Are the companies which make generative AI dreadful? Yes. Is the whole thing bad for people? Almost certainly. Should I use it anyway? Yes, but slowly.
My generative AI training partner and I are in the process of writing final evaluation reports for training we’ve done over the last seven months.
Starting on September 30, 2025 and ending on April 23, 2026, Anne Taylor and I – generally in a state of some nervousness – ran workshops for 173 people in three different cities. The workshops started with an introduction to generative AI and ended, six hours later, with building Custom GPTs.
There were presentations and interactive exercises and group discussions. In the group discussions, we heard one issue mentioned again and again: the fear that easy access to generative AI will make people lazy, or will stop them from thinking for themselves. Allied to that was a concern that young people are entering the workplace with only superficial knowledge of their subjects, or that learning on the job will be replaced (inadequately) by over-reliance on AI tools.
Using AI – with constraints
This is something I’ve thought about and written about in various different ways over the last couple of years. It’s certainly something I have experienced personally: I experimented early on with having generative AI write articles for me, and found that I disliked the process intensely. Once ChatGPT had written those 800 words, I found editing them to convey my own voice impossible: I hadn’t done enough thinking and tinkering and battling with ideas to have any voice at all.
These days, I use generative AI in very constrained ways, as an add-on to the things my brain, heart and hands do and make, however imperfect that might be.
What I’ve learned about what’s wrong with generative AI
Over the last couple of years, I’ve also done enough reading to have a clear sense of why it is that AI-generated writing is just not good enough.
It’s a common sense idea that human-made is best, and that machine-made is not quite as good. We know that we would pay (if we could) for handmade, tailored clothing rather than mass-produced items from a factory. But it’s not quite as clear when we enter the world of creativity (text, images, video). How do we judge if (say) a piece of marketing copy written by generative AI is better than a human-made version? There are three ways to think about that.
One: the creativity problem
An article in The Conversation, written by Ahmed Elgammal, Professor of Computer Science and Director of the Art & AI Lab, Rutgers University, takes a look at the limitations of AI’s “creative capacities”.
Elgammal says AI systems have been trained on massive collections of visual data and have been rewarded for producing results that closely match the patterns contained in those visuals. “Because they’re optimized to produce familiar outputs, they end up suppressing novelty. This, it goes without saying, doesn’t lend itself to true creative breakthroughs.”
No argument there. But it was this paragraph that stopped me in my tracks:
Furthermore, these systems don’t learn from a vast repository of data that encompasses the visual world and all human artistic outputs. Instead, the data used to train these models has often been curated to favor certain images and videos that are polished, clear and visually appealing. In effect, the training process teaches models not just what things look like, but what good-looking content is supposed to be. (my emphasis)
This then is the first thing that’s wrong with AI output – it reproduces something that humans have already made, and judged to be good enough. But, says the prof: “Good creative work involves pushing boundaries, not simply coming up with something that’s passable and palatable.”
Two: the agreeability problem
In addition to AI outputs being passable and palatable, they are also agreeable. Which sounds like a good thing. But as philosopher and educator Esmè V asks in a Substack article and related LinkedIn post about a discussion in her classroom: “When you hand your judgement to a system designed to agree with you, what exactly are you trading away?”
She reckons that when we talk to something that always tell us we are right, we trade away our ability to reality-test. “Real people, with real opinions, speaking into your life – that is not an inconvenience. That is how you stay calibrated.”
So AI outputs are passable, palatable and endlessly, uselessly agreeable.
Three: the shape-shifting problem
It’s often said that generative AI tools are left-wing biased (in this case, the left-wing as understood in the United States).
But a newly-published study says something different. As described in a LinkedIn post by Professor Petter Törnberg, one of the study’s authors, political bias in an large language model’s answer to a question will shift based on who the LLM thinks is asking the question. If it thinks the person is left-wing, it will give a left-wing answer, and vice versa. “Political bias in an LLM isn’t a coordinate on an ideological map. It’s a response profile that depends on who the model thinks it’s talking to.”
So, shape-shifting, sycophantic and palatable. Lovely.
Remind me again why I am using AI?
In spite of all that, generative AI has its uses (just as the internet has its uses, despite many problems with social media platforms, to name just one thing).
I use AI in particular ways because I know the ways in which it is problematic; the things I do are designed to minimise the blandness and yes-person-ness of generative AI outputs.
