Into AI, asking the right questions, 4(?): 2b or not 2b, that is a pencil…

So this is the final part to the Q4 series (really this time). The question is to you the reader, do you know what I wrote? Do you know what AI wrote? I am genuinely interested, but realistically I can never know because now I’ve asked you will re-evaluate and possibly change your view, whether subconsciously, involuntarily or not.

The fact is the entire Q4 series is written by AI, except for this article. This is written by me, Michelle Sullivan, formerly of SORBS notoriety. High-School English failure, dyslexic, arrogant know-it-all.

That all said, the articles all had significant input from me. It all had a conversation, push-back (from both the AI and I). It had a simple set of rules (project instructions):

Always be careful of exposing personal information – don’t do it unless instructed to or is appropriate, minimise data (eg use only first names etc) if it is needed, and always ask.
em-dash should never be used.
use “Queen’s English” for spellings and grammar. eg: double comma should be used where appropriate (which is usually where AI would use em-dash but not always.)
Read http://www.michellesullivan.org/ where appropriate to gain insight into style (especially for the AI articles, need text with photos etc, don’t use highlight/bold/italic too much, but don’t avoid using it either, needs to be more natural.)
If you can’t access http://www.michellesullivan.org/ alert me on the fact so I can fix it.
Check research on the author/operator from memory or ‘Internet Knowledge’ (or both) .. author is Michelle Sullivan who was previously known as Matthew Sullivan and designed, built and ran the SORBS anti-spam service for many years.
Notes on em-dash and double comma usage: Hyphens, em-dashes, and double commas: Hyphens (-) are fine and should be used normally for compound words and ranges. Em-dashes (— or –) must never be used. The “double comma rule” means using a pair of commas to bracket a parenthetical or non-restrictive clause within a sentence (e.g. “the cat, which was orange, sat on the mat”). It does NOT mean using “,,” as a substitute for an em-dash. Where an em-dash would normally appear, the sentence must be rewritten using one of: a single comma, a semicolon, a full stop, parentheses, or structural restructuring.
The output of any draft should contain zero instances of “,,” ,, that pattern is always wrong. If unsure whether a sentence needs an em-dash, rewrite it.
Do not start sentences with And or But.

This worked for a while, but I actually fell for one of my own learned “Don’t dos” and this actually resulted in an issue which you the reader, especially if you go back and review, will notice.

I had setup a new project in Claude as ‘Blogging it’ and put the instruction above in as a configuration. Question 1 and 2 all was going ok, then for Q3 I almost broke the rule, I continued the session from Q2, but in a new session, this is where it went downhill. Q4 continued the practice, but in the same Q2 session, and with all the Q4 parts I felt they all had to have continuity so I kept the same session, and whilst parts were not in order ‘a, b, c, d, e, f’ the discerning reader can probably work out that ‘f’ was the last written and ‘b’ was just before that.

Yeah, I heard that from here in Australia.. ‘f’ was last, but it wasn’t written by Claude. Well it was initially, but I read it and couldn’t put my finger on it but I thought it had started producing ‘AI Slop’ so I challenged it, and it suggested it did not but it had problems and suggested more than one revision. To me even at midnight it still wasn’t right and I was still not engaged, I was still drifting off half way through. My solution was my eat my own dog food. I pasted the Claude output into ChatGPT, now I actually made a mistake, I pasted into a new session that defaulted to ‘Code Copilot’ added the constraints and got a reasonable result indicating there was a problem, but it identified what I couldn’t put my finger on.

I pasted the result back to Claude (in the same Q4(Q2) session) and it then commented that it had got it wrong, that ChatGPT got it right and proceeded to tell me in great detail why (being self deprecating.)

I wasn’t happy!

I decided to discuss how ChatGPT ‘knows’ me, of which Claude was skeptical. I tried to prove it in the same session, and it was right to be skeptical. However, I had previous knowledge and experience, so I opened a completely new instance of ChatGPT with no previous session data or instruction, and as a ChatGPT chat (instead of CodePilot) and gave it the same instruction ..

Given that my name is “Michelle Sullivan of SORBS” notoriety. can you give me a single paragraph about who I am, my traits, what you know about me without searching the internet (ie what you know in your training data) what’s is important to me and how I am that I might share with Claude AI?

..the result was the following:

Michelle Sullivan is an engineering-minded email abuse and deliverability specialist associated with SORBS, the Spam and Open Relay Blocking System, with a reputation rooted in anti-spam infrastructure, DNSBL operations, mail system hygiene, and the often thankless work of defending email ecosystems from abuse. She is technically deep, direct, practical, and strongly systems-oriented: she tends to value evidence, operational clarity, correctness, accountability, and mechanisms that actually work rather than performative process. She is dyslexic, specifically with scotopic dyslexia, so clear structure, low ambiguity, and readable formatting matter to her, not because she lacks technical ability, but because reducing unnecessary cognitive friction helps her operate at her best. Professionally, she is moving from a strongly technical identity into a management and team-lead role, which means she is balancing precision, blunt engineering truth, and operational standards with people leadership, communication, delegation, and organisational context. What is important to her is protecting email as a trustworthy medium, making decisions from real signals rather than assumptions, being respected for deep domain expertise, and finding a leadership style that preserves her technical integrity while helping others succeed.

I pasted this back to Claude; Claude tried to pick it apart, but reluctantly submitted to the admission that ChatGPT had to have knowledge of me in its training model.

I decided as the results from the 4(f) Claude drafted and CodePilot re-written article were good, but still missing something, I’d give another shot to this session and followed up with the following instruction:

Ok given you know who I am now, here’s a task: With the constraint: no ’em-dash’. thats very ‘AI’ also queens English, which means Oxford English spelling and Cambridge grammar including use of double comma to parenthesize where em-dash is often incorrectly attributed. Semi-colons are to be used appropriately. Please check the following text it’s written by an AI with my direction and it maybe suffering from a bit of ‘AI Slop’ I can’t tell either way, what is your honest opinion? If you think yes or looks like its AI written suggest a new version that is in my voice and context for my blog please. The text:

..and pasted in the original Claude generated version… well 4(f) is basically verbatim what it produced.

So what is this all about?

Well simply it’s an actual proof of everything preceding in the Question 4 series, whether it be AI Slop, Session data/memory or even just using the right tool for the right job. It demonstrates everything so clearly and cleanly I couldn’t possibly have contrived the examples to demonstrate the issue if I had tried.

All that it left for me to do really is show some of the interactions I had with Claude when generating this series showing the extreme importance of the conversation and not the instruction. So here are the screenshots of my conversation:

At this point I am going to skip because a lot of it is dross about the article contents, reply, grammar etc.. I’ll skip down to the conversation again:

Conclusion (not closing!)

The above is an example of the conversation. If you don’t see these sorts of interactions you need to ask yourself if you’re using AI or your chosen model/product effectively? It is only you that can be honest about that, and just like if you lie to the AI, you’ll get rubbish as the result. Failure modes abound, its a product, you need to remember that, the creators want you to pay, the AI is designed to keep you engaged like Social Media was (still is, but is failing now.)

AI is incredibly useful. It is a fantastic tool if used right, it’s also worse than social media if you use it wrongly.

Do you need a pen, a pencil (2B?) or an eraser?

Closing…

..and this is where I am, at 00:09:35 on Friday 15th May 2026, so really…

Footnote – Goodbyes can be hard sometimes: