professional writing & creative consulting
Back in June 2024 Christopher Mims of the Wall Street Journal interviewed Dave about AI and B2B copywriting as part of a long piece about the impact of AI on freelance artists and writers. It’s especially appreciated that WSJ chose to use the promo headshot from Dave’s first book, when he was a much younger man:
For archival purposes, here’s a moderately ugly PDF of the article as it appeared in June 2024.
There are a lot of reasonable ethical concerns about A.I. in the classroom. So let’s focus on AI outside of the classroom. There are so many tasks that you, as a teacher, are asked to do that have little to do with teaching. A.I. can help you spend less time and effort on these, leaving you more time to focus on the classroom, your students, and the things that genuinely can only be done by you.
A brief reality check: just because we’re keeping A.I. outside the classroom doesn’t mean we’re magically making every risk and worry go away. This is new technology, and every new technology has had unforeseen consequences and downsides.
Two tips to help you avoid the clearest pitfalls we’re seeing with A.I. right now:
Now that we’ve laid some ground rules, let’s dig in.
There are lots of different “A.I.”s floating around out there, many embedded in products you already use (like Google Translate and the predictive text/autocorrect in your phone’s messaging app). But most of the time if folks are talking about “A.I.” right now, they mean “generative AI” or “gen A.I.” like ChatGPT (which generates written answers to prompts) or DALL-E (which generates images based on user prompts).
Text-generating “A.I.” like ChatGPT are built on Large-Language Models (or LLMs). An LLM is a statistical model created by digesting an absurd amount of writing (mostly the Internet in general—especially sites like Wikipedia, Reddit, and online news sources—plus a whole lot of digitized books). The result is a model of what people think, based on what they’ve said over the last several decades online (plus what people wrote in books for several hundred years before the Internet).
At its core, all an LLM does is predict some likely possibilities for the next word or phrase in a series. You give it “Jingle bells…” and it’s likely to return “…jingle bells, jingle all the way.” But it could also return “…Batman smells, Joker got away.” (A valid, if slightly lower probability response.) Or “…is a popular seasonal song.” It could even say something crazy, like “…are highly acidic and a good source of thiamine.” That’s not true, and doesn’t make much sense, but it’s grammatical and within the realm of a possible answer—just not a good one.
As the complexity of what you offer it (the “prompt”) grows, so can the complexity of the LLM’s outputs. (And likewise the probability of it saying something crazy or factually incorrect—a so-called “hallucination.”)
Search is about facts; LLMs are about vibes.
This is fundamentally different than what you are used to with computers. Until now, most applications you used were programs written as a series of commands. Those commands executed specific functions in response to specific inputs from you. The result was consistent outputs. Ask Siri the time, and it answers with the current time in your timezone.
The LLM underlying an A.I. like ChatGPT isn’t programmed. It contains no commands or rules or functions. The exact same input can result in wildly different outputs—just like asking two different people their opinion can get you wildly different answers. You can’t trust an LLM to give you facts or figures, or to follow instructions.
At first glance, that seems both annoying and useless: I google something because I actually want to know what time the sun will rise on March 1st or what year George W. Bush was born. I don’t want opinions; I want facts!
So here’s a good way to think about it:
Search engines (like Google) are built on programs that allow you to search the Internet for facts; LLMs (like ChatGPT) are built on models that let you converse with the Internet.
Search is good for fact-based activities (for example, looking up the sunrise and sunset times for the first day of each month in a given year). LLMs are awful at that task, but are great at helping you get a handle on how to nicely, politely, and professionally word a note whose gist is “your son is failing math because he mostly skips class and sleeps when he is here.”
“Prompt engineering” is the art of “talking” to an LLM so that it helps you get something done.
There are a lot of A.I. tools out there, with more (of wildly varying quality) hitting the market every single day.
But all you need to get started is a free OpenAI account. You can set this up right now:
If this is feeling overwhelming (most of us need one more account like we need one more hole in the head) don’t jump ship just yet! I’ve included a screencast of me running through the example below. Scroll to the bottom of this post, and you can follow along without any sort of commitment.
Your day includes many repetitive or tedious chores that eat up time (often your personal time) but have little classroom or educational benefit. Think weekly check-ins and newsletters, progress reports, emails, cookie-cutter lesson plans, and a whole slew of record-keeping / “paperwork” / “box-checking” tasks—the sorts of activities where a list of scores or a one-sentence summary would likely suffice, but isn’t considered sufficient. These also tend to be situations where your “tone” or “vibe” is extremely likely to be picked apart by an unhappy recipient—compounded by the fact that you are forced to get this done outside of school hours, often after a full day or week in the classroom.
It is, at best, a fraught slog. Many teachers resent these chores with good reason (I was once one of you: I taught for seven years before my son was born and I started working from home).
LLMs can help you get these tasks done more quickly with less stress. Here’s how:
Let’s imagine you need to write a one-page letter of recommendation.
Start by going over to ChatGPT: https://chat.openai.com/
Once you sign in, you’ll see a text box inviting you to “Message ChatGPT.” This is where you enter your prompt.
The golden rule of prompt engineering is:
“Collaborate, don’t command.”
If you tell ChatGPT “Write a letter of recommendation” it will spit out a perfectly formed, perfectly generic, perfectly useless letter of recommendation template.
Instead, try this prompt:
"I need help writing a letter of recommendation for one of my students. What do you need to know to help me?"
ChatGPT will reply with a list of questions.
Copy ChatGPT’s questions into an empty document, answer each, then paste those answers into the message bog as a reply.
(It’s ok to skip some of the questions; ChatGPT tends towards being both exhaustive and exhausting.)
Hit enter and see what it comes up with.
Don’t like what ChatGPT offered? Tell it so. Ask it to try again, with a prompt like:
“Try again, but be brief.”
or
“This is close, but I’d like it to sound more enthusiastic. Try again.”
You can continue this back-and-forth as long as you like, but don’t aim for perfect. Remember: our goal is to save time on tasks that aren’t at the heart if your work. Just get ChatGPT to draft a usable rough draft. Copy it, plop it into a fresh Word document, and clean it up yourself.
Here’s a three minute screencast of me using ChatGPT to write a letter of recommendation for Ms. Marvel:
Interested in learning more about how regular folks can make ethical use of A.I.? I’ll give you a heads up with new tips, tricks, and tools as I come across them.
This post is Just a Test .