How to write with (or without) AI language tools
Is AI part of a revolution in how we write at work?
When email was invented, people said “This is the end of work;” “What will we do if we don’t have to call and fax all day?” “What will we do with all the extra time email will save us?” Every technological development has threatened to end the workforce as we know it, to eliminate the need for work at all, to require less human intervention.
Sure, the industrial revolution meant that we needed fewer farmers, but we didn’t end up reducing the workforce overall. And email certainly didn’t end the need for workers–or phone calls.
AI language models are unlikely to be the paradigm that ends the need for work either. We didn’t stop needing to learn math just because we invented the calculator. We won’t stop needing to learn to write because we have predictive text.
What will change is what we do and how we do it. Based on history, we will need more workers, but those workers will need to have new skills to use the new technology. And the old technology may change, but it won’t disappear—for a while yet.
AI language models predict text based on past uses of text. One reason we will still need text is that humans have to create the content that predictive text is based on. No new content, no new learning–which means that eventually all AI language models will put out the same garbage because they won’t have new things to feed into them.
Humans have to keep the machine going.
Another reason we will still need text is that predictive models can only predict within general contexts. The AI doesn’t know your relationships with every person in your Contacts. They don’t know how you socialize with each other offline. At this point, they don’t even know all the ways you interact with each other online.
Predictive language works best for audience’s you’ve never met or whom you will never meet because we can apply standards to language between people who have never met. Cultural norms guide how we communicate before we know each other. These change over time–they actually change with each new generation of teenagers (along with slang and emoji use)--but the AI won’t know that unless it gets fed new information and content from that generation.
For example, my assistant last year sent out a template connection message to maybe 100 people. While the message she drafted worked well for most people, she sent the same message to one of my lifelong friends–we’ve known each other since she was born and I was 4 months old. The message was hilarious to that friend! And sounded like spam. She asked if someone had hijacked my account. The same message went out to a man who had asked me out on dates several times and I had declined. He saw the message as an invitation to pick back up on his pursuit. The same message went to one of my cousins, who I know but don’t interact with very often.
While the template could work for some people, even most people, it couldn’t be used for all the people in that contact list. My assistant didn’t have enough context to tailor the message to some of those people. And AI won’t be able to do that (yet?) either.
AI is good for rough drafts
For that reason, AI language models are going to be great for first drafts, rough drafts, the messy job of coming up with new ideas.
Prompt: Give me 10 possible titles for a presentation to construction workers on how listening skills can improve their safety in the field.
Me: No, those are terrible. Try again. Make them funnier.
Me: Make them more professional.
Me: Make them more strange.
See, the AI can generate over and over and over more content than I can dream of. Because it isn’t dreaming up anything. It is adapting from the trillions of data points of all the writing that it has access to.
But it doesn’t “know” if what it comes up with is good because good changes depending on the context.
This is similar to a calculator. I can plug in equation after equation into my calculator without it getting tired or worn out. But I have to know what I am asking for–just like the prompt for generative language has to be really specific–and I have to evaluate the result.
I can ask my calculator to figure out how many steps less than 10,000 I got each day for the last week. And then I can add those numbers up to figure out how many steps I “owe” myself. And then I can figure out how many steps over 10,000 I need to get for how many days to even out my step goal. I know what the numbers mean and how to use them. The calculator knows nothing except how to compute the math.
In fact, the AI doesn’t know a lot about language. I’ve asked it.
The AI doesn’t know when to use first-person (I) or second person (you) or first person plural (we) or third person (they/it). It doesn’t know when to use -ly words and when not to. It doesn’t know whether using “no” or “not” in a sentence is good or bad. It doesn’t have a system for understanding good or bad writing because 1) it hasn’t been coded with that information. It’s been coded as if all language were equal and 2) it isn’t a person and doesn’t know how people will *feel* about the words that it has produced.
What it does know? How to quickly come up with a million versions of the same thing that are grammatically correct, but not exactly human.
A professor I know told me that AI can write poetry and showed me a poem it had written. Was it technically a poem? Sure, if you think that the only requirement for a poem is that it has meter and rhymes (those are absolutely not requirements for poems). The poem had meter and rhymed, but it was terrible. What made it terrible was that it didn’t have anything interesting or unique in it. It was boring. It was sterile. It was inhuman. The professor teaches economics, not poetry, so he thought it was “good enough”. But “good enough” for what? To be read? To be enjoyed? To meet the basic (inaccurate) definition he had in his head of poetry?
AI is a tool like a calculator. It needs human INPUT and human intervention in its OUTPUT
Think of AI language models as AI language calculators. They might be able to write grammatically correctly, but because they follow rules that are programmed, they can’t actually think. They don’t know what they are doing. They don’t have context. They don’t have input unless you put it in. They don’t know what to do with the output. You have to do that.
