A few years ago I decided I would not use AI. Not for writing, not for research, not for anything attached to the business. I said it out loud to people who asked, and I meant it.
I broke that. I use AI now, including for parts of this article.
I did not change my mind. Almost every objection I had is stronger today than when I made the decision. What changed is everything around it, and then one year in particular that took the fight out of me.
What I was actually protecting
Before anything else: I have worked online for a long time. I build websites and apps, I shoot and edit video, I fly a drone, I run hosting, I have spent years on SEO. I have been publishing kettlebell material since 2009. So the refusal had nothing to do with being afraid of new software, and I want that clear early, because “old bloke scared of technology” is the easiest way to dismiss what follows.
Three things were behind it.
The first, and the biggest, had nothing to do with me. It was the scale of what this is doing to other people. Millions of them, in a lot of countries, most of whom were never asked and have no way of pushing back.
The second was the ability to work things out. Not only mine. The general habit of going and finding out, which I think is being quietly removed from a lot of people, including people who will never notice it happening.
The third was my own business, and I want it named last, because it is the smallest of the three and it is also the only one that eventually moved me.
The part that has nothing to do with me
Somebody wrote every book, every article, every tutorial, every study, every caption, every forum answer that these systems were built out of. Photographers took the photographs. Translators did the translations. Illustrators drew the illustrations. Developers wrote the code. None of them were asked, none of them were paid, and the result now competes with them in their own trade.
That is the part I could not get past. Not the technology. The consent.
Then there is the work itself. A junior writer used to get hired to produce the kind of copy that AI now produces in seconds. A junior developer used to get paid to write the kind of function that gets autocompleted now. Customer support, data entry, translation, basic design, the first rung of half a dozen trades. Those first jobs are how people used to get a foothold. You did the boring work badly, someone corrected you, and after a few years you knew something. Take the foothold away and the people already up the ladder are fine. The ones underneath never get on.
I keep seeing this argued as if it were a debate about whether a technology is good or bad. It is not that abstract for the people it lands on. Somebody loses an income they were relying on, in a specific month, with specific bills, and nobody involved in building the thing that displaced them ever had to look at them.
I did not want to be a participant in that. That was the vow. Not “I might get worse at writing,” which is a small and self-centred concern next to the rest of it.
The internet used to run on an exchange
There was a rough deal in place for about twenty years, and most people who published anything understood it without anyone writing it down.
You made something. Search engines indexed it. Somebody typed a question, found your page, and arrived on your site. Most of them read one thing and left. A few stayed. Of the ones who stayed, a small number eventually bought something, joined something, or hired you.
That was the exchange, and it supported an enormous number of people. Independent journalists, recipe writers, repair guides, hobbyist forums, small specialist sites run by one person who happened to know a great deal about one thing. Almost none of them got rich. Plenty of them made a living, or part of one, and the reason it worked was that the person who did the work had some chance of being found by the person who needed it.
We were one of those. Someone would type “why does the kettlebell hit my wrist,” land on an article I wrote about the clean, and from there find the books, the programs, the certifications, the app.
AI answers break that chain at the first link. The question gets answered on the results page, or in a chat window, and the visit never happens. The knowledge in that answer still had to come from somewhere. Somebody had to do the work, get it wrong, find out why, and write it down. The person asking may never learn that person exists.
It takes the traffic, then it takes the reason to come
Losing search visits is only part of it. The bigger problem is that AI can produce a version of the thing people were coming for.
Someone used to search for a kettlebell workout and end up on a site like mine. Now they type “give me a 30-minute kettlebell workout for fat loss” and get one. Someone used to look for a twelve-week program, compare a few, and buy one. Now they ask, and it arrives in four seconds, free, formatted, with a warm-up and a progression scheme and a confident explanation of why it is built that way.
I have spent over two decades on this. Books, certifications, thousands of hours with people who could not clean a kettlebell without smashing their own forearm. That work now sits next to a text box that answers instantly and costs nothing, and the text box was built partly out of material like mine.
It knows very little about kettlebell training, and that is not the main problem
When AI writes about kettlebells, I can see the seams. It confuses exercises that share a name. It takes a cue that belongs to one tradition and states it as if it applied to every version of the movement. It invents a mechanism, describes it fluently, and cites nothing.
