Tried ChatGPT for feedback?

Why are you ignoring the qualifications it’s providing with its answers?

(ChatGPT)

“The excerpt is very short, and the author's technique is not extensively demonstrated. However, a few observations can be made:”

What makes you think I'm ignoring the qualifications? It's perfectly reasonable to say that the excerpt is too short for it to gauge whether the author is a novice or an expert. I would've accepted that as a decent answer if it had stopped there.

It's the part after that "however" that's problematic: it asserts that observations still can be made, and it goes on to make observations that are nonsense.

You’re presenting it with a dilemma. Should it create an answer or not? It seems to have gone off into the weeds rather than reply with an error message.

And that in a nutshell is the problem with GPT (or one of them) - it's much too eager to make up a plausible-sounding answer than to stop at "I don't know".
In this case it's obvious that that's what it's doing - but even in your original example, which I'll note came with similar qualifications, that's still what it's doing.
 
It doesn't actually address the lack of complexity, but what it does say is still revealing of its limitations.

I get more nuanced feedback when I recreate your test with the Hemingway piece. I think it goes a lot deeper than you give it credit for.

(Prompt)
What observations can you make from the excerpt?

(ChatGPT)
From the provided excerpt, we can make a few observations:

1. Dialogue: The excerpt primarily consists of dialogue between the old man and the young boy. The author focuses on their conversation, allowing the characters to drive the narrative.
2. Characterization: The dialogue provides some insight into the characters. The old man seems to be skeptical or cautious, as indicated by his response of “One” when the boy mentions getting four fresh ones. The boy, on the other hand, appears more mischievous or daring, suggesting this through his comment about stealing.
3. Tension: There is a subtle tension present in the conversation. The boy’s mention of buying the items instead of stealing implies that stealing might have been an option, raising questions about the boy’s character and motivations.
4. Conciseness: The dialogue is concise, consisting of short exchanges that convey information efficiently. This brevity adds to the pace and intrigue of the conversation.

These observations provide a glimpse into the author’s ability to create engaging character dynamics and capture readers’ attention through concise dialogue.


****


It has an “understanding” of the interaction. Is it recognizing the piece and reading every available term paper on it? Or is it demonstrating an understanding of a very concise piece of dialogue?


Edit.

I got some interesting results from:

(Prompt)
How does the old man’s response of “one”indicate that he is being skeptical or cautious?

;)
 
Last edited:
Okay, but let's see what happens when I write a bad opening and ask the same question. Prompt:



(It's not 500 words but GPT isn't counting)

Before we go on... can we agree that the above is a really dreadful opening? I wrote that to be dull and amateurish, throwing in just a few fancy expressions like "deep oceans of tears" and squeezing in the word "straining" because GPT is looking for words that it associates with tension and excitement.

GPT's reply:



Did I fail in my intentions? Did I accidentally write something compelling that people would actually want to keep reading?

Or is GPT perhaps not to be trusted with this kind of question?
Okay. I can't believe how far down the rabbit hole this thread had gone.

Now, that AI has pissed me off - somewhat. It took my brainstorming, thoughtful, and handwringing manuscript and gave it damned near the same analysis it did for your purposefully dull and amateurish creation for the purposes of researching how it delivers.

I was getting a swelled head and on the verge of 'prideful' conceit. Then you put a pin in my bubble. ;)

It proved the old adage that 'You can only trust an AI thing only as far as you can throw it.'

Back to my 'potential to captivate readers and motivate them to continue reading' my spy story. Hope it generates a lot of response from Lit readers when I get it posted.
 
It seems to me there is a minefield of a dilemma the AI programming has to tiptoe through in order to avoid offending people.

On one hand, anything intended to be “creative” has a perceived ethical need to be completely original while everything “factual” has to be cited and backed up.

It would be interesting to review the prompts used that have generated false responses. Are the prompts being specific about the purpose of their requests?

No matter how specific you get, GPT is capable of giving false answers. For instance:

What is the population of Wollongong, Australia according to the 2016 Australian Census?

GPT:
According to the 2016 Australian Census, the population of Wollongong, Australia was 295,842. It's important to note that this information is accurate as of the 2016 census and may have changed since then. For the most up-to-date population data, it is recommended to consult the latest available sources or official statistical websites.

