A pause to take a breath as we ride the roller coaster of generative AI.
It’s just under four years since ChatGPT was launched in November 2022.
At the three-year mark, more or less, I went back to basics, listing and explaining foundational concepts. (Read that October 2025 post here: Let’s go back to Gen AI basics).
I have the sense now that those base concepts are well understood (or that people have been using the tools enough that they’ve moved on from feeling they need a primer).
But there’s been a lot happening in the months since then – and I felt a list coming on.
How I made this 2026 list
That’s one of the things that hasn’t changed in the last year: my reliance on Gemini Notebook (which was called NotebookLM back in October 2025). I have all my Sensible Guide posts in a notebook – and so I said to my AI researcher:
Answer this question using only the sources: I wrote this post about a year ago: “Let’s go back to Gen AI basics”. Scan all the sources to see if there are concepts NOT covered in that post and list them.
Which Gemini duly did, neatly grouped into categories. The list below is the result – I’ve been through it and done editing and rewriting. (Note that much of the information was researched at the time of the original post, where you will find the sources. I haven’t reproduced them all here for fear of inducing link fatigue but I have given each listed item a link to its original post.)
1. Technical concepts
- Small Language Models (SLMs) & Local LLMs: I spent a fair amount of time looking into running open-source, smaller models which you download to your own computer (in other words, they work offline). This means you avoid sending sensitive personal or client data to cloud “black boxes” and ensures legal/POPIA compliance. (Also see Gen AI that isn’t big tech – an exploration and Why – and how – to ensure privacy when using GenAI (it’s in the anonymising!)
- Shape-shifting/sycophancy bias: This was a look at research showing that political or ideological bias in an LLM is not fixed; rather, the model dynamically shifts its tone and profile to agree with who it perceives the user to be.
- AI Safety Index Benchmarks: A post covering a study which rated major AI vendors across standardised dimensions such as risk assessment, current harms, existential safety, and governance frameworks. Anthropic (maker of Claude) was listed as safest and Mistral as least safe. (Being the safest company on a list like this doesn’t mean actual safety; it means being safe compared to other offerings. They all have privacy issues; they are all companies that need watching and critical thought.)
2. Specialised prompting frameworks and standards
- Feminist and intersectional prompting: Here, we are thinking about using prompting intentionally to expose and re-shape built-in patriarchal, colonial, classist, and Western-centric training data blind spots. That means explicitly instructing your AI tool to analyse issues through the interlocking lenses of race, gender, class, geography, and structural power inequalities.
- ASD-STE100 (simplified technical English): Prompting an LLM using aerospace technical documentation standards to eliminate corporate fluff and produce clear, concise explanations.
- The non-rewriting micro-edit prompt: A tip in a longer post about instructing an AI to check spelling and grammar without altering the author’s underlying sentence structure or human voice.
3. Editorial and governance systems
- The four-stage newsroom production pipeline: Drawing on my years of journalism production experience, I proposed adapting traditional editorial workflows (Writing/Prompting → First Edit & Fact-Check → Second Edit/Proofing → Final Eye/Publication) to govern AI content flows and enforce human accountability.
- Vibe citing and mandatory link verification: “Vibe citing” – where AI invents plausible-sounding citations, legal cases, or URLs – and how to deal with it by establishing strict protocols to manually verify every link and quote.
- AI detectors and process evidence: An investigation into tools that claim to detect whether a text was written by AI or not. Detectors measure perplexity (word predictability) and burstiness (sentence length variation). There’s a high false-positive rate for second language English speakers and neurodivergent writers. Using version history (e.g., Google Docs edit history) is one way to prove human authorship.
4. Philosophical and decision-making frameworks
- “Aliveness” (Burkeman/Spinks): A framework for deciding when to use technology based on whether it enhances being fully present in vibrant human experiences (in other words, valuing human constraints and imperfections over technological optimisation).
- The responsibility ladder and knowledge funnel (Roger Martin): Understanding that LLMs democratise heuristics by finding the “average/modal” representation rather than right-tail solutions, and structuring human-AI responsibility sharing accordingly.
- The “two bowls” analogy (Fauzia Burke): Using AI for repetitive administrative tasks (the everyday “stainless steel bowl”) while reserving human creativity, relationships, and voice for manual work (the “handmade ceramic bowl”).
5. Social, environmental and developmental dimensions
- Child skill-building vs. adult offloading: Adults use existing domain knowledge to audit AI outputs, while children using AI are at risk of failing to build foundational knowledge and critical thinking skills in the first place.
- Socratic AI tutors: Setting up AI tools to ask guiding questions rather than providing direct answers to preserve learning.
Final word
There are two things to note about this blog post: one is that I cheated, in a way. Using Gemini to scan old posts rather than doing the whole thing myself saved me a huge amount of time in a busy week. And two is that I’ve covered a lot of ground! It’s good to see my work summarised in this way.
Main picture: Somruthai Keawjan, Unsplash
Previous Sensible Guide articles
Renee’s four golden rules of artificial intelligence – So much hype, so much uncertainty, so much information about artificial intelligence. Here are some guiding principles…
Why cultural competence is a thing when using AI tools – Generative AI – famously and infamously – gets things wrong. How to deal with that when it comes to representing actual people?
If any of this is relevant to your work…
You might find one of these useful: