Why Em Dashes are the least of your AI writing worries

Cliche - A phrase used so often it has lost its original impact, meaning and creative power.

I guess, on reflection, it was inevitable that we would hit this point with using LLM’s for writing.

They are literally designed to produce - and ultimately invent - cliches. By stacking up billions of words according to predictive logic, they themselves become predictable.

Em dashes are the loadbearing tell but you can add contrasts, empty words and clumsy juxtapositions to an increasingly widely recognised list of AI writing signatures which, increasingly obviously, switch readers off.

If everything you read is ultimately the output of one writer (or machine) then of course its impact rapidly dwindles. Novelty and the component parts of personality remains an essential driver of interest for us humans.

So…writers, rejoice?

Maybe not yet. The sense of satisfaction we feel at the idea maybe AI can’t yet lay claim to a future monopoly on writing will probably be short-lived as it becomes clearer that AI writing is only the tip of the iceberg. If people hate the writing AI produces then people won’t read it and perhaps then people won’t do it in the first place (although I fear this is unlikely to be true in many mass-market contexts).

But none of this means they will use it any less to inform their work.

Which is why the bigger challenges sit around how we collectively handle the role of AI-thinking.

What’s AI thinking?

For the purposes of this piece, it’s the outcome of an exchange between an LLM user and the LLM. It can be visible - in the form of a document or artefact or it can be invisible - in the form of an idea or assumption.

I’m fairly confident I’ve spotted the vapor trails of AI thinking in the media on multiple occasions this year as an argument or analytical approach applied by a writer carries an eerie familiarity to my own exchanges with Claude on a similar topic.

In one conversation exploring the likelihood of an AI stock market crash I was encouraged to apply a monitoring framework that bore a striking - and unacknowledged - resemblance to the one maintained by a widely read AI and technology observer. There’s no evidence of cause and effect but a similar sensation crept up on me when asking about the future of the BBC license fee and seeing the same arguments offered up by Claude, in print, in similar language, by the weekend.

If our lines of thought become slowly homogenised as we blend them with AI’s pattern based ones then this problem will manifest at a deeper and different layer of our social behavior than if we all get fed up of reading the same writing.

I should note here that there are many different degrees of AI thinking - they can range from single prompt, single output documents or takeaways to mammoth undertakings spanning hours or even days of work captured in a huge deck or even interactive webpages. All of which carry different amounts of user comprehension and intent in their end form. Sometimes it’s possible to make several cognitive leaps in your thinking during a session working with Claude but if you can’t capture and relay those to the reader when you attempt to do so with an AI generated output then communication suffers and challenges arise.

AI thinking in the workplace

Dealing with AI thinking is a zeitgeist challenge. For those of us who work with AI quite a lot it is possible to make large leaps forward in your understanding of, perspective on or attitude to a particular issue during a session working with Claude. This can involve using large amounts of contextual information, sourcing new data or synthesising several strands of thought.

For individuals this feels like, and in my view often is, progress, but for organisations it only counts as worthwhile if other people can benefit from this AI thinking. It’s not for nothing that the term ‘meat-proxy’ has begun doing the rounds on social media to describe somebody who uses AI in a way that reduces themselves to a mere conduit between the LLM and the recipient of their ‘work’.

One of the most time consuming parts of modern work is translating our own thinking into something that others can engage with in order to make informed decisions. This often involved creating decks, strategy notes and briefings that attempted to set out the operational context from the perspective of the author.

AI outputs are now frequently being used to try and shortcut this process and for many of us it is messy and painful. As Rootcause CTO Ben Phillips sets out in his own piece, sharing AI writing levies additional ‘cognitive taxation’ on the reader by demanding they overcome some of the inherent problems with AI writing themselves.

This is an important point. None of us wants to make life more difficult for our colleagues. But is there scope for sharing any AI writing at all?

Ben proposes a set of new norms. We’re going to be exploring these at Rootcause but I think the most important thing to focus on is the importance of being able to explain your thinking and to be respectful of the time and expertise of others. A short, human authored cover note for any AI-generated content feels like an essential bare-minimum.

But norms do not exist in isolation within any organisation. Right now if I share a branded diagram with annotations drawing on a wide range of internal documents then there is a sub-conscious expectation that the content is high-value and near finished. This expectation is analogous to what we’re seeing with the collapse of trust in institutional brands - we used to trust institutions that looked like they were trustworthy because they have the resources to communicate in ways that looked polished and conformed to the norms of a mass media era.

But if creating diagrams or artefacts (or even brands and the ephemera of an organisation) is now far easier, then that assumption may collapse. With the way LLMs are being designed It may become quite normal to share something that looks comparatively polished by today’s standards as the equivalent of messy bullets and notes.

One of the things you hear from those at the forefront of AI innovation in software is that software could soon become ‘throwaway’ - spun up for a particular purpose in a particular moment and then discarded.

There’s potential for something similar in how we exchange our AI thinking.

For some of us the ability to ‘work with the gist’ is helpful, even if it exposes us to content that is uncertain, inaccurate or even contradictory but for others clarity is king as it enables implications to be identified and planned for with maximum efficiency.

I’m lucky enough to work in an environment where the culture is one of joint exploration and discussion around the impacts of AI. We see every day the power of this technology but in the realm of writing and thinking we are not immune from its shortcomings.

Our Drumbeat Machine has generated A LOT of AI writing - literally millions of words - of analysis of news and social media. In that time we’ve had one consistent piece of feedback from our clients which can be summarised as - make it shorter, the writing hurts my head.

That’s something we’re in the process of fixing. It’s reassuring to me that as feedback goes this particular piece has been in circulation for hundreds of years.