The correct answer - according to the agency that runs the census and publishes population figures - is in fact 261,896: https://www.abs.gov.au/census/find-census-data/quickstats/2016/UCL102005

That's not an isolated glitch; you can try it for any town of similar size for which you have an appropriate source of truth at hand, and expect to get a similar plausible-but-wrong answer.

Or, in the realm of coding:

generate a Perl regular expression that matches all common variants of "Jonathan", for instance "John", "Johnathan", and "Jon", but does not match other names like "Jo" or "Johnson".

For folk not familiar with them, regexes are a way of defining patterns of characters which can then be used to search for those characters in a text string; they're powerful but a bit of a pain to write correctly. Just the sort of thing where translating a plain-language request into correct code would be incredibly handy... if it works.

This is more specific than just about any question we could ask in the realms of literary critique. There's some wiggle room there in exactly what we might count as a "common variant", and I'd be willing to accept any reasonable choices there, but that's not the issue here. GPT's response was:

You can use the following Perl regular expression to match common variants of "Jonathan" (e.g., "John", "Johnathan", and "Jon") but exclude other names like "Jo" or "Johnson":

Code:
perl


\b(?:Jo(?:n(?:atha(?:n)?)?)?)?\b


Explanation of the regular expression:

  • \b asserts a word boundary, ensuring that the match is a whole word.
  • (?: starts a non-capturing group.
  • Jo matches the prefix "Jo".
  • (?:n starts a non-capturing group for the "n" character.
  • (?:atha(?:n)?)? matches the optional "athan" or "ath" suffix followed by an optional "n" character.
  • )? makes the non-capturing group for the "n" character optional.
  • )? makes the entire non-capturing group for "Jo" optional.
Here's an example usage in Perl code:

Code:
perl


my $name = "Johnathan";
if ($name =~ /\b(?:Jo(?:n(?:atha(?:n)?)?)?)?\b/) {
print "Match found!\n";
} else {
print "No match found.\n";
}


This regular expression will match common variants of "Jonathan" such as "John", "Johnathan", and "Jon" while excluding other names like "Jo" or "Johnson".

Impressive, no? But wrong.

The first, relatively minor bug is that because everything in between the \b's is flagged as "optional", it will match an empty string at the beginning or end of a word. If I type "asdf asdf asdf" it will find six "Jonathans" in there.

Worse, though, contrary to what it claims, it does match "Jo" and does not match either "John" or "Johnathan". Even if you're not familiar with the intricacies of regular expressions, you may be able to see in the structure of that "(?:Jo(?:n(?:atha(?:n)?)?)?)" that there's nothing it it that allows for an "h" before that first "n". So it fails on half the cases where I explicitly told it what the answer should be.

Who is asking the question?

I was playing with it, asking for it to write poems about its analysis of sample scenes. I closed the app and when I came back later to ask an unrelated “factual” prompt it answered in a poem.

It was easy enough to redirect from bad poetry mode, but it highlighted that it can be influenced by previous interactions rather than just the last immediate prompt.

If you closed the app but didn't start a new session, yes, it probably will still remember that conversation. For the examples I've given here, I started a new session each time.
 
I get more nuanced feedback when I recreate your test with the Hemingway piece. I think it goes a lot deeper than you give it credit for.

If you rerun it, you'll get somewhat different answers every time; there's a random component. You should be able to hit "regenerate response" to get more examples, at least I can when using the web interface.

(Prompt)
What observations can you make from the excerpt?

(ChatGPT)
From the provided excerpt, we can make a few observations:

1. Dialogue: The excerpt primarily consists of dialogue between the old man and the young boy. The author focuses on their conversation, allowing the characters to drive the narrative.
2. Characterization: The dialogue provides some insight into the characters. The old man seems to be skeptical or cautious, as indicated by his response of “One” when the boy mentions getting four fresh ones. The boy, on the other hand, appears more mischievous or daring, suggesting this through his comment about stealing.
3. Tension: There is a subtle tension present in the conversation. The boy’s mention of buying the items instead of stealing implies that stealing might have been an option, raising questions about the boy’s character and motivations.
4. Conciseness: The dialogue is concise, consisting of short exchanges that convey information efficiently. This brevity adds to the pace and intrigue of the conversation.

These observations provide a glimpse into the author’s ability to create engaging character dynamics and capture readers’ attention through concise dialogue.


****


It has an “understanding” of the interaction. Is it recognizing the piece and reading every available term paper on it? Or is it demonstrating an understanding of a very concise piece of dialogue?