I catch that because I have spent years on it and because I wrote a lot of the reference material I am checking against.
Somebody looking for their first program cannot catch it. That is what concerns me now, more than any question about how good these systems eventually get. To displace an expert, AI only has to sound like one to a reader with no way of telling the difference, and it clears that bar comfortably today.
There are three separate situations here and they get mixed together constantly. Today, in my field, AI is not good enough to trust unsupervised. Also today, plenty of people trust it anyway and train on what it gives them. And the entire industry, not one company, is spending extraordinary sums to close that gap. Whatever any individual company says publicly about keeping humans in the loop, the direction of the work is toward systems that handle more of the job with less human involvement. That trajectory is easy to observe.
So the question I keep failing to answer is what happens the day it produces a genuinely good program. I will come back to that.
What worries me more than any of the money
Every useful tool removes a task, and removing tasks is usually good. I do not do long division on paper. I do not remember phone numbers. I have not learned a route from memory in fifteen years, because the phone tells me where to turn. I am not going to pretend I am above any of it.
There is a difference between removing a task and removing the looking.
When you go after an answer properly, you find one article saying one thing, then another that disagrees. You read the study and realise the article misrepresented it. You pick up a term you did not know, which changes what you are able to search for next. Eventually you work out that the question you started with was the wrong question. Most of the learning happens inside that mess rather than at the end of it.
Prompting removes nearly all of it. Ask, receive, move on. Convenient, and I use it too, which is exactly why I am uneasy about it.
There is a second problem sitting inside the answer. Real knowledge has disagreement in it. Studies conflict. Sample sizes are small. Mechanisms get proposed and later abandoned. A good part of what I was taught about training twenty years ago has since been revised, and some of what I teach now will be revised too. That is what science is supposed to do.
An AI answer flattens all of that into a clean paragraph with no visible seams. The uncertainty is gone and what is left reads like settled fact. If enough people stop asking who says this, what the evidence is, and who disagrees, and ask only what the AI says, the relationship between people and knowledge changes shape, and not in a direction I like.
Who decides what the answer says
These systems are not neutral instruments floating above the world with access to objective truth. They are built by organisations. People choose what goes into the training data, what the rules are, which answers are acceptable, which are refused, which sources get retrieved and ranked, and how the system behaves when the evidence is genuinely contested. Those are decisions, made by someone, for reasons.
Nothing sinister has to be happening for that to be a problem. Commercial pressure exists. Political pressure exists. Bias exists, including bias nobody in the building noticed. Mistakes exist, and at this scale a mistake reaches millions of answers before anyone catches it.
Now put that beside the direction things are moving. A large share of the population is starting to get its answers from a small number of systems. When a handful of organisations sit between people and what they take to be true, the ownership and the incentives of those organisations stop being a technology story. I do not think we should be relaxed about it, and you do not need a conspiracy for it to matter.
So why am I using it
Because refusing achieved nothing, and then 2026 happened.
Take the first part on its own. Our organic search traffic dropped heavily. Not gradually, and not because we stopped publishing. The path that used to bring people to us narrowed while I stood there with my principle intact. That refusal did not bring back a single visitor, stop anybody generating a workout, protect one job, or cost any AI company a cent. The only measurable effect it had was on me.
My objection was mostly about other people, and it held up for as long as the damage was mostly happening to other people.
Then last year arrived. Personal and family issues I am not going to write about here. Our dog died. And I was told I have heart conditions I was born with, after a life of eating properly, not smoking, and training nearly every day. You can do everything right and still be handed something you had no say in whatsoever.
The vow did not wear down over that year. It ended in an afternoon. The thought was closer to “f— it” than to anything I could defend in an article.
Three things gave way at once, and I could not tell you which mattered most.
My priorities changed. I was no longer prepared to spend energy on a fight that had produced nothing in years, when there were things in front of me that actually needed it.
I could not afford the principle. The business had to work.
And I stopped believing the resistance had ever registered anywhere. Not with Google, not with any AI company, not with the people whose jobs I was worried about. It was a position held by one man in Greece, and the world had not noticed.
So I feel forced into it, in the economic sense. Nobody is standing behind me telling me to open a chat window. The environment my work exists in changed, and I either adapt to part of it or go down with the version of the internet I preferred. I would rather not have been put in that position.
