C. So far we’ve seen that what is written and what is read may be entirely different depending on whether the writer can influence the reader into reading in the ″voice″ he used when he wrote. And we’ve seen that whether he does that is at least in part influenced by punctuation. Logically, we should be able to extend that to the length of sentences, phrases. Those things don’t change depending on whether it’s the written or spoken word.

D. No, they don’t – in theory - but in practice they do. I think we need to think about the quasi-mechanical process by which a machine processes the written word.

We give it a passage and it interprets it. Let’s take this from Lindisfarne’s Winter Song:

When winter's shadowy fingers
First pursue you down the street
And your boots no longer lie
About the cold around your feet

First we wonder: is ″winter″ a proper noun? That depends on whether Winter has personality. It’s almost the same argument as to whether a computer has the capacity to think or is merely mechanical.

Or so you might think but if you were in the audience at the Magic in the Air concerts in 1977, you would know, from the way the word was sung, that in the song Winter is a malevolent force against which man has no defence and is left only with the power to question, to challenge others as what they are doing to counter that malevolence.

If you listen to the original recording, there is no attack on ″winter″ - in fact, it’s almost resignation. The feeling is not one of anger or a desire for resistance because, hey, it’s winter and it comes and goes every year and it does what it does whether or not we like it.

With the capital letter, Winter is a force of nature; without it it’s just a dreary season.

Who on earth is going to programme that? Who is going to spend the time looking at each word, each letter in each line of each stanza of each song from each songwriter there’s ever been and will ever be. And if they do, unless they have access to the original writer, how is what they programme going to be anything other than their opinion which may, or may not, be what the songwriter intended?

C. That’s another side-track, isn’t it? That the analysis depends on the interpretation of the programmer or programmers. That of itself raises a question: can we rely on a single programmer? And as a separate question, can we rely on a committee? Or do we have to program all the opinions? If we program all the opinions, do we have to prioritise them?

D. That’s a fundamental question as to the value of products that perform analysis, isn’t it? Isn’t that where we start to get into trouble on diagnostics – regardless of the industry in which those diagnostics are performed, be it engineering, medicine, financial services even law.

B. Are we talking about uncertainty?

A. In our very first conversation, we talked about yes, no, maybe. Yes and no is easy: that’s the basics of the simplest of algorithms. The maybes are inherently uncertain and they include the unknowns that, if they aren’t identified and programmed for, give unintended, if not unexpected, consequences.

D. We usually end our chats with a conclusion. Do we have one from this discussion?

C. I don’t think we do. I think we’ve outlined a fraction of the questions that need to be asked and we are as far as it’s possible to be from reaching a conclusion.

A. We can conclude one thing : the direction we are travelling in is that it appears that the only route to what some might consider certainty is through superficiality and simplicity and that anything beyond mechanical calculations and yes and no opens doors to too many options which in most cases cannot be protected against or provided for. That’s contrary to the entire point of both ″artificial intelligence″ and ″machine learning.″ Is that a conclusion we are happy with? I’m not.

B, C and D ″not at all.″

D. I feel Imaginary Friends 4 coming on.

----------------------------