You Ask, I Answer: Can AI Write as Well as a Human?

In today's episode, Deborah asks, Can AI write as well as a human? Well, the answer is it depends on what the human is writing. Um language generation models, natural language generation models, like uh OpenAI's GPT uh series of models uh or the T5 transformers from Google can write reasonably well in a in a very general sense, but the more specific the task, the better they get, the more guidance you give them, uh, and more guardrails you put on them, the better they get. Something like uh a complex original work, like say a novel, answer is no. Machines cannot uh write as well as a human, not without an extraordinary amount of training uh and uh resources to the point where it's not cost efficient.

Uh you would you'd be better off hiring a novelist to write a novel than you would be to try and and train a machine, even if the machine could generate, you know, a thousand pulp novels, the amount of editing time and QA that would have to go into it would effectively negate any gains you got out of it. Now, if it's marketing swill, the answer is yes. And the reason for that is a lot of what we write in marketing is junk, right? It's not very good quality content. It's actually it's pretty appallingly bad content, and it's easy for machines to replicate crap.

Think about a bell curve, right? On one end you have uh low quality content, on another end you have high quality content, and in the middle is a bunch of mediocre stuff. Uh as machines evolve, they go from they can replicate total garbage, right? Because that's pretty easy. Um, and then they start to iterate and get it slightly better.

And now they're kind of at the the mediocre uh phase, right? They can write mediocre marketing swell. Can they write good quality copy or great quality copy? Not really, not without again a lot of training to the point where uh it is not cost efficient to do that, unless you're a company that is dedicated to to doing that. But uh if you were to look at the metrics, in fact, let's do this.

Let's bring this up here. Um what we're looking at here is uh four sets of readability scores for a lot of marketing uh copy. So this is uh articles, blog posts, um press releases. Essentially it is content marketing stuff, right? And not very good.

And this is a hundred and thirty thousand of these that uh we have stored in the trust insights database. We have four quadrants. We have the s jargon score, uh, which is the smog score, simple measure of gobbledygook. Um we have the flesh concade readability index, we have the Coleman Liao index, and the Flesh Concaid grade level. What you notice here is that there is a definite skew towards uh the right side for three graphs and uh sort of a bell curve in the middle.

So let's walk through these briefly. The jargon score, uh lower is better, right? So you want uh that content to be easier to read than harder to read. And in this case, there's a very definite kurtosis or lean um towards the uh harder to read side. It means that our content is full of jargon uh and a lot of it.

On readability, uh the readability score uh one hundred is is great, zero is bad. Uh and we see a bell curve there in in the middle, you know, fifty, fifty five, sixty is where uh most content is. So it's it again it it's mediocre content, right? Um there's very, very, very little on the far side here of the readability index that says, yeah, you got great content. Um very thin.

On Coleman Liao, again, this is uh similar to grade level. You can see there's a tremendous amount of very difficult to read content on uh the far right-hand side, and then a whole absent field, and then on the flesh concade grade level, uh, we see that marketing content is around ninth ninth to eleventh grade content because of the jargon, because of the the amount of stuff that we put in our copy that is difficult to read, that is um extensively polysyllabic, which means that we use real big words. Um, think about the uh the list of corporate buzzwords that we love to use, and you can see that reflected here in this data, that this data very clearly shows we love our our fancy complicated language. When you have language like that, when you have uh you know buzzword bingo, uh we have templates and and an awful lot of things like you know, basic blog posts are very templated, when you have press releases, uh announcements, they all follow copy that is so formulaic that it is very easy for machines to replicate and probably do a better job than the majority of humans because when these natural language generation models are trained on language. They're not trained on just one specific domain.

They're trained on as much language as possible. Um, the most recent version of uh GPT-3 was trained on something like 175 billion parameters, uh, which is a massive, massive, massive amount of text. It's basically the bulk of the readable uh text uh in English online. That means that when these machines go to generate language, they will naturally use more vocabulary, a little bit more uh lexical diversity is the technical term, than uh a human would, because in a lot of cases, uh humans will just copy and paste the last thing they did. I used to work at a public relations agency, and I would literally watch you know a junior associates just copy and paste from one press release to the next, change the name and the uh of the company and the name of the CEO, but effectively they all say the same thing.

You know, the the chief executive says, you know, we're so excited or pre uh proud or pleased or release our new version, whatever, whatever, whatever, uh, and we're a flexible, scalable, industry leading, agile, you know. Can a machine replicate that? 100%. 100% a machine can replicate that and do better than a machine than the humans do. So can AI write as well as a human?

It depends on the context, but for sure, um, machines can now write at I would say the mediocre level, right? The they could they've got bad down, they've got ehh down, they've got uh they're they're getting mediocre down now. As each model improves, as the technology improves, they will eventually get down good, good writing, right? And for those companies that have the uh strongest infrastructure and the h you know greatest level of resources, some will get great writing down. What does that mean for you as a as a marketing practitioner?

It means that you've got to be improving your skills. If you are on the bad side of uh writing, you probably need to stop writing for marketing and look at a different profession because the machines can already crank out swill better than you can if you're in the the meh to mediocre you better up level those skills take some courses do you know some writing workshops do anything that whatever you can do that will improve your skills and get them to good right um if you're a good writer you know keep working to become a great writer but whatever you do you cannot stay static you cannot just rest on your laurels at wherever you are in this continuum because the machines are advancing faster than than we are collectively will there come a day when you push a button and a machine spits out a novel probably we're already seeing we're prototype examples of this with GPT 3 the open AI uh model is it good yet no um but it definitely shows what is possible uh what is what is theoretically possible uh and what is possible today is easily achievable in in five to ten years maybe less depending on how fast compute uh power goes up so that's the answer to this question AI can write as well as humans who are bad at their job. AI can write as well as humans who are mediocre at their job and AI will eventually write as well as humans who are good at their job. So your job as a human is to become great at writing so that you stay ahead of the machines. If you have follow up questions, leave them in the comments box below.

Subscribe to the YouTube channel in the newsletter, and I'll talk to you soon. Take care. Want help solving your company's data, analytics, and digital marketing problems? Visit TrustInsights.ai today and let us know how we can help you.