How I actually use it
I do not sit down and type “write me an article about kettlebell training.” I have decades of material already: published articles, books, course scripts, coaching notes, positions I have argued and defended in public. I write the outline. I decide the structure and the claims. I pull the source material from what I have already written. Then AI helps me get it into readable shape.
Grammarly with a much longer reach. Same category of work, larger scale.
Then the real job starts, which is hunting for what it broke. It hallucinates. It removes a qualifier that was carrying the entire meaning of a sentence. It adds a mechanism I never claimed. It turns “may” into “will,” which in a subject involving injury and biomechanics is not a stylistic difference. It smooths a specific statement into generic marketing language. I correct it, explain what it got wrong, and get back another version that is wrong somewhere else.
There is a second use I did not anticipate, and it might be the most valuable one. I write directly. Too directly. I can come across as offensive when I did not intend to be, and defensive when nobody attacked me. Somebody says something that annoys me, I write the full reply, and I am one click from sending it. Now I stop and have it checked first. It will tell me I can make the same point without the part that starts a fight. It sits between what is in my head at that moment and the person on the other end, and it has saved me from myself more than once.
Related to that, I sometimes misread Reddit. Jargon, sarcasm, in-jokes, phrasing I take at face value when it was not meant that way. I can check what somebody probably meant before answering the version I imagined. That is not asking a machine what to think. It is making sure I understood a human before I reply to him.
AI helped me write this article, but “helped” makes the process sound far smoother than it was. The hardest part was the section you just read, and specifically the claims about what AI is doing to people’s traffic, their income and their work.
Almost every one of those started with the machine telling me I was wrong. Not after checking. It disagreed first, then I argued. I explained what I was seeing, pointed at what it had left out, and asked questions until it understood the claim I was actually making. And then it would turn around, tell me I was right, and produce the research proving it. Research it could have gone and found before it disagreed with me.
That happened repeatedly, on exactly the claims where the industry that built it is the one being accused. I am not going to tell you what was behind that, because I cannot see inside the thing and neither can anyone else. I can only tell you it took a fight to get the evidence out of it, and that the evidence was there the whole time.
Every claim here is one I will defend if you challenge it.
It does not always save me time
The productivity story is oversold, at least for the way I work.
If I trusted the first output it would save enormous amounts of time. Prompt, generate, copy, paste, publish. You could produce a hundred articles a week that way, and plenty of people are.
I do not trust the first output, particularly on subjects I know well, which are the only subjects I publish on. So I go through it line by line, take out the invented parts, put back the details it dropped, argue about a single word, and go around again. Finishing an article this way can take ten times longer than posting whatever came out first.
The result is better than what I would have written alone. The machine did not supply the idea. I had something to say, and it helped me say it more clearly than I usually manage.
The best part is that I can swear at it
I did not see this benefit coming.
I can call it every name I know. I can ask how the hell it arrived at something that stupid. I can tell it the last four attempts were garbage and that it has ignored the same instruction four times, which it has.
Nobody gets hurt. Nobody screenshots it. Nobody writes a post about how I treated them.
I shout at the machine, and then I speak to the human like a reasonable person. That is probably better for everyone than the arrangement I had before. I just hope this never gets to the Skynet part, and that there isn’t a log somewhere with my name at the top of it.
This is not just my opinion
I have made claims here about jobs, about traffic, and about influence. They are measured, so here is the measurement, including the parts that cut against me.
Jobs. The Stanford Digital Economy Lab, using ADP payroll data covering millions of US workers through June 2026, found no evidence of widespread economy-wide job displacement. That should be said plainly, because it contradicts the loudest version of this argument. What they did find is that employment among workers aged 22 to 25 in AI-exposed occupations now sits about 19% below where it would be had it kept pace with similar-aged workers in less exposed jobs. The gap has widened steadily since they first documented it in 2025. It runs through reduced hiring rather than firing, and it concentrates in occupations where AI automates the work rather than assisting it. The damage is landing at entry level, which is exactly where people get in.
Training on other people’s work. In June 2025 a US federal court ruled that training an AI model on lawfully acquired books can be fair use, and that downloading and keeping pirated copies is not. The resulting settlement, given final approval in July 2026, came to $1.5 billion covering 482,460 books, roughly $3,000 per work. As somebody with a shelf of his own titles, I have a direct interest in where that line ends up.
