Stop Multicasting Slop

Should you use AI for writing, presentations and communications?

My current hot take is that AI makes my own writing worse and it can get in the way of my thinking, but it is a very useful reviewer. I do use AI for basic documenting and reporting tasks. There is nuance to this, so I'll elaborate.

When you communicate one-to-many, your time is less important than your audience's. The bigger the audience, the greater the burden. On a personal blog, you can be forgiven for writing for yourself and your own thinking, but when the audience is effectively captive like a conference keynote or when you write for a print publication, you either make use of people's time or you waste it, and in doing so waste the opportunity. Hundreds or thousands of people sat reading/watching you (and usually more online) deserve respect. Slack messages in large companies often count here too. Hold yourself to account.

Slide decks are one of the worst offenders. It appears as if many of the AI pilled think themselves above the craft, or without the time for it. Perhaps it's a lack of belief that they could do better. Conveying information in talk format is incredibly difficult, most of us are bad at it. AI can improve the visual aesthetics of a deck beyond what you might be able to achieve yourself (though paying artists for graphics is a great option, and their work will be noticed), but you control the information density, the exact content and final output. The art of making niche or complex information simple to digest remains. Refine your craft, don't recoil from it and don't let AI replace your thinking.

Ironically, whether you use AI or not is irrelevant. People experience the exposed interface (in this case a document or presentation), not the process that got you there. People regularly conflate their subjective and objective self, and wonder why people don't react to them as they are. The you they imagine is the sum of interactions, expectations and prejudices they transpose on to you. There are almost certainly more people that believe in this version of you than there are people that know what you experience internally, so external communication really matters if you care about influencing this. In short, your job as a writer/presenter is to hone that interface, but how you get there is for you alone. Good writing is possible with AI assistance, but knowing what good writing is is the same challenge either way. People will only see the output. You still own it if an AI writes for you.

The sunk cost fallacy also plays a large part: people struggle to throw out bad drafts and instead waste time refining bad human or AI output. It's often quicker to get rid of bad content and optimize the process of re-drafting. I encourage you to try. Getting to 90% there from a cold start might not help if the final 10% costs more time than it takes to create the whole. Prompting current models for certain creative tasks can feel like a tennis match stuck at deuce: you're close to winning, but slight misfires prevent you making any concrete progress. Recognise and avoid this if you can.

A few things matter especially in one-to-many communication:

  • Efficacy/efficiency: actually transmitting what you wanted to communicate clearly and with respect for the recipient's time.
  • Structure/presentation: is the information density right, does the audience deepen their understanding progressively as they give you more of their time.
  • Tone/character: does anyone enjoy the experience of your communication? Does it help them digest the material? Will they remember it?

What using AI does is make it easy to do these things poorly, at scale. For example, sloppy slide decks almost universally have poor information density, bad titles, surprising text wrapping and confusing terminology. Those things cannot be effectively overlooked even if you are a great speaker, because the act of looking at them is already costly for the audience's attention. I have sometimes lost respect for speakers when subjected to this. I just don't blame AI. These are choices a person has made.

In written text, slop tends to look like known AI tropes, repetition, odd structure, confusing terminology and often either very staccato or very verbose sentences. Readers have to think harder to read slop content versus well written content, and so will increasingly just stop.

Some of this is down to fundamentals of how models work, they are probabilistic machines in the end, and so are unlikely to produce output from the high entropy1 space (most of which is garbage), but as a writer it's the rare but good high entropy parts that offer the most value - the parts that are not just the same as everything else but somehow unique. Sure, you still need to wrap those parts in common language, with familiar mechanisms to aid in comprehension, but at its core the motivation to read most content is driven by an expectation there is something of value contained in it.

The craft is orthogonal to the AI usage part, it's all about knowing what a good end result is when you see it. Don't multicast2 slop.

Footnotes

  1. In the information-theory sense: a measure of how unpredictable an outcome is. See Entropy (information theory). ↩

  2. A networking term for sending one transmission to many recipients at once, i.e. one-to-many communication. See Multicast. ↩