To use AI language models effectively, we have to become experts in how to put in the right stuff to get the right result that we can actually use. If I don’t know basic math, then I can’t use a calculator. If I don’t know basic language, then I can’t use an AI language model either.
These tools can’t write *for* you. They can only write *with* you. You still have to know why you are writing, who you are writing for, and what you want to have happen as a result of the writing. You still have to make choices about purpose and audience and context. The AI can show you a million different permutations of the exact same prompt, but you still have to pick the one that will work best.
So how are you going to do that?
By understanding what to put in and what to do with what you get out. And by letting the AI help you do the math faster than you could ever do yourself.
INPUT
For INPUT: the AI needs to know the same things you would need to focus on to start writing. Here are some good questions to generate your initial prompt:
Medium: What are you writing? An email, a script, a letter, a memo, an agenda, a 3 minute presentation, a 30 minute presentation. Each content type has different requirements, expectations, and rules. You need to tell the AI what kind of thing it’s supposed to produce with as much detail as possible. What is it? How long is it?
Content: What is it about? You need to tell the AI what the content of the text is on.
Audience: And who is it for? The audience affects how much information and at what level. How much do they know? Are they 3rd graders or experts? Are they experts in another related field but not this one? Are they people you know or people you don’t? Who is the decision-maker them or you or someone else? How many people? What do they have in common? How are they different? What do they care about? What do they value? What motivates them? How are they going to feel about the message content? Good-is it something they want? OK–is it something they are responsible for already? Bad–is it something they don’t want or didn’t expect?
Purpose: What is the intended outcome? You’ll need to tell the AI what you want the audience to do with your message. Are they studying for a test? Is this content they need to remember? Do you want them to know it? Or do you want them to do something with it? Do you want them to take an action or write you back or engage in another way?
Context: Where or how are you delivering the text? Will it be a commercial on the radio or on tv or on a YouTube video? Are you sending the email to work or through LinkedIn or Facebook or to their personal email address? How formal or informal should the message be?
Organization: Do you want to start by telling them what the content is about from the beginning? Or do you need to provide a little context, explanation, or background before going into it?
If you’ve taken a class with me, you’ll recognize that many of these questions come up in the Planning process I teach. The questions you ask yourself before you start writing are the same things that AI needs to know in order to come up with the best possible draft of what you want it to write. The more you can feed it, the better the rough draft(s) will be.
If you haven’t taken a class with me, I teach the planning questions in many of my seminars and online classes. Here’s a brief handout with the list of questions as I’ve taught them to humans.
OUTPUT
For OUTPUT: You still need to review the generated text to make sure it meets your goals for the intended audience. The most important thing is that writing doesn’t work the way that speaking does. But AI doesn’t know that. It doesn’t know what the differences are. It doesn’t know what tone is or how humans create it or hear it. You still have to review the words to make sure they do what you want them to do.
Big picture, you’ll need to review the message for these aspects:
Complete: Is the message complete? Does it address all the things you want or need it to? Does it provide all the information the audience will need? Always go back to the 5Ws and How for Completeness. Does the message tell the audience who, what, where, when, why, and how? Remember, the AI can only produce based on the information you have provided. If you haven’t told it when the tests are scheduled, then it can’t produce a calendar showing the tests on the schedule.
Clear: Is the message clear? Will your target audience understand it and be able to act on it? AI doesn’t know who knows what. It knows “everything”--or at least thinks it does. You have to make sure that what the AI generates makes sense and can be understood by the people you are writing to.
Concise: People don’t read every word. We skim. But AI does read every word. All of them. Everywhere. And it reflects the way all people write, whether they are concise or not. A lot of AI generated text is wordy because most people write wordily. I blame the education system which is always asking students to write 500 words or a 3 page essay. We are taught to write to a word or page length rather than writing useful content, however long that takes. Ultimately, people ended up learning to write more words than necessary. AI imitates that habit because it doesn’t know how to write any better than any human. It can just produce more writing faster than any human. You’ll need to review the AI generated text to remove excess language wherever possible. You can ask AI to reduce it or cut it down, but it doesn’t know what to take out or why.
Nice: AI knows how to be nice, but it is also weirdly, overly polite. You need to make sure the tone of the writing matches the actual relationship. Can you use “Hey” with this person or not? AI doesn’t know. Can you use emojis with this person or not? AI has no idea. Will the person read “Thank you” as nice or passive-aggressive? AI is clueless. You have to read the language and think about how it is going to sound to your audience.
Again, if you’ve been in class with me, you’ll know that these are strategies I teach. And if you haven’t, take a look at my online classes where you can learn this stuff!
AI can make us faster by doing the hardest part of writing for us: coming up with the rough draft. But it can only do that if we put in the right things as in the INPUT and correctly revising the OUTPUT to get the outcomes we want.