Search clicks. The Pew Research Center tracked the real browsing of 900 US adults across 68,879 Google searches in March 2025. On results pages with an AI summary, people clicked a normal result 8% of the time. Without a summary, 15%. Clicks on links inside the summary itself: 1%. Sessions ended on 26% of pages with a summary against 16% without. Google disputes the methodology and says total organic click volume has stayed roughly stable, and you should weigh that when you read the number.
Traffic, by publisher size. Chartbeat, measuring more than 2,500 publisher sites, found Google Search pageviews down 34% between December 2024 and December 2025. Split by size, small publishers running 1,000 to 10,000 daily pageviews lost 60% of their search referral traffic over two years. Mid-size lost 47%. Large publishers lost 22%. The smaller you are, the harder it lands. That data covers news and media publishers rather than fitness sites, so treat it as the closest available proxy for what independent publishers are living through, not as a measurement of us.
The traffic did not move somewhere else. Referrals from ChatGPT grew more than 200% over the same period, and all AI chatbots combined still account for under 1% of publisher referrals. Optimising for the chatbots does not recover what search stopped sending.
What publishers expect next. The Reuters Institute surveyed 280 media leaders across 51 countries and reported an average expectation of a further 43% traffic decline over three years.
Influence. Eleven of the largest technology and AI companies, together with their trade associations, spent $41.8 million lobbying the US federal government in the first half of 2026. That is over $226,000 a day, working out at roughly one lobbyist for every 1.5 members of Congress. Four years ago several of the AI companies near the top of that list had no federal lobbyists at all.
One more thing, since I am arguing for checking sources. Two of the items above involve Anthropic, and Anthropic makes the AI that helped me write this article. I did not leave them out for that reason, and I am not going to pretend the tool I used is a neutral party in its own story.
The question I cannot answer
I still do not like any of this. I would prefer the arrangement where a person who did the work had a reasonable chance of being found by the people who needed it, and where nobody’s livelihood was rearranged without being asked. That version is going, and my opinion of it changes nothing.
What I have landed on is narrow. Use it to investigate, to organise, to find the weak point in your own argument, to say clearly something you already understand. Stay responsible for the result, because when my name is on an article I am the one who has to defend it, and “the AI wrote that bit” is not a defence I hope I have to give.
There is still a question underneath all of this that I have not resolved.
Right now I can tell you not to trust an AI-generated kettlebell program. That argument only works while the programs are flawed, and I have no reason to assume they will stay flawed.
So what do I say the day it produces a genuinely good one? Well structured, correctly progressed, appropriate for the person asking, free. Do I still tell people they should learn from someone who spent twenty years earning that knowledge? Is that a real argument at that point, or is it me protecting my own position?
The bigger version is the one I actually lose sleep over, and it is the same question asked of every trade at once. When the answer is always available before the argument starts, where do people get the practice of finding out, being wrong, and changing their mind? Who still bothers to learn something properly, when learning takes years and asking takes seconds? What happens to the income of every person who spent a life becoming good at one thing, and to the ones who were never given the chance to start? What happens to a society where a large share of what people believe arrives already formed, from a handful of systems owned by a handful of companies?
Nobody knows. The people building it do not know either, which is worth sitting with, because they are building it anyway.
I would like to be wrong about all of it. So far I have not seen much that suggests I am.
References
- Brynjolfsson, E., Chandar, B., & Chen, R. Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab, revised August 2026.
- The Authors Guild. Bartz v. Anthropic settlement: what authors need to know.
- JURIST. Judge approves record $1.5 billion AI copyright settlement involving Anthropic, July 2026.
- Pew Research Center. Google users are less likely to click on links when an AI summary appears in the results, 22 July 2025.
- Chartbeat. Pageviews are down, but AI’s impact is complicated.
- Axios. Small publishers hit hardest by search traffic declines, 17 March 2026.
- Search Engine Journal. Search referral traffic down 60% for small publishers, data shows, reporting the same Chartbeat breakdown.
- Press Gazette. Global publisher Google traffic dropped by a third in 2025, reporting the Reuters Institute Journalism and Technology Trends and Predictions 2026.
- Issue One. Lobbying disclosures reveal Big Tech spends more than $226,000 per day to buy influence, 21 July 2026.