why I write by hand (and when)

0600 on a late-October-standard cold, dark, and rainy Danish morning. Up with the baby in a sling, an hour before anything else happens. Once the early wakeup mind fog clears: why not start the day with a splash of writing? Many of my blog posts require at least some research, and my entry point to that is typically DuckDuckGo, so I fire off a query and...

...I'm reminded why so many (myself included) are fed up with traditional web search. First page: 10 results; 1 is a relevant webpage, 1 is a relevant Wikipedia article, 1 is a relevant Reddit thread, and the other 7 are pure slop. Someone got their hands on some SEO keyword data and cranked up the robots to barf out words.

Roughly zero thought, editing, or curation went into those pages. This is painfully obvious from reading the first couple of lines, at which point a visceral ick comes over me and I make for the door. Sometimes the title and preview snippet alone are enough to get the ick.

Sure, there's the AI summary box. I don't like the AI summary box. It replaces this distribution of mostly-slop with a digested summary of said distribution of mostly-slop, sort of like eating a thrice-recycled hamburger. It also means the websites that "contribute" material to that summary get 50% less traffic.

I'm aware there's a bit of cognitive dissonance here: I don't feel the same way about, say, self-checkout in supermarkets. But for me this isn't an economic argument: I don't care about the revenue from my website; I don't make any! I just prefer it when people read and engage with what I write, rather than writing into a void to indirectly enrich billionaires.

Which is why, when it comes to anything I write for a human audience, I still insist on the purely human approach: hands on keyboard, no AI writing assistants in sight. (In the waning days of 2026, this is a bold claim requiring proof. I can't offer proof per se, only evidence, but we'll come back to that later.)

If I want living, breathing people to read and engage with what I write, I should write something worth reading and engaging with. It's that simple. Slop is not worth either: when you read slop, you're reading a pale shadow of human thought, and you know it. Damn, do you know it - that's the ick I mentioned before, an uncanny valley of content.

💡
If you're looking for better slop detection criteria than "I know it when I see it", Wikipedia's Signs of AI Writing is well worth a read. TL;DR: it's not just em-dashes and delves; AI tells evolve rapidly, and extend into every facet of the writing (word choice, structure, semantic markup, even how they claim to adhere to platform policies!)

Of course, we've seen this before with blogspam and listicles and patent medicines that sound like D&D items: as long as repositories for the printed word have existed, unscrupulous actors have scrambled to fill them with slop for personal gain! So this is not a new problem. The novel part is that GenAI can produce plausible-looking prose, and do so quickly. Yesterday's ghost-written marketing content or keyword soup on a domain-parked website has been replaced by large volumes of syntactically correct (if bland) human language whose semantic inaccuracies and omissions are only evident on a close read.

There is a growing asymmetry between the time needed to create slop and the time needed to identify said slop. I'm reminded of the early days of email, which were rife with spam and chain letters and THAT PERSON WHO LOVES SENDING ALL CAPS MESSAGES and missing half the thread because someone didn't Reply All. Until we have equally strong slop detection and filtering, and equally strong social taboos around worsening this asymmetry (the workslop classic: "I expect all 20 of you attending the meeting tomorrow to read this multi-page slop document beforehand"), we can expect this to be an equally painful user experience for most.


OK, but where's my evidence?

Yes, I promised evidence that my writing isn't AI-generated. I like to think you can tell, that your taste and my quirks are clear enough demonstrations of this fact. I'm also an engineer at heart, and would probably reject this explanation if given by someone else.

So let's answer this the late-2026 software-brain way: with more AI! It takes one to know one, as they say - and there are now tools, such as Pangram and GPTZero, based on models specifically trained to estimate the probability that an input is AI-generated.

One specific thing to note about purpose-specific models in the AI landscape is that they often command a hefty premium over generic LLMs. The Pangram API costs $0.05 / 100 words. For comparison, frontier models are ~$2 / M input tokens right now. Using the "1 token = 0.75 words" rule of thumb, this is $2.67 / M input words. So Pangram is ~200x as expensive as your average frontier LLM.

As someone who uses open-weights models instead, though, a better comparison would be something like ZeroGPT API pricing, which is $0.042 / 1000 words. I pay ~20x less than frontier, so for me this is ~300x as expensive.

I should be careful to note: LLMs (and other generic models like jev) perform fine at a wide variety of tasks; purpose-specific models perform better at a specific task, and not at all on others. Tasks where accurate classification (or prediction, or inference, or...) has high value, or where inaccurate results have high costs, may well benefit from paying the premium.