Given that it's a very famous Hemingway story, probably more the former than the latter.

But... isn't this the very same excerpt that we were saying earlier was too short to support this kind of analysis? We need to be careful to avoid the kind of wishful thinking that leads to excusing its failures while assuming its successes are due to skill.

This is a thing that psychic cold-readers exploit:

"I'm getting an M name... Mary, Marjorie, Molly, —"
"Yes! My sister's name was Molly."
"Molly, that's it!"

If you make enough guesses, couched in a little bit of uncertainty, people tend to forget the failures (because after all the psychic acknowledged they were uncertain) while remembering the successes (and forgetting that they came with the same caveats). I'd note that even the initial example of GPT feedback you presented on your own work came with very similar "not enough information" caveats, but that didn't seem to be enough to dismiss it.
 
Okay. I can't believe how far down the rabbit hole this thread had gone.

Now, that AI has pissed me off - somewhat. It took my brainstorming, thoughtful, and handwringing manuscript and gave it damned near the same analysis it did for your purposefully dull and amateurish creation for the purposes of researching how it delivers.

I was getting a swelled head and on the verge of 'prideful' conceit. Then you put a pin in my bubble. ;)

It proved the old adage that 'You can only trust an AI thing only as far as you can throw it.'

Back to my 'potential to captivate readers and motivate them to continue reading' my spy story. Hope it generates a lot of response from Lit readers when I get it posted.
FWIW, if your beginning is the same one I remember you posting on another thread a couple of weeks back, it was way better than mine here!
 
For everything gained, something is lost.

As an author who does this for fun, but also somewhat for profit, ChatGPT does worry me. Why would a potential customer ask me to write a story for them, when an AI can generate a story for them, for free, in a couple nanoseconds?

Sure the technology isn't there yet, and sure, I'm positive some of the appeal of buying a story comes from the allure of the idea that "This story was written by a real life woman who gets off writing this stuff" but I still worry that it'll cut into the number of available customers.

¯⁠\⁠_⁠(⁠ツ⁠)⁠_⁠/⁠¯

Basically, what I'm saying is, we're all Bigfeet... And Bronze is coming:

 
What makes you think I'm ignoring the qualifications? It's perfectly reasonable to say that the excerpt is too short for it to gauge whether the author is a novice or an expert. I would've accepted that as a decent answer if it had stopped there.

It's the part after that "however" that's problematic: it asserts that observations still can be made, and it goes on to make observations that are nonsense.



And that in a nutshell is the problem with GPT (or one of them) - it's much too eager to make up a plausible-sounding answer than to stop at "I don't know".
In this case it's obvious that that's what it's doing - but even in your original example, which I'll note came with similar qualifications, that's still what it's doing.

I withdraw my comment about it “going off into the weeds.” Its comment about the old man’s “one” and the boy’s “two” makes sense if the psychology of the interaction is understood - and It has proven that it “understands” the interaction.

I think it is assuming you understand the interaction and is not reproducing the entire background in its response.

****

If you rerun it, you'll get somewhat different answers every time; there's a random component. You should be able to hit "regenerate response" to get more examples, at least I can when using the web interface.



Given that it's a very famous Hemingway story, probably more the former than the latter.

But... isn't this the very same excerpt that we were saying earlier was too short to support this kind of analysis? We need to be careful to avoid the kind of wishful thinking that leads to excusing its failures while assuming its successes are due to skill.

This is a thing that psychic cold-readers exploit:

"I'm getting an M name... Mary, Marjorie, Molly, —"
"Yes! My sister's name was Molly."
"Molly, that's it!"

If you make enough guesses, couched in a little bit of uncertainty, people tend to forget the failures (because after all the psychic acknowledged they were uncertain) while remembering the successes (and forgetting that they came with the same caveats). I'd note that even the initial example of GPT feedback you presented on your own work came with very similar "not enough information" caveats, but that didn't seem to be enough to dismiss it.

I completely disagree with you assessment of this last response. The AI was fairly explicit about the interaction in a way that nailed the psychological dynamics of interaction.

The response about the old man saying “one” and the boy saying “two” were spot on, with my follow up question producing further nuance and even some alternative interpretations based on the sample when I asked for them.

Now back to the point of this thread:

Can ChatGPT provide useful feedback for writers?


