Categories
Technology Work

I asked ChatGPT to write this. Here’s what happened next.

The other day The Washington Post ran an article titled, ‘Type in your job to see how much AI will affect it.‘ So I did. First I typed in Delivery Driver, which changed to ‘Light Truck or Delivery Services Drivers’ from a pre-populated drop-down menu.

It then displayed an orange dot and a green dot, based on research about two types of artificial intelligence: language (orange) and images (green). Here’s where those dots appeared:

As you can see, with regard to language and images, my job as UPS driver is likely to be less impacted by artificial intelligence than the average job. (It’s not clear whether that average is within the broader category of ‘Production and transportation,’ or across all occupational categories.)

These results make sense to me. I already rely on smart technology – especially the algorithmic software that optimizes my daily route in two-dimensional space. UPS also makes extensive use of other technologies that help it plan and execute the movement of some 20 million units of work (‘UOW’) every day. That includes obtaining the three-dimensional measurements of each of those UOWs.

But technology alone cannot make critical decisions about how to prioritize my deliveries and pickups. For that I must use the considerable knowledge that I’ve accumulated about local context and emergent conditions. My ability to read addresses, directions, or traffic signs are not likely to be improved by language-based artificial intelligence.

Also, my job is intensively physical. Unless AI is embodied in a very sophisticated robot, delivery companies will still need humans to schlep their customers’ packages from origin to destination.

I conclude, then, that my vocation is probably pretty safe from the imminent rise of the machines.

But what about my avocation? What does the future hold for bloggers?

When I typed in ‘Writer,’ it changed to ‘Writers and Authors.’ The resulting graph looked like this:

Not surprisingly, the results for Writers and authors show up on the positive side of the graph. Image-based artificial intelligence is expected to have a higher-than-average impact on these jobs by a factor of more than one (whatever that means).

The corresponding orange dot did not appear, and I couldn’t conjure it for the life of me, despite hovering on every orange dot I could see on the graph. I have to believe that language-based AI will affect the jobs of Writers and authors at least as much as image-based AI, if not more so.

Regardless, the possibility of artificial intelligence impacting how I write blog posts makes intuitive sense to me.

Which made me wonder: what’s possible right now? Could I just hand my blog over to a smart machine and watch it generate content for me while I sit back and sip my soda? Could I type up a few talking points and feed them into ChatGPT and voila! ?

So I signed up for a ‘research version’ of ChatGPT, developed by OpenAI. The ‘stable-release’ version of this chatbot was released on May 12th; the basic model is free of charge.

After logging into my new account, I started my first chat:

At this early stage, the interaction feels just like those I’ve had with other chatbots I’ve interacted with at various companies’ websites. Like the one at Amtrak, which is rated by several websites as one of the best online chatbots.

And up to this point our conversation is lively and unproblematic. I’ve started to notice that it uses the same or similar words and phrases that I use, redirecting them back toward me as a way to return the volley.

This was news to me. I thought this software was being fed a steady diet of internet content. Or could be, if the situation warranted.

But! According to my conversation partner all was not lost.

At this point I pasted into the chat box quite a few paragraphs – twelve, to be exact – I’d copied from my most recent blog post. Which prompted this response:

And with that, ChatGPT returned to me a lightly revised version of those same twelve paragraphs. Here’s a side-by-side comparison of the first portion:

MY ORIGINAL TEXTChatGPT’s REVISED TEXT
Recently, in the space of about a week, the news media have reported on no fewer than four separate incidents where an honest mistake or innocent misunderstanding led to gunfire.In recent times, the media has been filled with unsettling reports of incidents where innocent mistakes or misunderstandings have resulted in tragic encounters involving gunfire.
In Kansas City, a sixteen-year-old boy rang the doorbell of a house where he was sent to pick up his younger siblings, except he was on the wrong street. Despite being shot in the head by the 84-year-old homeowner, he miraculously survived.One incident took place in Kansas City when a sixteen-year-old boy, who had mistakenly arrived at the wrong house while trying to pick up his younger siblings, was shot in the head by the 84-year-old homeowner. Miraculously, he survived despite the grave injury inflicted upon him.
Just outside Austin, Texas, two teenage girls attempted to enter the wrong vehicle in a supermarket parking lot. Both were shot by a 25-year-old man in the passenger seat, with one suffering serious injuries; thankfully, she’s expected to recover.Another distressing occurrence unfolded just outside Austin, Texas, where two teenage girls inadvertently tried to enter the wrong vehicle in a supermarket parking lot. Unfortunately, their mistaken action led to both of them being shot by a 25-year-old man seated in the passenger seat. While one of the girls suffered serious injuries, fortunately, she is expected to recover.

It went on like that through each of the remaining paragraphs. The only differences were in word selection or phrasing. The overall structure of paragraphs and sentences was exactly identical.

I saw this as Strike Two. Strike One had been ChatGPT’s inability to access my blog directly and get a sense of my style and ‘voice’ without my having to force-feed it. Now I saw what ChatGPT had done as nothing more than a revision – not even an edit.