💡
Though here I wonder: if the cost disparity is 200-300x, could I simply call a bunch of different LLMs and treat them as weak learners in an ensemble model? This likely only makes sense for classification tasks over larger datasets, where the upfront cost of human labelling, training, etc. can be amortised over many calls. I won't explore this here, as my dataset (~50 blog posts) is too small for this approach to be viable.

which tool to use?

I've written ~60k words on this blog to date. At Pangram API rates, this would be ~$30, or ~$24 at bulk API rates. ZeroGPT is considerably cheaper at ~$2.50, but it has a minimum credits purchase of $9.99; still, that's cheaper than Pangram, so I'll go with it.

💡
Total actual cost: $2.64, including $0.17 spent on a test run with 2 posts. I guess I now have $7.35 in ZeroGPT credit, valid indefinitely 🤷‍♂️

before the results, a disclaimer

No AI detection tool is 100% accurate, nor are their detection models infinitely robust to attempts to disguise AI writing. This repo, for instance, offers a proof-of-concept using Unicode spaces to confuse ZeroGPT.

Every AI detection tool also exhibits false positives, meaning they will occasionally flag human writing as AI-generated.

If your use case for AI detection involves punishing authors in consequential ways (e.g. plagiarism detection), don't rely exclusively on the output of AI detectors to do so.

results

I analyzed 51 posts using ZeroGPT, with the help of Mimo V2.6 Pro (plan mode) and GLM 5.3 Flash (build mode) running via OpenCode to create API client scripts.

Each response from the ZeroGPT Detect API includes these fields, which I've used for the analysis:

  • response.feedback: a human-readable description of ZeroGPT's overall judgment (e.g. "Your Text is Human Written")
  • ai_pct: the estimated percentage of the text that is AI-generated
  • response.h: specific sentences flagged ("highlighted", hence h) as AI-generated

Looking at response.feedback, I have:

  • 48 posts judged as human-written
  • 2 posts judged as "Most Likely Human written, may include parts generated by AI"
  • 1 post judged as "mixed signals, with some parts generated by AI"

My word-weighted average ai_pct is 1.18%, which is pretty low - but not zero! 5 of 51 posts have ai_pct > 0:

  1. adversarial wordle (36.7%)
  2. Schengen ain't what it used to be (21.2%)
  3. alignment (6.8%)
  4. goodbye, big tech (5.9%)
  5. forests and trees (4.3%)

Curiously enough, it's not #1 that gets flagged as "mixed signals", but #2.

#3, #5 were written back in 2023. As I mentioned above, false positives are a reality of these tools; the most famous example is possibly ZeroGPT flagging the Gettysburg Address as AI-generated. So perhaps I'm in good company, and these posts have achieved Founding Fathers-level ubiquity? checks analytics nope.

#1 starts by quoting a description of the game Wordle from Wikipedia - and indeed, all the response.h sentences come from that description.

#2 quotes the Schengen Borders Code, but here response.h sentences do not come from those quotes. In this case, the post itself is heavy on data analysis, which leads me to use a lot of evaluation adjectives (purported, minimal, empirical, solid, multiple) and hedging clauses (likely, purported, feels, trending, found minimal support).

For #4, here are the sentences flagged by ZeroGPT:

{
  "h": [
    "While it's true that their track record on privacy has been reasonably solid, recent forays into scanning iCloud / iMessage content , facial recognition , and AI call this reputation into question.",
    "In practical terms, that means you can't really avoid LinkedIn if you want to land a job in Denmark - not unless you have a solid professional network and are comfortable limiting your search to smaller companies that won't toss out your application when they can't find your LinkedIn profiles.",
    "💡 Reducing your Big Tech footprint isn't an all-or-nothing exercise.",
    "There may well be platforms you can't easily leave - but by being aware of your footprint, you can make informed decisions about where you're willing to compromise.",
    "If I were to leave Denmark, for instance, I'd likely re-assess my LinkedIn usage.",
    "Browser What I use now: Chrome, Safari Viable alternative: Firefox Steps to migrate: install Firefox import profile data from Chrome set Firefox as your default browser delete Chrome Even though Firefox has been steadily losing users and has low overall usage (~2-3% of total), it remains a robust open-source browser and a solid option to help protect your online privacy .",
    "And finally, some areas where I don't need the service at all: TV: my wife and I used to share her parents' Netflix subscription until they stopped supporting long-distance account sharing - and quite honestly, we haven't missed a platform whose CEO once declared they \"compete with sleep\" .",
    "taxi: we stopped using Uber a while back; they're notorious for underpaying drivers and ignoring local regulations .",
    "In most places we visit, it's easy enough to call a cab - or just take public transit, or even bikeshare or walk."
  ]
}

Sundry things that jump out at me: I love dashes for separating clauses - and always have. (How dare the robots take this from me!?) Many of these sentences have a sort of "generic summary" tone; even though I wrote them myself (I promise!), they're bland enough prose to be written by any other passably competent author, bot or human. (Maybe a gentle nudge to further develop my own writing style / voice?) I like to cite specific stats (e.g. "2-3% of total"); bots love this too. goodbye, big tech is my longest blog post to date (4844 words!), and is fairly formulaic, in that 70-80% of the article is semi-structured exposition on alternatives to Big Tech; maybe anything long enough with some structure trips the AI detection?

All of this is mere speculation. I have no way to know for sure why these sentences were flagged. I'm not sure what to think of the emergence of magic machines: systems composed of nondeterministic agents and models, whose inner workings we can never know but only intuit or guess at. Explainability has not kept pace with functionality. Every day we move further away from systems whose function we can inspect, understand, and explain; those powers become the exclusive province of AI labs and model providers, on whose royal lands we peasants may not poach. (Not that they can explain these functions either!) We lurch onwards towards the dystopian visions of Rainbows End, in which physical and digital artifacts alike have No User-Serviceable Parts Within.

some thoughts / reflections

I claim to write all my posts by hand, 100% human, no AI. ZeroGPT flags 3 of 51 posts as potentially containing AI-generated sentences, and estimates my overall content corpus as 1.18% AI.

But what does "100% human" mean? If I quote external sources, and those sources were AI-generated, some AI detectors will count that against me. I've also used AI in non-textual content twice: the game code in adversarial wordle was my first attempt to use GenAI in a more complex coding task; servant leader, peer leader contains an AI-generated diagram midway through the article. (These instances are not measured above; I only analysed text, not code or images.) As with debates over open-source code: some sentence constructions are fairly standardised, especially when it comes to data analysis; here it may be practically impossible to avoid these constructions without abandoning analytical clarity.

So maybe I'm OK with being measured as 1.18% AI by the robots. I've learned a bit about my writing style and the challenges of authoring original work in this brave new world.


where to from here?

In the spirit of There's never been a better time to go into math from 3Blue1Brown, one reflection and one question.

The reflection is a simple one: high-quality human writing is more important than ever. As it becomes easier to produce noise, the value of signal rises. Good human writing provides signal. Great human writing amplifies and clarifies signal.

Now, I'm decidedly not an AI vegan. AI does have its place in the writing / thinking process. It makes a decent smart search engine, helping surface pages, articles, papers, etc. to read in more detail. It can also produce fairly useful summaries of long texts, similar to how academics might scan paper abstracts to help shortlist papers for closer reads.

I know, I know - I heaped disdain on AI summaries in web search above; what's the difference? When you search, any resulting AI summary uses a prompt constructed behind the scenes, based only on your short query. When I prompt as part of research, I'm typically asking longer and more nuanced questions, with the intent of finding sources. I'll then go and actually read those sources. AI summary as destination is a perilous and foolish void of thought - but the same was always true: of entering a search query and only looking at the results page; of looking at Wikipedia and not curiously following any of the cited sources; of searching the library catalogue and, say, only using the publisher keywords and marketing blurbs.

I'm not under any illusions: AI will still ignore some of my nuance, misinterpret what I say, and respond with irrelevant nonsense on occasion. Sometimes the irrelevancies are glaring; sometimes they are frustratingly subtle. Given that reality, I'd rather be able to correct with follow-up prompts, much as I might correct coding agent output through multiple rounds of review, than to accept a canned summary as truth. One must learn to work with the grain of the material one has.

💡
Speaking of material: I've been thinking recently about how AI-driven programming removes a sense of materiality from my work. Much as a carpenter might experience different woods, sizes of piece, etc. each in their own unique way - backend, frontend, DevOps, data, mobile, etc. work each had their own feel before AI; after AI, they no longer do. I wonder about this as a contributing factor to cognitive debt, and as a flattening of experience and specialisation that seems to me one plausible source of existential dread in the profession right now.

...but, you might say - but: what if AI writing does become good enough to be indistinguishable from human, or perhaps even better (by whatever definition); what then? In this case, I'd still argue - much as Grant Sanderson does in his 3Blue1Brown video - that the act of writing itself still has value. It trains the writer to think complex thoughts, and to be capable of expressing them. In all but the most accelerationist / doomerist of scenarios, we will continue to need as many living, breathing humans who can think and express complex thoughts as we can get.

Writing is also a sort of theory building, as all exercises of language are: how do I sequence my thinking, such that each part builds naturally on the last? Not only that, but it hones our theory of mind: our cognitive empathy for others as thinking beings. Who is my audience, what would help them understand my point, and what might their response be? These are questions worth asking, especially in uncertain times.