I love this post because it shows that we can quantify the impact of AI in an objective way. (As opposed to just saying, truthfully but vaguely, that human communication is being radically transformed.)
At the same time, I think that AI effects on word choice and semantic similarity are exaggerated, because they've been essentially cherry-picked.
McLuhan predicted a global village. What the internet has given us instead are isolated villages, each evolving conventions of writing that many "villagers" over-rely on. Even before the rise of AI, this resulted in cliched word choice and excessive semantic similarity.
The studies you cited aren't very sensitive to this, because they glom together too many "villages".
For instance, thinking back to the 1990s and my earliest experiences as a social science professor, I recall highly cliched ways of describing human development, whether the writer was an undergraduate (one set of cliches) or a scholar with subpar writing skills (another set of cliches).
The fact that more and more people nowadays "delve" into a topic or "underscore" key points only tells us that AI has society-wide impacts on the language. It doesn't follow that the quality of writing in any particular village is eroding (though of course you could cherry-pick examples).
So, maybe AI is merely intensifying an existing problem? I'm guessing most people would disagree...
I have fewer qualms with LLMs used to improve writing because there was SO MUCH bad writing out there before, especially on websites. Unfortunately, the text that's generated with LLMs is easy to spot and easier to dismiss.
I get most annoyed with small businesses using AI for image creation for product posting or menus. If a new restaurant opens and its menu is all AI-generated, there is a 0% chance I'm visiting until I see actual photos of the food.
"AI has arrived to turn human expression into an optimisable science—no different from coding or medicine." (Apologies, I had to correct your Americanism of optimise :D)
This has bothered me for a while so I have intentionally configured my Copilot and Claude accounts to review and analyse my own writing style and to intentionally make small grammatical and spelling mistakes - like accidentally skip small words, or fat finger a few words with letters in wrong places - and to respond in a natural way. I detest the over the top empathetic responses, and have instructed the agents to never respond in that way.
In the end it's not always perfect, but I get a result which sounds "better" if you smell what I'm cooking?
I think this will be the next movement in human-ai interactions - teaching your AI model to sound more like an individual based on your own patterns of communication. If I have thought of this, then no doubt others are doing something similar. I foresee a world where it's not a dystopia of sameness and sedation, but a world where communication and content adapts as a reaction against those ideas. A more hybrid-grey area where the sameness appears where it is 100% necessary, but in less formal arenas - like emails, or slack messages, or social media posts, etc. - the content generated by ai will be akin to the user that prompted it.
But who knows? No one really. Least of us older pre-digital generations.
I love this post because it shows that we can quantify the impact of AI in an objective way. (As opposed to just saying, truthfully but vaguely, that human communication is being radically transformed.)
At the same time, I think that AI effects on word choice and semantic similarity are exaggerated, because they've been essentially cherry-picked.
McLuhan predicted a global village. What the internet has given us instead are isolated villages, each evolving conventions of writing that many "villagers" over-rely on. Even before the rise of AI, this resulted in cliched word choice and excessive semantic similarity.
The studies you cited aren't very sensitive to this, because they glom together too many "villages".
For instance, thinking back to the 1990s and my earliest experiences as a social science professor, I recall highly cliched ways of describing human development, whether the writer was an undergraduate (one set of cliches) or a scholar with subpar writing skills (another set of cliches).
The fact that more and more people nowadays "delve" into a topic or "underscore" key points only tells us that AI has society-wide impacts on the language. It doesn't follow that the quality of writing in any particular village is eroding (though of course you could cherry-pick examples).
So, maybe AI is merely intensifying an existing problem? I'm guessing most people would disagree...
Great piece. The point about the subtler ways LLMs reshape how we punctuate and phrase things, not just the obvious slop, really lands.
I have fewer qualms with LLMs used to improve writing because there was SO MUCH bad writing out there before, especially on websites. Unfortunately, the text that's generated with LLMs is easy to spot and easier to dismiss.
I get most annoyed with small businesses using AI for image creation for product posting or menus. If a new restaurant opens and its menu is all AI-generated, there is a 0% chance I'm visiting until I see actual photos of the food.
On the em dash thing. Now there's a way out -- where humans can use it again. Check out https://realemdash.org/.
"AI has arrived to turn human expression into an optimisable science—no different from coding or medicine." (Apologies, I had to correct your Americanism of optimise :D)
This has bothered me for a while so I have intentionally configured my Copilot and Claude accounts to review and analyse my own writing style and to intentionally make small grammatical and spelling mistakes - like accidentally skip small words, or fat finger a few words with letters in wrong places - and to respond in a natural way. I detest the over the top empathetic responses, and have instructed the agents to never respond in that way.
In the end it's not always perfect, but I get a result which sounds "better" if you smell what I'm cooking?
I think this will be the next movement in human-ai interactions - teaching your AI model to sound more like an individual based on your own patterns of communication. If I have thought of this, then no doubt others are doing something similar. I foresee a world where it's not a dystopia of sameness and sedation, but a world where communication and content adapts as a reaction against those ideas. A more hybrid-grey area where the sameness appears where it is 100% necessary, but in less formal arenas - like emails, or slack messages, or social media posts, etc. - the content generated by ai will be akin to the user that prompted it.
But who knows? No one really. Least of us older pre-digital generations.