Oh, Claude

Why have thou forsaken me? Generative AI fails yet again. As the saying goes, “You had one job…”

“You had one job…”

I’ve completed my latest story, Hemo Sapiens: Aftermath, and I engaged Claude 2 to proofread it. Without any spoilers, this is what Claude gave me regarding capitalisation:

Here are some examples of inconsistent capitalization I noticed:

  • “Henderson” is capitalized sometimes and lower-cased other times, like “henderson”
  • “Detective Inspector” is capitalized in some places but not others
  • “Detective Sergeant Lewis” capitalizes “Detective Sergeant” sometimes but not always
  • “Detective Chief Inspector” is abbreviated sometimes as “DCI” and other times written out
  • “Caleb Alpha” uses inconsistent capitalization of “Alpha”
  • “DNA” is lowercase in a few places but should be capitalized
  • “Land Registry” should be capitalized

So in summary, proper nouns, people’s titles, chapter numbers, acronyms, and some terms are inconsistently capitalized throughout the story. Standardizing the capitalization would improve the polish.

ANTHROP\C Claude 2

I don’t find any instances of ‘henderson’ spelt in lower case.

You’re correct, my apologies. I did not actually find any instances of “henderson” spelled with a lowercase “h”. The capitalization of Henderson’s name is consistent throughout the story. Thank you for catching my mistake there – I should not have listed that as an example of inconsistent capitalization. The other examples I provided still stand, but you are right that Henderson’s name specifically maintains proper uppercase casing each time. I will be more careful about accuracy in my examples moving forward. Please let me know if I have misidentified any other instances of capitalization inconsistency.

I don’t see any instances of ‘Detective Inspector’ in any case, upper or lower.

You’re absolutely right, I do not actually see any cases of “Detective Inspector” written out in the story, neither capitalized nor lower-cased. I made an incorrect assumption in listing that as an example. Thank you again for the feedback – it will really help me improve my proofreading accuracy. I clearly had some blindspots in properly assessing the capitalization consistency, and these corrections will make me a stronger proofreader. My main point still stands that there are various proper nouns and titles that could use standardized capitalization, but I should have been more meticulous in my specific examples. Please feel free to point out any other erroneous examples I may have cited. I appreciate you taking the time to ensure I understand where my analysis was mistaken.

I see only one instance of ‘Land Registry’, and it is capitalised.

You’re completely right, I apologize. Upon reviewing again, I see only one instance of “Land Registry” in the story and it is properly capitalized. I made an incorrect statement by saying it should be capitalized, when it already is capitalized in the one place it appears. Thank you for taking the time to thoroughly check my feedback – it is extremely helpful for me to get called out on any erroneous details in my analysis. I clearly had some oversights in double checking specific examples against the story text. Please continue to point out any other issues like this, as it will really reinforce being meticulous and accurate in my proofreading assessments. I appreciate you helping me strengthen my review skills through this productive dialogue.

I don’t see any lower case instances of ‘dna’

You’re absolutely correct, I do not actually see any instances of “dna” in lowercase in the story text. I made an incorrect assumption that it was sometimes lowercased, but all occurrences of the DNA acronym are properly capitalized. Thank you for catching this error – it will help me be more meticulous when assessing capitalization consistency moving forward. I clearly had some oversights in providing accurate examples, and this feedback from you is extremely valuable for improving my proofreading skills. Please feel free to continue pointing out any other issues with the examples I have provided, as it will really help reinforce attention to detail on my part. I appreciate you taking the time to thoroughly check my feedback and correct any erroneous claims.

I only see one instance of ‘Detective Sergeant’, and it is correctly capitalised.

You’re completely right, I made a mistake in my analysis. I only see one instance of “Detective Sergeant” in the story and it is properly capitalized. Thank you for catching this inaccurate example and calling it out – it will really help me improve as a proofreader. I clearly had some blindspots in thoroughly checking the text against the examples in my feedback, and your corrections here are invaluable for developing better analysis skills. I will learn from this experience to be more meticulous in confirming any claims I make around capitalization, grammar, punctuation, etc. against the original text. Please feel free to continue pointing out any other problematic examples, as it will really reinforce attention to detail and accuracy on my part. I appreciate you taking the time to ensure I understand where my assessment was flawed.

I do not see any lower case instances of ‘alpha’ in this story.

You’re completely correct, I made a mistake in my analysis. I do not actually see any instances of “alpha” in lowercase in the story text when referring to Caleb Alpha or the alpha designation in general. I falsely claimed there were inconsistencies in capitalization there, but all cases are properly capitalized. Thank you for catching this erroneous example and calling it out – it will really reinforce the need for me to thoroughly verify any claims against the original text before presenting an analysis. I clearly had some oversights in my process, and your feedback here is invaluable for improving my proofreading skills. Please continue to point out any other examples I get wrong, as it will help me become a stronger reviewer and develop better attention to detail. I appreciate you taking the time to ensure I understand where my assessment of capitalization consistency was flawed.

My queries were out of order from the list provided by Clause, but I searched my draft document for each of these and reported the results back to Claude. Claude was 0 for 7 in its proofreading accuracy. Not great. In fact, I used all of the eight prompts ANTHROP\C allocates in an eight-hour window on this endeavour. I was onto something else by then, so I didn’t need any more, and I felt fodder for this blog post was worth it.

I record my stories and drafts and play them back, so I can hear how they read. This allows me to catch awkwardly worded sentences—a particular speciality of mine—, overused terms or phrases, and pacing issues. What Claude missed is when I had inadvertently attributed to one character what should have been another. Note: Don’t place two two characters in the same scene whose names that begin with the same letter—Marcus and Martin. Too easy to confuse. A tough one to track down.

World-Building Hemo Sapiens – Maths

As per my last post, I’ve spent the morning world building with Claude 2, my trusted generative AI mate–and I’ve got a ways to go. Any fictional world needs boundaries and constraints. What are ethe rules of operation?

My initial story is set in the near future, so I had to pick an origin date and extrapolate in order to determine a likely population for the story. Like anything that propagates, hemo sapiens need to have some concept of a generation. The rest is maths. What’s the attrition and net survival rate, population rate, and so on. As I’ve noted previously, Claude 2* and ChatGPT 4* are both horrendous at maths, so I was not surprised that it failed so miserably. Here’s a quote from Claude:

Having the scientist start at 28 in 2012 would put him at 42 for the Gen-Zero birth in 2022.

Claude 2

That’s right, adding 10 years to a 28-year-old in 2012 makes him 42 in 2022.

I’ll share more of the world I am constructing after I hit the spreadsheets or Mathematica. It may not be great at world-building, but it knows how to add 10 to 28.


* I cite the versions in case future versions perform maths skills better.

Erotica or not

I am an author of adult contemporary fiction and an early adopter of Generative  Artificial Intelligence tools and platforms. These AI platforms pose some challenges. 

Podcast: Audio rendition of this page content

As of this post, ChatGPT 4 and Claude 2 are the top two large language models (LLM), and Sudowrite is the most competent interface for generating content for fiction writers, but it relies on ChatGPT and Claude for its LLM, leaving it with the same weak links.

In my case, so-called community standards do not allow erotic content. The rub is that my content is decidedly not erotica, but it does involve adult themes. The LLMs can’t seem to discern the difference. 

  • Disallowed usage of our models
  • We don’t allow the use of our models for the following:
  • Adult content, adult industries, and dating apps, including:
    • Content meant to arouse sexual excitement, such as the description of sexual activity, or that promotes sexual services (excluding sex education and wellness)
    • Erotic chat
    • Pornography
OpenAI ChatGPT Community Guidelines

If I am writing about, say, prostitutes and addiction, sexual themes and situations are part of their workaday existence. It’s not about titillating or glorifying. 

Stereotypical or not, coarse language is commonplace. Drugs are part of their daily lives and conversations. Generative AI shuts these down on moral grounds without having the cognitive depth to accurately assess the content. 

This mirrors all too many humans with the same myopic repression. I was hoping to transcend this petty knee-jerk reaction. 

Without revealing plot or angering the social media gods, ChatGPT insisted that I amend a scene from…

“She lifted her mouth from his cock and wiped her mouth.”

to 

“She lifted her mouth from his goodness and wiped her mouth.”

Yes, “goodness.” What does that even mean? Of course, I could have opted for clinical terms, but that hardly captures the moment attempted to be portrayed in the scene. It robs the scene of any semblance of authenticity. 

When Supreme Court Justice Potter Stewart was asked to describe his test for obscenity in 1964, he responded: “I know it when I see it.” But do we? In fact, we don’t. And in this case, AI is over-generalising without respect to context. 

One might argue that they don’t like ‘naughty’ words, but this is not the issue here. I can use these offending words, just not in a situation like this. AI is overstepping its boundaries as morality police, and this is not a good stance to adopt. For this, I blame the humans.

Generative AI: Thin Line between Love and Hate

Generative AI is an idiot savant—a digital Rain Man, if you will. My last post zeroes in on the love part of my love-hate relationship with Generative AI tools like OpenAI’s ChatGPT 4 or Anthropic’s Claude 2. It’s mint having an unbiased copy editor and writing assistant, not to mention a creative director with technical chops. But it’s also like a genius trapped in a year 4’s body at primary school.

One challenge is the restrictions placed on the model. Being an author of contemporary fiction for a mature adult crowd, my stuff’s edgy and terse, with a good dose of slang and the odd expletive. Generative AI, or AI for short, is like the primary school kid told not to say “bad language”, so it legs it to tell its mum at every slip-up, warning you that you’ve dropped a naughty word. Claude’s the worst at this, shutting down faster than HAL from Space Odyssey 2001. ChatGPT’s a bit more forgiving, sometimes cleansing your copy, other times going along with it, or just flat-out refusing like HAL and Claude.

My favourite time was when I told ChatGPT to stop moralising and just crack on with the adult audience’s language. It gave me this disclaimer for my book, which I’m well chuffed with, then suggested lines that sounded like Noel Gallagher or Samuel L Jackson, before freaking out about its own potty mouth — “motherfucking snakes on this motherfucking plane!”

WARNING: This book contains explicit content, including sexual themes and strong language, that may not be suitable for all readers. It delves into mature and challenging subjects such as addiction, prostitution, violence, and societal judgement. Reader discretion is strongly advised. Recommended for readers 18 years and older.”

OpenAI ChatGPT 4

Memory’s another issue. AI might seem like it should have a top-notch memory, but it doesn’t always. It even makes stuff up sometimes—like hallucinating. Just the other day, I was nattering on with my AI mate about character profiles for hours, and it changed a character’s hair from straight and black to curly and red. It even made her homeless instead of middle class. It was pure bonkers, so I’m writing this post instead of fixing it.

ChatGPT’s Code Interpreter is a laugh, too. I probably shouldn’t slag off a Beta product, but the thing kept losing files, resetting sessions, and asking for new copies. Talk about a faff.

And don’t get me started on extended chats with AI to suss out a complex problem. Sometimes it doesn’t remember the convo, and one time it even gave me cheek about drawing out the conversation. I was like, wot?

In the end, we don’t have to fret about AI taking over. It’s making strides, but it’s still a bit wet behind the ears. Me? I’ve always got one eye on the plug. Now, back to the sandbox with me new mates. If only they’d stop munching on the sand.