Here’s Why ChatGPT Can’t Replace Technical Writers
For starters — It doesn’t understand how to crochet
Instructional design and technical writing are two halves of the same coin, guiding the user through how a specific product functions. Part of the editing process involves a writer envisioning themselves as the end-user. They look for any pitfall or confusing wordage that might trip up a novice user.
With the rise of AI, every company is experimenting with the implementation of AI for writing positions. — Why pay a writer for something an AI can create for free? Because not all writing is the same.
An AI can’t become a technical writer because it doesn’t have the physical and social aspects that are critical to a technical writer’s success. Here are a few issues with using AI for technical/procedural writing:
- Half of a technical writer’s role involves interviewing and talking to Subject Matter Experts (SMEs) about a specific product or service.
- ChatGPT’s output is instant based on a prompt, requiring someone to craft a detailed initial prompt to even leverage AI tools to write technical writing instructions.
- AI can’t apply logic to a specific set of instructions to find issues with real-world applications. Any instructions generated will be generic and created based on other instructional material.
As a point of testing, I’ve decided to have ChatGPT create real-world instructions for a crocheted hat. Crochet was selected as it’s something I have expertise in and can analyze the output for accuracy compared to similar patterns.
What is Crochet?
Crochet is a yarn craft that involves holding a crochet hook with one hand and yarn with another. By manipulating the hook through the yarn, hand-crafted items are made like clothing or home goods. This is different than knitting because crocheting usually involves working on one stitch at a time with a hook, while knitting involves working on many rows at once with two knitting needles.
Patterns are written by experienced crocheters for others to follow. The creation process often involves crocheting an item multiple times, recording the stitches and number of rounds used, and then having others test the pattern for any problems that the creator didn’t realize. It’s a long process that is very precise.
For AI to replicate the pattern-making process, it would need to ignore physically creating an item and having people test it. — It misses the two critical parts of making a pattern because, regardless of how it looks written, it needs to translate into an actual tangible good.
Parameters:
To start, I prompted ChatGPT 3.5 to “Write a hat crochet pattern using double crochet.” ChatGPT provided me with material requirements, abbreviations used, and a 22-row instructional pattern. — All that a typical pattern would provide.
One important note is that it did not identify which crochet terms were used (US vs. UK). This is important because US and UK terms are different, often with different techniques using the same name. A “single crochet” in US terms is called a “double crochet” in UK terms, but the US terms call a different stitch “double crochet.” If these terms are mixed or the user is unaware of which to follow, it can lead to the entire project coming out very differently.
As well, it didn’t include a gauge or expected width of the hat. These are helpful inclusions so that someone following the project can know if their personal tension or technique is making the hat larger or smaller than how the pattern creator made the original hat. When someone crochets, they pull the yarn and, depending on how tight their tension is, it can make a stitch taller or shorter. This can severely impact the final size because of how tight or loose these stitches are. Plus, people have different sized-heads further necessitating a guide for how to accurately size the hat.
But that aside, the pattern roughly makes sense with the first half and a hat could be created from it.
First half — Pattern
Round 1: Start with a magic ring, ch 2 (counts as first dc here and throughout), work 12 dc into the ring. Join with a sl st to the first dc. (12 dc)
Round 2: Ch 2, 2 dc in each st around. Join with a sl st to the first dc. (24 dc)
Round 3: Ch 2, dc in next st, 2 dc in next st repeat around. Join with a sl st to the first dc. (36 dc)
Round 4: Ch 2, dc in next 2 sts, 2 dc in next st repeat around. Join with a sl st to the first dc. (48 dc)
Round 5: Ch 2, dc in next 3 sts, 2 dc in next st repeat around. Join with a sl st to the first dc. (60 dc)
Rounds 6–12: Ch 2, dc in each st around. Join with a sl st to the first dc. (60 dc)Note: Adjust the number of rounds in this section if you want the hat to be longer or shorter before starting the decrease rounds.
So far so good
The pattern told you the final number of stitches in each round, making it easy to stay on track. Without this, it can be difficult to tell if a stitch was missed or if an extra one was added. Mentioned adjusting the rounds based on how long you need the hat. This can help modify the hat if a more slouchy hat is wanted.
It’s a solid pattern for someone to follow so far. And while I can find some issues, as mentioned, it’s okay for a beginner hat or even a novice pattern maker. I wouldn’t be upset to find this pattern on someone’s website.
Second Half — Pattern
This is the part where things take a weird turn.