Thinking that that might be useful to other people, I decided to see if I could codify the steps I take when I use AI.
To that end, I fed five blog posts that I’d written that touch on this (links at the bottom of the post) and asked NotebookLM for “a practical, scaffolded and sequential set of steps that people can use when working with generative AI”. I got, as usual, a somewhat long-winded response.
So I went over to Google’s Gemini (which is now, wonderfully, integrated with NotebookLM) and asked it this: “I want readers to come away with a printable checklist they can use when working with generative AI so they can preserve their humanity and creative edge.”
Here then is that checklist, edited, rearranged and annotated by me (and available here as a printable PDF):
First some ground rules:
Rule 1: You will always, always, always have to audit the output an AI tool gives you.
Rule 2: If you don’t understand a task well enough to audit the AI’s output, you should not be giving that task to a machine. Stay in your lane!
The checklist
First steps
- [] Triage: How important is the output? Is it small and personal, or is it mission-critical business-related? (This affects what you do with the output: skim it and use it, or interrogate it.)
- [] Minimise: Have I reduced this task down to the bare minimum I need? (Ask for an outline; not the whole article. Or for three questions important to the problem you are trying to solve; not the entire solution).
- [ ] Manage the process: Be in charge. Think through the problem before starting – what are my ideas about this, what do I actually want, in what format. What don’t I want? (Note – you don’t need to do this to get a recipe for supper; you do need to do it to generate a content strategy for a business.)
- [ ] Privacy: Does what I am about to enter into the AI tool contain PII (Personally Identifiable Information) or trade secrets? (If so, take them out.)
The human-in-the-loop audit
- [ ] Verification: Have I checked the “facts” the AI provided? If I can’t, is there someone I can ask?
- [ ] Tone check: Is this “workslop” or does it have my own human voice? (Try reading it aloud!)
- [ ] Bias check: Does the output rely on stereotypes? Does it work for my audience (who might, for instance, be second-language English speakers)?
- [ ] Extend thinking: Can I push the AI to do better? (Ask it to combine ideas in “extreme” ways or adopt specific personas to unlock creative sparks.)
Optional: Professional growth checks
- [ ] Tutor vs ghostwriter: Am I using the AI tool to help me understand the subject, or just to finish the task?
- [ ] Manual mastery: Could I do this task without generative AI? If not, should I learn how to do it?
- [ ] Time allocation: How will I use the 15 minutes I just saved? (e.g., reading a book, a walk, or a deep-focus planning session).
- [ ] New skill: Did I learn one thing this week the old-fashioned way (no AI assistance)?
Final thought: just slow down
The seductiveness of generative AI doesn’t lie just in its pleasant (yet bland) personality. The speed with which it can get things done is the true quicksand. A quick recipe is a godsend; a report for your boss that got generated in five minutes could be a career-killer if you can’t actually answer questions about it.
I give the final word to independent software developer and coach Mario Zechner who has this to say about using AI agents to help when writing code (warning – some rude words):
The point is: let the agent do the boring stuff, the stuff that won’t teach you anything new, or try out different things you’d otherwise not have time for. Then you evaluate what it came up with, take the ideas that are actually reasonable and correct, and finalize the implementation.
And I would like to suggest that slowing the fuck down is the way to go. Give yourself time to think about what you’re actually building and why. Give yourself an opportunity to say, fuck no, we don’t need this. Set yourself limits on how much code you let the clanker [agent] generate per day, in line with your ability to actually review the code.
That may be about writing computer code, but his words are wise: the best way to use generative AI is slowly, resisting the temptation to go faster than you yourself are able to go.
Main picture: Frames For Your Heart, Unsplash
Previous Sensible Guide articles
Who is using Gen AI, and what for? Lessons from the training room – Training people is always a two-way street: a look at what’s been learned in the training room about who is using Gen AI, and how.
Is Gen AI making us lazy? And other 3am questions – We’re living through deep changes in the way we interact with technology. If you are worried that Gen AI is making you lazy, here are some ways to approach that…
Productivity – what is it, can AI help? – Being busy is not the same as being productive. That’s a cliche – but sometimes cliches are useful.
The end-of-year Gen AI post: be afraid, be thoughtful – An end-of 2025 post, reflecting on a year of generative AI and the emerging dangers – but also to think about what to do about that.
What about the young people in the age of GenAI? – What should employers be doing about the young people they hire in the age of generative AI?