Alasdair Stewart

Ha, and there is a another rule of three at end of that second image didn't originally notice. Once start highlighting the patterns it definitely reads more and more like they have fed it training data to avoid some usual tells, but it's more 'masked' than 'removed'.

Interestingly, noticed in finding these examples there is more variation in number of items in places. Some of these though are rule of three with 'and so on' or 'even ...' (and variant phrases) stuck on at the end. Also outlined in blue in 2nd image the repetitive phrases can also get in lists.

ChatGPT's love of rule of three - including in 2nd image a rule of three at the end of a rule of three. 3rd and 4th images are examples of bullet-point list rule of three.

Ah, it looks like it's "etc" if in brackets and "and so on" when not using brackets.

The switch-up between em-dash asides and bracket asides is more noticeable across whole response than paragraph by paragraph, but first image shows an example. Second image has em-dashed rule of three aside, this is within same response as first image that has bracketed rule of three aside.

Italics and bold variations. Can see in second image the new found fondness for brackets as well.

Colons are the new em-dashes.

Overall, it reads like its been trained to vary its writing more and reduce some of the most obvious glaring tells that received lot of online and media attention - but this has resulted in an inconsistent style where it shifts between new and old patterns.

- In longer responses there is similar more pronounced inconsistencies in how it structures sections - which was always a tell - and smaller aspects of writing such as whether uses "etc" or "and so on".
- It also seems inconsistent in whether uses bold or italics for emphasis within same response.

- In addition to using "colon: list, of, items", there is lot of needless colon use in general where ChatGPT 5 and older models would have used an em-dash.
- For asides, it varies its use of em-dashes and (asides in brackets) but with no clear consistency of which it uses based on type of aside.

- There is more variation in sentence length, but - while a noticeable improvement - the 'variation' itself feels a bit too rigid.
- When instructed to use paragraphs instead of endless lists it uses "Some random point it generates: a list; of various; phrases; separated by; semi-colons; and so on".

From early testing, it looks like ChatGPT 5.1 uses em-dash significantly less. A lot of the usual tells someone has copied/pasted genAI text remain - such as obsessive rule of three - along with a few potential new ones.

'These are marketing statements driven by profit-motive and ideology rather than empirical evidence and formal proof. In short, these tech CEOs claims concerning “super intelligence” and “AGI”are manifestly bound with their financial imperatives and not rigorous science.'
www.iccl.ie/wp-content/u...

Stop overhyping AI, scientists tell von der Leyen | Euractiv Open letter criticises the Commission President for remarks earlier this year – when she anticipated AI would “approach human reasoning” in 2026

Amazing title for this open letter:

"Open letter: retract your unscientific AI hype"

www.euractiv.com/news/stop-ov...

How Grammarly Launders AI-Generated Content Grammarly’s Authorship tool is billed as a powerful way to prove that your words are authentic. Even though it allows you to easily launder AI-generated text.

Grammarly's rapid descent into an outright plagiarism tool continues. Grammarly will now misleadingly report text rewritten by its AI 'humanize' feature as "Typed by a Human".

www.plagiarismtoday.com/2025/11/06/h...

The way to fix the issue if you experience it? 1. Change your default printer 2. If that doesn't work, remove the driver for the printer that is causing NVivo to crash.

Because this then calls your print driver, if your default printer - even if it's a print to PDF printer driver - is one that NVivo doesn't interact well with, it can crash when you try to open files, view your coding, etc.

Behind the scenes NVivo on Windows is largely built on top of Microsoft frameworks. That is why NVivo has always tended to be able to display Word documents more accurately than competitors. However, when displaying Word docs NVivo sometimes does so by effectively doing a print preview...

In NVivo training I always emphasise that with any issues encountered "It's not you, it's NVivo".

Today an NVivo bug that is a perfect example of this - and I thought was long gone - came up at a training event today.

This bug causes NVivo to crash because... NVivo doesn't like your printer...

How do folk survive commuting to work everyday? Had to leave house at 7am today and just got back 11 and half hours later and feeling half dead.

GitHub - jesseduffield/lazygit: simple terminal UI for git commands simple terminal UI for git commands. Contribute to jesseduffield/lazygit development by creating an account on GitHub.

No idea why I slept on it for so long, but finally decided to give lazygit a go - and it has immediately surpassed my prior git workflow of relying on random aliases I created over the years & praying Control+R can find lesser used commands from my terminal history.

github.com/jesseduffiel...

Always a warning sign when see terms like "deliver efficiency" used in relation to genAI.

Impossible Challenge - Find a pro-genAI journal article that doesn't read like genAI was used at least in part to write it.

There are some decent articles, yet they still have scattering of genAI phrases. Come across a few though that are questionable & won't be surprised if they end up retracted.

Literature Is Not a Vibe: On ChatGPT and the Humanities | Los Angeles Review of Books Rachele Dini discusses OpenAI’s “A Machine-Shaped Hand” and an academic sector in crisis.

Very much enjoyed reading this. Perfectly timed since I'd assigned the ChatGPT short story as a reading for an AI lecture this week.

Initial work on visual refresh of the SGSSS NVivo Guidance pages now done.

sgsssonline.github.io/nvivo-guidan...

"Knowledge is power". With a little coding knowledge - whilst there is an initial learning hurdle, which can be anxiety inducing for students - you can easily build skills that supercede dependence on LLMs. Giving up on teaching students how to code by ceding knowledge to LLMs disempowers students.

Snippets instead show students:
- Whenever you are doing anything remotely repetitive, you can likely automate the worst of it.
- You do not need to memorise code, only how to modify examples/templates, which snippets provide in readily accessible form.
- Snippets are easy to create and customise.

Promoting dependency on expensive and unreliable tools makes data analysis less accessible. If students don't learn how to code, how can they have a sense when LLMs are spewing out incorrect bullshit? The default response behaviour of LLMs is also dire, and risks promoting further dependence.

There is disturbing number of journal articles promoting LLMs as making quantitative analysis 'more accessible', removing need for students to learn how to code, and similar arguments.

My response - "Have you never heard of snippets?"

Amid misinformation, Wikipedia's volunteer community stands out. Editors uphold neutrality, verify sources, and log every edit. 

Without corporate ads, its scale and transparent governance make it a beacon of trust. Learn more ➡️ w.wiki/7zfT