Decrease Rounds:
Round 13: Ch 1 (does not count as a st), sc in each st around. Join with a sl st to the first sc. (60 sc)
Round 14: Ch 1, sc in next 8 sts, sc2tog repeat around. Join with a sl st to the first sc. (54 sc)
Round 15: Ch 1, sc in next 7 sts, sc2tog repeat around. Join with a sl st to the first sc. (48 sc)
Round 16: Ch 1, sc in next 6 sts, sc2tog repeat around. Join with a sl st to the first sc. (42 sc)
Round 17: Ch 1, sc in next 5 sts, sc2tog repeat around. Join with a sl st to the first sc. (36 sc)
Round 18: Ch 1, sc in next 4 sts, sc2tog repeat around. Join with a sl st to the first sc. (30 sc)
Round 19: Ch 1, sc in next 3 sts, sc2tog repeat around. Join with a sl st to the first sc. (24 sc)
Round 20: Ch 1, sc in next 2 sts, sc2tog repeat around. Join with a sl st to the first sc. (18 sc)
Round 21: Ch 1, sc in next st, sc2tog repeat around. Join with a sl st to the first sc. (12 sc)
Round 22: Ch 1, sc2tog around. Join with a sl st to the first sc. (6 sc)Fasten off, leaving a long tail for closing the hole. Use the yarn needle to weave through the remaining stitches, pull tight to close the hole, and secure the end.
While the pattern started strong, this is no longer a crocheted hat pattern.
Final number of stitches
To start with, I have to mention the number of stitches in the hat. A magic circle is a technique used to start a circular crochet project without a large hole in the center. Like the diameter of a circle, a circular crochet project usually starts with 6–12 stitches in the magic circle (depending on the project) and slowly increases each round. A round is marked by completing a full circle around the initial magic circle of the crochet project.
For a hat, this stitch increase stops once the hat is wide enough for a head to fit. The rounds then continue, without an increase, for the sides of the hat. It never decreases again unless the hat is truly too wide.
A final hat with 6 stitches as the diameter, wouldn’t fit on someone’s head. You’d have to push the entire base into itself so the final diameter is from the sides, with 60 stitches around. This doesn’t make sense though because you’d lose half of the work you’ve done to even wear the hat. (Plus, as I’ll mention below, single-crochet stitches are shorter than double-crochet stitches so it might be even smaller than expected.)
Number of Rounds
The number of rounds is also too short, a hat I recently crocheted was 30 rounds and was a little small for my head (much longer than 20 in this pattern). While the pattern mentions to “adjust to length” before starting the decrease, it never mentions how many or any sort of guide. It is hard to gauge how many additional rows to create unless there is context to the statement. How does the decrease impact the length of the hat? — If too many are added before the decrease, it would cause the hat to be significantly longer than anticipated.
Type of Stitches
If you remember, my initial prompt asked for “a double-crochet stitch” hat. This was followed in the first part of the pattern, but was quickly scrapped in part two for single-crochet stitches. This is important as stitches look different. A single-crochet is half as tall as a double crochet and looks different. You aren’t supposed to just switch unless you are following a specific pattern.
I assume that ChatGPT got confused and mixed up a decrease with a hat brim. Some hat patterns use single-crochet for a hat brim because of its different look. This is done by first crocheting a hat base (for example using double-crochet stitches), and then making a long chain to attach to the bottom of the hat to fold over using single-crochet stitches. The stitch difference is visible, but can flow nicely because they are two different parts of the hat.
Final Object
It’s not a hat. Part of the problem with relying on ChatGPT to create a pattern is that it doesn’t have a real-world foundation. As mentioned, pattern makers spend hours upon hours finalizing and critiquing patterns so that they are easy to follow and make sense. ChatGPT never actually made this pattern after writing this pattern.
If made, this pattern would look closer to a cylinder than a hat. It starts with a circle, then multiple rows as tube-like, and ends with another circle.
Conclusion
It wrote a decent pattern in the first half, but because it doesn’t truly understand what a crochet pattern is, it generated what it expected the second half to be. ChatGPT generates text, it doesn’t think and understand the text, only the patterns associated with it.
The same goes for generating technical or procedural information. ChatGPT was and still is trained on mountains of procedural text from previous pieces of technical writing. It can write the instructions for how to change the oil in a car because many people already have detailed the process.
If a new type of car came out tomorrow which ran on water and needed that water replaced every 30 days, ChatGPT wouldn’t be able to write specific instructions because people haven’t already written them. Someone would need to describe the process to ChatGPT to even interpret first before generating instructions. While AI can generate or guess how to do something, it won’t always be the right way.
Only a human, who can investigate a process and ask questions, can truly write and finish a piece of technical or procedural writing.
