I love music. I always did. Music has always moved me, and I can’t remember a single period of my life without some soundtrack. I remember guitar solos and where the string section comes into the front; I know where there’s a drum fill and how the tune modulates in hundreds of songs. I know this all in my head. Unfortunately, the only thing I can play is music recorded by others. On Spotify.
With such passion for music and a total inability to create any, it’s probably no wonder that I went for the second best thing: a CD collection of just about 1000 albums (as I’m writing this, I wonder if this last sentence requires an explanation for some younger readers). It took two decades to assemble this collection and one music-streaming product to make it ridiculously obsolete. Anyway, music was my thing, and I expressed my love for it by buying CDs (and curating some mixtapes). And as anyone who spent so much time in music stores back in the 80s and 90s knows, any music CD has a cryptic 3-letter code printed on it with a combination of the letters ‘A’ and ‘D’.
The SPARS Code is a classification system defined in the early 80s by the Society of Professional Audio Recording Services. Its goal is to provide the listener with information on how the music was produced: using analog or digital equipment. Each of the three components of the code marks a step in a typical music-production workflow: recording, mixing, and mastering. Each of these activities can be done using analog or digital equipment and is therefore marked with an ‘A’ or a ‘D,’ respectively. So, a CD with a ‘DDD’ code imprinted on it contains digitally recorded, mixed, and mastered music. On the other hand, a CD marked with an ‘AAD’ was recorded and mixed using analog equipment and then digitally mastered.
This classification didn’t mean much to most people. I guess most people didn’t know what these letters represented or even notice they existed. To some, however, these letters actually meant a lot. Whether digital was better than analog or vice versa was debatable; nobody could claim that one is objectively better. But in some audiophile circles, everybody seemed to have an opinion in this debate, which made this information important enough to maintain.
With the rising popularity of AI-based tools like Chat GPT, it is increasingly likely that the content we come across is written by or with the help of AI. As a reader, I feel it is important to know how AI was used to create the text I read. I want to know to what extent AI was involved in generating the ideas manifested in the content. Writing and reading are two important parts of communication; not knowing who you communicate with just doesn’t feel right. As you begin to read this text, you don’t know me. But the more you read, you create a mental model of me or at least some parts of me; we communicate as two people, even if the conversation is mostly in your head. Communication is personal, even when it is as unidirectional as reading a text posted on the Web, so knowing to what extent a machine was involved in creating it is imperative. Much more so than classifying the music creation process as Digital or Analog.
It doesn’t matter if you resent AI-generated content or like the idea of it; if you fight it or are excited by the opportunities it creates. Knowing what it is you are consuming must make some difference, especially when it comes to professional content. A SPARS-like classification system can help us understand the nature of what we read. Like the SPARS code, a Human-Machine Collaboration (HMC) Code can break down the writing process into three primary activities: Research, Synthesis, and Writing. Each of these activities can be done by a human or machine. As you can guess, in the HMC Code, these activities will be marked with an ‘H’ or an ‘M,’ respectively.
Unlike the SPARS code, the borders between a human’s act and a machine’s output are blurred. There could be many nuances and various degrees of Human-Machine collaboration. What I propose might be too simplified or even naive. Still, it gives you a glimpse into how much human thought was poured into creating what you read. Regardless of what you will eventually make of it, knowing that is valuable. So before we move on, know that the article you are reading is HHH (researched, synthesized, and written by a human, which in this case would be me).
Good writing is all about conveying ideas and insights; Ideas and insights are always built upon raw material: previous ideas, pieces of data, and any other information we consume. That is why much (if not most) of non-fiction writing starts with research.
Research is not a technical activity. First, you have to decide what it is you are looking for. Even a simple web search can yield very different results depending on your query. No less important is the need to validate the data you come across in your research either by verifying it is from a trustworthy source or by cross-checking it against data from other independent resources. When the data you collect is incorrect (let alone deliberately misleading), your ideas become questionable too.
One of the most basic uses of text-based AI tools like Chat GPT is for research — a replacement for the tedious task of sifting through search results, reading the sources one by one, and evaluating each. One does not have to be a radical AI supporter to realize the potential of AI-generated summaries for the productivity and efficiency of this preliminary phase of content creation.