What’s more, at no point did I see this revision as an improvement on the original. It was just a variation, and a wordier one at that. Another swing and a miss. But it was too soon for me to give up. I tried to regroup:

I was not deterred in my optimism. Nor apparently was my partner:

With that, I was ready to launch into my main objective:

I’m a pretty fast typist. I type at least 60 words per minute when I’m transcribing basic text; I can probably type about 30 wpm when I’m composing in real time.

ChatGPT easily had me beat. Its response times were typically just a few seconds, even for a response that went on for several paragraphs.

But even though this conversation unfolded in just a few minutes, I had to pause to attend to other parts of my life. I wondered if that would be a problem.

That was yesterday. What follows is from today. I didn’t start a new chat, just entered more text in the chat I’d started before:

That’s right, dear reader. Somehow ChatGPT had tricked me into already doing much of the work myself.

At this point I simply copied and pasted everything I’d composed in the blog post so far, which is everything you’ve just read through. But since my text was broken up into segments between screenshots, I had to enter it in discrete chunks of varying length.

This was a problem. ChatGPT felt compelled to digest each chunk separately, rather than waiting for all of them before weighing in.

For example, when I fed ChatGPT the single introductory paragraph before my first screenshot, it offered this feedback:

By contrast, for the seven-paragraph segment that appears after the first couple screenshots, it lapsed into what I characterize as ‘research-paper mode.’ It was no longer responding directly to what I’d written; now it was seeking to expand upon it, using a style that was inconsistent with my voice or the tone of the post.

ChatGPT was nothing if not completely transparent about its own limitations!

As for me, I kept my head down. I just copied and pasted all the text I’d written, ignoring my screenshots as well as its responses to the separate segments until after I was all done. Only then did I review the chat log to see what all had transpired.

What I discovered was a letdown. The responses to each segment in turn became less and less helpful to me.

As I was bringing ChatGPT current on the work I’d done thus far, I was no longer asking it to reflect on new, outside information. Now I was feeding back to it the very conversation we’d had only yesterday. Unsurprisingly, then, it was giving me the exact same or very similar responses as it did the first time around. Like a hall of mirrors. Or a dog chasing its tail.

Still, its enthusiastic support never wavered. Its willingness and ability to serve as my sounding board and conversation partner were inexhaustible. Always eager to please, ChatGPT offered this mea culpa:

At this point I felt like ChatGPT had brought me full-circle, and there was nothing more I could expect from it. Strike Three, I’m Out.

As I reflected on this experience with language-based artificial intelligence, I initially felt disappointed. I thought I could set ChatGPT up with a few instructions and guidelines and let it do its thing. Like putting some oats and water in a bowl, setting it in the microwave and pressing ‘START.’ Even if I had to stop and stir every so often, eventually – soon! – I’d have my oatmeal.

But why did I feel this way? After all, I’m finishing up a post that I started just today. I have never – not even once – composed a blog post in a single day. It usually takes me several days, or weeks, or even longer. I still have drafts of unpublished posts that I’ve started several months or years ago.

What’s more, this post takes the very shape that ChatGPT suggested at almost the very beginning. Recall:

‘We can begin with an introduction that sets the stage for the experiment and highlights your motivation for seeking assistance from an AI-language model.’

Check.

‘Then we can delve into the process, describing how you interacted with ChatGPT, the ideas and prompts you provided, and how we collaboratively shaped the content.’

Check.

‘Throughout the post, we can incorporate elements of your actual conversation with ChatGPT to add an authentic touch to the narrative.’

Check.

‘Additionally, you can include your thoughts and reflections on the collaboration process and the final outcome of the post.’

Check. And mate.

How did it know?

Or – maybe – how did I not know?

ChatGPT had helped me more than I’d wanted to give it credit for. Sure, it had put me through my paces. Maybe I needed to learn as much about it as it needed to learn about me? Maybe it was – had to be – an iterative process for both of us?

If so, then it seemed only fair to let ChatGPT have one more look at what I’d written, and one final opportunity to provide feedback:

You’ll notice that this time I’m much more mindful about prepping ChatGPT for what’s to come, and more direct and explicit in what I’m asking from it. I’ve learned from our previous interactions.

I copy all of the text segments and paste them into a single blank Word document. Then I copy all the stitched-together segments and paste them back into the chat box. In a matter of seconds, this appears:

Happy writing indeed.

3 replies on “I asked ChatGPT to write this. Here’s what happened next.”

So, ChatGPT from my perspective does resemble the reference picture from Terminator — precise, functional and devoid of flesh and with metal bones. But still kind of intriguing to look at.

I’m grateful that you have job security and blog security.

And I think you may have taught the chatGPT a few things…

Thanks Martin!!

Working closely with AI like this would have me screaming hysterically and running out of the room. In short, it would drive me nuts (but it would be a very short drive!) Stick to blogging by yourself. It will be more productive.

Yeah the more I play with the ChatGPT and other similar products, the more I’m convinced that the technology just isn’t there quite yet. Either that, or I haven’t figured out how to use it effectively – which is a very real possibility! Thanks for reading and commenting Mom!

Leave a Reply