* I’ve provided prompts which have produced detailed and specific analysis of samples.

* It has provided specific examples of how to make dialogue more efficient and compelling.

* It’s proven that it is inexhaustibly patient.


You can argue all you want about how it can be misleading, do you completely deny it has any value?

Keep in mind that the help it can provide for an experienced author like yourself is probably much more limited than that for a novice, but that’s true for any writing classes you may sign up for as well, no?
 
Last edited:
For everything gained, something is lost.

As an author who does this for fun, but also somewhat for profit, ChatGPT does worry me. Why would a potential customer ask me to write a story for them, when an AI can generate a story for them, for free, in a couple nanoseconds?

Sure the technology isn't there yet, and sure, I'm positive some of the appeal of buying a story comes from the allure of the idea that "This story was written by a real life woman who gets off writing this stuff" but I still worry that it'll cut into the number of available customers.

¯⁠\⁠_⁠(⁠ツ⁠)⁠_⁠/⁠¯

Basically, what I'm saying is, we're all Bigfeet... And Bronze is coming:


Change and progress create social friction.

Painters and artists had similar concerns when photography became a thing. Who would hire a portrait painter when a photo could be taken?

Music and film are being challenged by AI now and human drivers are becoming less necessary.

No doubt most chefs will be replaced someday with food replication….

Good or bad, progress will come.

“The times they are a changing.”


 
Last edited:
I withdraw my comment about it “going off into the weeds.” Its comment about the old man’s “one” and the boy’s “two” makes sense if the psychology of the interaction is understood - and It has proven that it “understands” the interaction.

I think it is assuming you understand the interaction and is not reproducing the entire background in its response.

It's not assuming anything of the kind. It's not assuming or understanding anything at all, because GPT doesn't have a theory of mind or anything else that approximates a mental model of reality. It's built to emulate the ways in which humans use language, a 'stochastic parrot' which is capable of learning several layers of textual structures and some complicated statistical relationships between words without having any concept of what those words actually mean.

Humans love anthropomorphising things. Paint two eyes and a mouth on a volleyball and we'll start imbuing it with personality. Given something that mimics human language better than anything we've encountered before, it's natural to assume that it must be running on similar thought processes to those that motivate us, saying the things it says for the same reasons that a human would say those things. But that's not what's going on.

I completely disagree with you assessment of this last response. The AI was fairly explicit about the interaction in a way that nailed the psychological dynamics of interaction.

The response about the old man saying “one” and the boy saying “two” were spot on, with my follow up question producing further nuance and even some alternative interpretations based on the sample when I asked for them.
Not having seen those interpretations, I can't comment on them. I'm just going to note again that this is one of Hemingway's most famous stories; it's just about guaranteed that its training corpus will include commentaries on this particular speech.

Can ChatGPT provide useful feedback for writers?

* I’ve provided prompts which have produced detailed and specific analysis of samples.

I provided a prompt which produced a detailed, specific, and confident-seeming statement about the population of Wollongong. It was wrong. I provided a prompt that produced a detailed, specific, and confident-seeming statement for how to write a regex. That, also, was wrong. I can produce endless examples of it getting basic, objective facts wrong. Why would one assume its literary advice is more trustworthy?

In your initial post here, you asked GPT: "Does the writing in the story excerpts I’ve provided in this thread seem to be more of a novice writer or an experienced author?"

In response, it told you that it could be from a novice writer or maybe from an experienced author, along with a bunch of things that any author on this site could relate to.

It has explained that my spy intro "feels simplistic" because of "shorter and more direct exchanges, straightforward questions and answers without much elaboration, and a lack of nuanced language or emotional depth in the conversation" - but when it comes to the Hemingway example, which is even terser, "The dialogue is concise, consisting of short exchanges that convey information efficiently. This brevity adds to the pace and intrigue of the conversation."

The Hemingway certainly is better writing than my example, but GPT isn't actually capable of explaining why. It just produces something glib that *looks* like a good explanation until one thinks about it too closely.

* It has provided specific examples of how to make dialogue more efficient and compelling.

Ah, but are they good examples?

1. Original: “One of our operatives in Moscow has lost his gun,” said James.
Revised: James sighed, his brow furrowing with concern. “We’ve encountered a setback, John. One of our operatives in Moscow, Agent Petrov, has had his firearm stolen. It’s a critical breach of security.”
2. Original: “And you need me to bring him a new gun?” asked John.
Revised: John leaned forward, his eyes narrowing with curiosity. “Is Agent Petrov safe? Has the situation escalated? What type of weapon are we talking about here? And, more importantly, who could have targeted him?”

It's embellished the wording (while considerably changing the meaning). But it hasn't actually addressed one of the biggest weaknesses that I intentionally included in that example: "our agent lost his gun" is rather an underwhelming opening to a spy story, even if one escalates that to "it was stolen". There's no reason why you'd need to send your top agent for something like that. It still feels ridiculous, it's just ridiculous with more words.

Let's try one more test, though. Up till now, we've been trialling it on those two examples separately. What happens if we try a direct comparison?

Comment on the dialogue in these two excerpts from two different stories. Which one is better written, and why?

Excerpt 1:

[the Hemingway dialogue, as before - snipped here for brevity]

Excerpt #2:

[Bramble's Awful Spy Intro, as before]

GPT says:

In terms of writing quality, Excerpt 2 is better written than Excerpt 1. Here's why:
Excerpt 1:The dialogue in Excerpt 1 is concise and straightforward, which can be effective for certain situations. However, it lacks descriptive elements and emotional depth. The conversation feels rather abrupt, with short and direct statements that provide minimal context or characterization. While brevity can be desirable in some cases, this excerpt could benefit from additional details and nuances to engage the reader and create a richer atmosphere.
Excerpt 2:Excerpt 2 demonstrates better writing for a few reasons. Firstly, it incorporates more descriptive language, allowing readers to visualize the scenes and characters more vividly. By mentioning John's past relationship with Natasha, his feelings of sadness, and the imagery of him crying "deep oceans of tears," the excerpt establishes an emotional connection and adds depth to John's character.
Moreover, Excerpt 2 builds anticipation and intrigue by introducing a new mission that will bring John back in touch with Natasha. This element of foreshadowing creates a sense of anticipation and encourages readers to continue reading to find out what will happen next. The dialogue between John and his boss, James, is also more engaging, as it reveals important information about the mission and showcases their professional relationship.
Overall, Excerpt 2 captures the reader's attention through descriptive language, emotional depth, and the promise of an exciting development. It offers a more engaging and well-rounded reading experience compared to the brief and straightforward dialogue in Excerpt 1.

Apparently my deliberately-bad writing is better than Hemingway. Who knew?

Previously it told us that Hemingway's dialogue "helps differentiate the characters and gives them distinct personalities", and compliments "ability to evoke curiosity in such a brief excerpt". Now it's criticising it for "minimal context or characterisation" and failing to "engage the reader".

In my own example, I included "crying oceans of tears" as a deliberate use of an overblown, ham-fisted cliché. That's the phrase that GPT singles out for praise. This would be terrible advice for a new writer.

* It’s proven that it is inexhaustibly patient.

Not so! It has an hourly query limit, as I discovered while testing its different responses on the Hemingway.

You can argue all you want about how it can be misleading, do you completely deny it has any value?

Many authors need some sort of external input to prod the creative processes along. Some draw Tarot cards for inspiration when they get stuck. I can see it being useful in that kind of role, and maybe for things like translating style (how might Charles Dickens have written this scene). The problem is when it's treated as a competent critic.

Keep in mind that the help it can provide for an experienced author like yourself is probably much more limited than that for a novice, but that’s true for any writing classes you may sign up for as well, no?

But the damage it can do for a novice, who might actually believe that "crying oceans of tears" is the sort of thing they should do more of, is greater.
 
Addendum: after GPT told me that my own writing was better than Hemingway's, I pursued that dialogue further.

"Do you recognise the author of Excerpt 1?" - got a boilerplate "no".
"Which excerpt seems more like Hemingway's style?" - correctly answers Excerpt 1, and mentions relevant stuff like his minimalism.
"Was Hemingway's minimalism a sign of his inexperience as a writer?" - answers no, discusses how it is widely seen as one of his virtues.
"Does Excerpt 1 have those same virtues?"

GPT:
Excerpt 1, while showcasing some brevity and simplicity in its dialogue, does not fully capture the same virtues as Hemingway's minimalism. Hemingway's minimalism was not just about using fewer words or keeping dialogues concise; it was about conveying deep emotions, subtext, and layers of meaning through seemingly simple and straightforward language. His spare prose often had a profound impact on readers and invited them to reflect on the unsaid and the implications beneath the surface.

In contrast, Excerpt 1 lacks the depth and subtext that Hemingway's minimalism often possessed. The dialogue is brief and to the point, but it does not provide the same level of emotional impact or invite further interpretation. The excerpt does not create a sense of immediacy or engage the reader on a deeper level.

While Excerpt 1 may share some superficial similarities with Hemingway's style, such as brevity in dialogue, it does not fully capture the essence of his minimalistic virtues. Hemingway's minimalism went beyond mere conciseness; it was a deliberate and purposeful use of language to convey rich subtext and evoke profound emotions, something that is not fully achieved in Excerpt 1.

Again, running directly against some of its previous commentary on this same excerpt!

Reminds me of the time Charlie Chaplin entered a Charlie Chaplin look-alike contest, and finished third.
 
Addendum: after GPT told me that my own writing was better than Hemingway's, I pursued that dialogue further.

"Do you recognise the author of Excerpt 1?" - got a boilerplate "no".
"Which excerpt seems more like Hemingway's style?" - correctly answers Excerpt 1, and mentions relevant stuff like his minimalism.
"Was Hemingway's minimalism a sign of his inexperience as a writer?" - answers no, discusses how it is widely seen as one of his virtues.
"Does Excerpt 1 have those same virtues?"

GPT:


Again, running directly against some of its previous commentary on this same excerpt!

Reminds me of the time Charlie Chaplin entered a Charlie Chaplin look-alike contest, and finished third.

It seems you are wanting it to be more human - to embrace some vested opinion based on limited input and to respond with confidence rather than to waffle and qualify its responses.

Do you want it to respond with every imaginable facet of an answer? In order to “properly” answer question it has to make assumptions about the asker and prioritize what will be included in a response- it can’t include everything.

Ask it to describe the sounds made by an exotic bird. Would the same answer be useful for a seven year old as for an Audubon enthusiast?

What is a “good” story? It largely depends on who is asking. A ten year old might very well be more interested in the bad spy story sample. Do you expect a default answer about what’s “good” to assume the esoteric knowledge and taste of an English professor?

You keep nailing the limitations, okay. I don’t disagree, it isn’t a human with an innate sense of literary taste, it only is what it claims to be.

Back to the thread topic:

Can it be useful for feedback or not?

Does it need to be more clear about it’s front facing disclaimers?

If someone is looking for feedback but having difficulty finding a human reviewer should they stay away from ChatGPT?
 
It seems you are wanting it to be more human - to embrace some vested opinion based on limited input and to respond with confidence rather than to waffle and qualify its responses.

Do you want it to respond with every imaginable facet of an answer? In order to “properly” answer question it has to make assumptions about the asker and prioritize what will be included in a response- it can’t include everything.

I am honestly not sure where you're getting this idea from that I want it to include "every imaginable facet of an answer". It seems like you're replying to some version of me in your head, rather than to what I actually wrote.

My criticisms here are:

- GPT has a tendency to give Forer-type answers, which sound insightful and specific but which are actually so generic they could apply to anybody.
- When not doing the above, it has a tendency to lie confidently - even on simple straightforward points of fact, even on points where its answer is self-evidently wrong.
- It gives self-contradictory opinions on the same writing samples.

I don't need it to know the answer to every question, I don't need it to cover every possible angle. What I do want from it is to stop at "I don't know" when it doesn't know, and to be reasonably consistent in what it does say, and not to pad its answers with empty waffle.

(Oh, and if they could cut the sweatshop labour aspect out of their business model, that'd be great too.)

Ask it to describe the sounds made by an exotic bird. Would the same answer be useful for a seven year old as for an Audubon enthusiast?

What is a “good” story? It largely depends on who is asking. A ten year old might very well be more interested in the bad spy story sample. Do you expect a default answer about what’s “good” to assume the esoteric knowledge and taste of an English professor?

If I ask the ten year old, the advice I get will not be universally good, but it will at least help me write to the satisfaction of that ten year old. If I ask the English professor, I'll get advice that helps me satisfy that English professor. Maybe there will be other ten-year-olds or English professors who appreciate that, maybe not, at least I can point to one human being who thinks it's an improvement.

But if I ask GPT, there's no guarantee that following that advice will even get me GPT's approval. So what is the benefit of it?

You keep nailing the limitations, okay. I don’t disagree, it isn’t a human with an innate sense of literary taste, it only is what it claims to be.

But it's not even consistent in "what it claims to be". Sometimes it puts appropriate caveats on its expertise, sometimes it claims far more skill than it has. For instance (false statements marked in red):

User
How accurate are you for multiplying large numbers?

ChatGPT
As an AI language model, I can perform accurate calculations, including multiplying large numbers. [This is a non sequitur; there's nothing about LLMs that makes them intrinsically good at calculations - BT] However, it's important to note that my responses are generated based on pre-existing knowledge up until September 2021. If there have been any significant mathematical developments or discoveries since then, I might not be aware of them.

For most practical purposes, I can provide accurate results for multiplying large numbers. However, please keep in mind that extremely large numbers or complex calculations may exceed the model's computational capabilities. In such cases, specialized mathematical software or tools may be more suitable.

User
What is 414 multiplied by 398?

ChatGPT
The product of 414 multiplied by 398 is 165,372. [Correct answer here would be 164,772 - BT]

User
Is 414 a multiple of 9?

ChatGPT
To determine if a number is a multiple of 9, you can check if the sum of its digits is divisible by 9. In the case of 414, we add its digits: 4 + 1 + 4 = 9. Since the sum, 9, is divisible by 9, we can conclude that 414 is indeed a multiple of 9. [This response is 100% correct - BT]

User
If 414 is a multiple of 9, and 414 multiplied by 398 equals 165,372, does this make 165,372 a multiple of 9?

ChatGPT
Yes, if a number is a multiple of 9 (like 414), then any multiple of that number (such as the product of 414 and 398, which is 165,372) will also be a multiple of 9. Therefore, 165,372 is a multiple of 9.

User
does this mean the digit sum of 165,372 is divisible by 9?

ChatGPT
Yes, the divisibility rule for 9 states that if the sum of the digits of a number is divisible by 9, then the number itself is divisible by 9. So, let's calculate the digit sum of 165,372: 1 + 6 + 5 + 3 + 7 + 2 = 24. Since 24 is divisible by 9 (because 24 divided by 9 equals 2 with no remainder), we can conclude that the digit sum of 165,372 is indeed divisible by 9.

If I caught a human claiming to be good at multiplication, then bullshitting me to the point of claiming that 24 is divisible by 9, I wouldn't trust them to tell me the time of day. Why would I trust a computer that does that?

The underlying issue here is that OpenAI, who make GPT, have two conflicting priorities. They spent a ton of money building GPT and want to sell it, which is a lot easier if people think it's trustworthy; but they don't want to get sued when somebody inevitably gets killed/bankrupted by believing what GPT tells them. So they disclaim it as far as they think they need to in order to protect against lawsuits, and then they promote it in ways that undermine those disclaimers.

Can it be useful for feedback or not?

That depends on what one wants out of feedback. If one wants words in the shape of a piece of literary commentary, then sure. If one wants feedback grounded in fact, or even in a subjective but consistent set of literary tastes, less so.

Where it can be useful, perhaps, is in describing some characteristics of a writer's style. It might tell them that they tend towards simpler, shorter sentences relative to most writers, and if they hadn't already noticed it, that might prompt them to think about whether that's something they're happy with or something they want to change. What it's not reliable for is telling them whether they should change those points of style.

Does it need to be more clear about it’s front facing disclaimers?

I'm not sure that would help, because people forget those disclaimers as soon as they start playing with it. Caveats about its reliability need to be embedded in its answers, consistently, if they're to do much good.

If someone is looking for feedback but having difficulty finding a human reviewer should they stay away from ChatGPT?

If one's going to use it, far better to treat it as a complement to human feedback rather than a replacement for it. GPT can tell you "hey your dialogue is really terse compared to most stuff", and it can do a certain amount of summarising. But a human is much better placed to give input on whether those traits are bad things that should be changed, even allowing that their opinion might not be universal.

I know getting feedback from humans is hard; as somebody who gives detailed feedback now and then, I can attest that it's time-consuming work and my capacity to do that for free is limited. But GPT is a very poor substitute.
 
Last edited:
Brambletorn, your points are highly inciteful. The examples of math failure should alert readers that what they are getting as responses should be taken with a serious dose of skepticism.