The problem is that some AI tools mask the origin of the data. Even if they don’t, the mere use of an AI tool creates a bias toward the results it provides: we believe it is superior to a direct web search, and so we might be more inclined to trust its output and waive (even unconsciously) the need to validate the results. When you take whatever AI throws at you as reliable and use it to generate your own thesis, you can quickly fail.
Of course, research done or validated by a human doesn’t necessarily result in a more reliable outcome. I can blindly trust the first result I get on Google and use it as a fact without double-checking it. When you think of it, taking whatever we get from a web search as given is literally delegating the Research activity to a Machine (even if Google Search is not considered an AI compared to the tools released in the past year). The condition for marking the Research part of the HMC Code as Human-based should be that a human has validated the data.
Will human-validated research always be better than a machine-based output? I doubt it. I can certainly be tricked or not thorough enough, even when I have the best intentions to validate the data before I use it. But marking the Research component with an ‘H’ means I take responsibility for it. It is not a guarantee that my research was good, just like a digital recording could be way worse than an analog one depending on the professionalism of the sound engineers. But it is a statement about the type of validation I conducted on what a black-box algorithm has produced for me.
Simply quoting existing ideas and raw data might have value in some cases. But professional non-fiction writing aims higher: to synthesize a personal insight or opinion based on the raw data.
My writing doesn’t always include novel insights, but they are always my insights. I rarely quote a piece of data or a preexisting idea without saying anything about it. I might argue with it or extend it; I might refine it or add some nuance; I might use it as inspiration and apply it in a surprising way (like using the SPARS Code as a generative metaphor and trying to apply it to a completely different domain). For better or worse, this is where the value of my content is. When I read other people’s content, this is the value I am looking for: their personal synthesis of information (whether this information is new to me or well-known). And I, for one, would like to know whether a human has synthesized these new ideas or a machine.
Again, I’m not saying humans will always have better insights than AI tools. There are plenty of examples where the opposite is arguably true. But as a reader communicating with what appears to be another person’s mind, I’d appreciate knowing whether I read the outcome of human reasoning or the output of some statistical manipulation on data scraped from the Internet. I don’t know what type of insights I will trust more, and I assume my answer will change on a case-by-case basis. But knowing what kind of mind is behind the thesis seems super relevant to this judgment call.
The Synthesis component in the HMC Code can indicate whether a human or a machine processed the information and was responsible for generating new insights and ideas.
The third component of my proposed HMC Code classifies the actual writing. A few months ago, this might have been science fiction to many, but since the launch of Chat GPT, it is clear that many writing tasks at various levels could actually be delegated to a machine.
I already wrote about the value of avoiding delegating writing to AI-based tools, but I’m not naive: many people will use AI for writing texts for them. Whether I read an article or engage in text-based communication with someone I don’t see, I’d like to know who I am communicating with. I want to know whether I am reading a text crafted by a machine or a human, regardless of who did the research and the reasoning.
Of course, if a machine is responsible for the entire creative process, from research to writing, I am not communicating with a person at all. But even if the ideas are based on human reasoning, the craft of writing adds depth and influences the reader. It might seem less important than coming up with the thesis, but I will risk being seen as biased when I say: it isn’t. I vote for transparency about this component as well. I can imagine (although not necessarily like the idea) reading a book written by AI; I cannot imagine doing so without being aware of this fact.
I’m not sure the idea of the HMC Code is practical. There are a lot of gray areas, and I deliberately didn’t go into potential nuances and variants that make this classification far from trivial. Even if everything is well-defined, enforcing such a classification and preventing abuse is practically impossible. It is also safe to assume that such a code, even if it is enforceable, doesn’t solve all problems and every critical question regarding the origin of the content we consume.
If anything, this is a rough sketch of a fundamental and initial step I believe we must take. Soon, the world will be flooded with texts written by or co-written with machines. It might be a disaster or the most significant innovation ever. Either way, as a human, I must know who’s on the other side of the text: who takes responsibility for the quality of raw material used, who came up with the insights and ideas in the piece, and who crafted the actual words and sentences.
We deserve to know who created the content we consume.