Writing
I used to have a blog as a graduate student and since then I have not written regularly anything that is strictly non-technical and in a longer format. Most social media content does not count. In the age of AI generated text, I think writing becomes even more valuable. Old style brain to keyboard word by word production.
I recently had a long conversation about this topic with a colleague at Rice University about AI generated text and what are the norms surrounding that. My colleague suggested that since I have some expertise in AI that perhaps I should take the lead in telling our communications team and other colleagues about what are good practices and what should be the norms and pitfalls of using AI in their daily use.
My short answer: I don't think I can do that yet. The norms around it are evolving as we speak.
First, I use AI to help in writing and I do not discourage my students from using AI to help in their writing. I am not generally against that. I have heard several times the argument that these tools are egalitarian as they allow a non native speaker to write better in a given language. Some scientific conferences have adopted policies allowing the use of AI in writing peer-reviews as long as its use is restricted to grammar and other issues in the use of language. Good ideas should not be dependent on native-proficiency on a given language.
Get Vicente Ordonez’s stories in your inbox
Join Medium for free to get updates from this writer.
On the other hand, we often overestimate how much we are using AI tools in ways that go much further than fixing grammar and language use. AI tools have preferred ways of paraphrasing things and preferred rhetorical devices and sentence constructs that are not easy to spot at first. However, these are increasingly easy to spot for people who are in-the-know or people who just seem to have certain aesthetic sensitivities. This is not just about em-dashes. Here is one sign that people might already be internalizing, it is already recorded in Wikipedia:
Large language models often use the rule of three.[25] AI writing may overuse the rule of three, especially when using adjectives, leading to writing which feels verbose but is lacking in depth.[26]
There is a really pervasive, abundant, and bizarre overuse of this rule in all kinds of writing, prose and communications these days. This might not feel annoying at first, second or third glance but over time it might leave the reader offended, annoyed and frustrated. I hope this captures the point. If you have read text like this recently, this is why. Another thing AI models like to do? Construct sentences that contrast one thing with another. Not only this is annoying it can also make readers not to continue. This is not just happening in writing but often we hear it in speeches who went through writing first. This is not just about the surface and quality of the final text but it is often also about the content of the text. Ok, those three last sentences were the last example to exemplify the point of why some people react to AI writing as something radioactive. I think I am more tolerant but these signs are the easiest to detect. However, there are subtler cues that I do not believe even the most capable models today escape from.
Why do AI models do this? Knowing what I know in how these systems are trained. Several factors. They are trained on large amounts of text from the Internet and other sources, written by a diversity of people with multiple writing styles. The language models being statistical machines, learn how to write in an "average" style as a result. There is also the "post-training" that happens in these models. In their last stages of training, they are trained with user ratings for how pleasing is their output using human feedback. This is not an average of the entire Internet, as the sample of humans that provide human feedback is a lot less diverse than people on the entire Internet. I recommend Alex Hern's article in The Guardian about the preference of AI models in the use of the word "delve" and what might be its origins. Without full transparency on the training process, we can only speculate.
Another thing we discussed with my colleague. AI text detectors, such as Pangram, and do they really work? I will side here with Tuhin Chakrabarty's opinion in this WIRED article:
“Pangram should not be the de facto judgment,” he says. “But I think your own discretion coupled with Pangram’s judgment cannot be wrong.”
I get a lot of email in my position as a professor, many from students but not only from students. The most annoying thing I face is getting emails with lots of text that clearly the person sending not only did not write but more critically, the person did not read. This is perhaps the most egregious misuse these days.
My final message to my colleague was that I think we are at the beginning of an era. Where is it acceptable to leverage AI for writing and to what extent is a personal judgment call at this point in time. Maybe norms will emerge naturally based on incentives and disincentives. For instance, I personally don't want to gain a reputation as someone who writes stuff with AI that I have not even read or thought about myself. My opinion is that you should not use any AI even for grammar or typos in spaces where people generally expect to hear your thoughts in your own voice.
This post is an example of that.