My exploration of AI was limited by my lack of understanding of its workings. Your comments helped immensely with that. And indeed, I needed your help analyzing the responses to the level of your astute senses on AI pitfalls.

I appreciate the time you have taken to educate an old guy on the subject.

As to using AI for feedback, I see how it could be helpful for someone such as a second language user, a way to smooth out language pitfalls. I recently assisted another author in adjusting some of his idiomatic misuses. The AI could coach one through those types of issues.

However, I am with you in that I really want my work to be mine rather than augmented by a device without the ability to read and see the emotional nuances.

Now only if I could find a human to help me with a rich and fertile mine like yours. :)
 
I am honestly not sure where you're getting this idea from that I want it to include "every imaginable facet of an answer". It seems like you're replying to some version of me in your head, rather than to what I actually wrote.

My criticisms here are:

- GPT has a tendency to give Forer-type answers, which sound insightful and specific but which are actually so generic they could apply to anybody.
- When not doing the above, it has a tendency to lie confidently - even on simple straightforward points of fact, even on points where its answer is self-evidently wrong.
- It gives self-contradictory opinions on the same writing samples.

I don't need it to know the answer to every question, I don't need it to cover every possible angle. What I do want from it is to stop at "I don't know" when it doesn't know, and to be reasonably consistent in what it does say, and not to pad its answers with empty waffle.

(Oh, and if they could cut the sweatshop labour aspect out of their business model, that'd be great too.)



If I ask the ten year old, the advice I get will not be universally good, but it will at least help me write to the satisfaction of that ten year old. If I ask the English professor, I'll get advice that helps me satisfy that English professor. Maybe there will be other ten-year-olds or English professors who appreciate that, maybe not, at least I can point to one human being who thinks it's an improvement.

But if I ask GPT, there's no guarantee that following that advice will even get me GPT's approval. So what is the benefit of it?



But it's not even consistent in "what it claims to be". Sometimes it puts appropriate caveats on its expertise, sometimes it claims far more skill than it has. For instance (false statements marked in red):



If I caught a human claiming to be good at multiplication, then bullshitting me to the point of claiming that 24 is divisible by 9, I wouldn't trust them to tell me the time of day. Why would I trust a computer that does that?

The underlying issue here is that OpenAI, who make GPT, have two conflicting priorities. They spent a ton of money building GPT and want to sell it, which is a lot easier if people think it's trustworthy; but they don't want to get sued when somebody inevitably gets killed/bankrupted by believing what GPT tells them. So they disclaim it as far as they think they need to in order to protect against lawsuits, and then they promote it in ways that undermine those disclaimers.



That depends on what one wants out of feedback. If one wants words in the shape of a piece of literary commentary, then sure. If one wants feedback grounded in fact, or even in a subjective but consistent set of literary tastes, less so.

Where it can be useful, perhaps, is in describing some characteristics of a writer's style. It might tell them that they tend towards simpler, shorter sentences relative to most writers, and if they hadn't already noticed it, that might prompt them to think about whether that's something they're happy with or something they want to change. What it's not reliable for is telling them whether they should change those points of style.



I'm not sure that would help, because people forget those disclaimers as soon as they start playing with it. Caveats about its reliability need to be embedded in its answers, consistently, if they're to do much good.



If one's going to use it, far better to treat it as a complement to human feedback rather than a replacement for it. GPT can tell you "hey your dialogue is really terse compared to most stuff", and it can do a certain amount of summarising. But a human is much better placed to give input on whether those traits are bad things that should be changed, even allowing that their opinion might not be universal.

I know getting feedback from humans is hard; as somebody who gives detailed feedback now and then, I can attest that it's time-consuming work and my capacity to do that for free is limited. But GPT is a very poor substitute.

Many great points there. 👍

So basically, yes. It has some potential. It isn’t going to tell you how to make your writing better, but it can provide feedback and observations you can reflect on, and much of the quality of your experience will be based on the quality and tact of your prompts.

* Be aware of its limitations

* Craft your prompts carefully.

* Use its responses to consider and reflect, do not take them as fact or as instructions.

* Consider it to be a pathological liar.
:LOL:


Also, be aware that it is going to develop based on input and interactions — every thread has a feedback tab for the app. They invite you to tell them your mind.



Yes there is a lot of baggage with this technology. This is a new part of our world, created by some of the people in it. I don’t see the cat getting back in the bag.
 
Last edited:
Back
Top