Some things are supposed to be hard.
Take learning. Any attempt to deepen your knowledge or improve your capabilities reaches a point where you have to really grind to reach the next degree of mastery. You are changing the structure of your brain1 as you engage in sustained, hard learning over time, so whether your goal is being able to code, sell, cook or cartwheel, part of the process is inevitably going to suck.
Even something you enjoyed from the first moment went through phases that demanded more from you than you felt you could give. But if you loved it — or had a strong enough stimulus in the form of wanting to pass, excel or be able to continue buying food — you put your head down and pushed through it.
Can you think of anything you're genuinely good at now that didn't go through a phase like that? I can't.
The same used to be true of every human endeavour, but we're living in an age where — in the digital realm at least — "mastery" can be publicly demonstrated without being earned.
When AI tools enable us to write symphonies without knowing the difference between a crotchet and a quaver, or build fully functional apps and great-looking websites with a few prompts, is all of the effort we've become so accustomed to still worth it?
Progress?
Since AI figured out how to generate images of hands with only five fingers all in the right places, we've been throwing our problems at the binary ether and marvelling at how quickly we walk away with useable outcomes. From generating email drafts and websites to designing a complete genome for a brand new virus2 (I'm sure that's not going to go wrong in any way), AI is removing friction in a manner that makes the possibilities seem endless.
If it helps us achieve greater things faster and cheaper, that's progress, right?
In some ways, yes, absolutely.
But the more friction we outsource to AI, the less competent we get.
I can’t help but think of this scene as I write this.
By less competent I don't mean that we'll achieve less; if anything, the opposite is true.
As AI dissolves the timelines of discovery and we hurtle into a future we've only previously been able to imagine, process is being traded for progress in ways that are already starting to cost us.3(1)
The aforementioned biological apocalypse aside, there are plenty of other ways we're potentially building our own AI-shaped pyre:
- employment4a
- economic inequality4b
- ecological damage4c
- labour exploitation4d
- automated discrimination4e
- critical judgement4f
- human connection4g
- shared knowledge4h
- trust in information4i
- creative diversity4j
We don't yet know the total bill humanity will pay for our race towards ever-better AI. But in some ways, it is already starting to come due.
Calculators
Progress has historically been the domain of people standing on the shoulders of others before them, reaching beyond the limits of their field and into the future. But those shoulders are starting to look distinctly binary, and the hard work is, more often than we'd like to admit, becoming a matter of figuring out the right set of skills, .md files and prompting away until we get a result that looks good enough.
This moment in history is comparable to the advent of the calculator,5 the personal computer6 and the internet.7 All three technologies enabled us to offload arduous mounds of mental computation and procedural drudgery to technology, and were met with cries of panic similar to what we're hearing today.
And what people feared it would take with it.
The calculator and the computer reduced how much time we spent on manual work, but we still needed to understand what we were trying to achieve — and often have access to expertise from different disciplines if the project was complex enough — to deliver something we could be proud of.
The internet gave us instant connection and access to oceans of information at whim, but we still had to understand, compile and parse it into something useful.
AI is following a similar pattern, but as it has improved from silly pirate poems and seven-fingered hands to the primary infrastructure behind trillion-dollar valuated companies, there is one fundamental difference that sets it apart from the previous revolutions.
Expertise, time and effort are no longer prerequisites to creating incredible things.
We still rely on the hard work of those before us.(2) But instead of hours of conversation, compiling sources, making and fixing mistakes and finally synthesising something new — often with a list of citations a mile long — a lot of new, noteworthy work coming out today is significantly lighter on attribution. And briefly mentioned somewhere in passing is the inevitable tagline:
"Built with AI."
That we are standing on the shoulders of other humans who poured their lives and various fluids9 into figuring things out hasn't changed.
However, we no longer always know who they are, or how much of their work was gobbled and regurgitated by the machine to give us the sleek summaries and solutions we use without a second thought.
Calculators, But Bigger
The scale matters, obviously. I'm not going to drop a prompt into Claude and wake up to a million-dollar company tomorrow morning (although building the entire tech stack behind one is no longer outside the realm of possibility). But the argument still applies as complexity increases.
Industrial R&D and materials science investigations where source material spans multiple languages? From months to minutes.10
Predicting nearly every known protein structure11 on earth? From many months or years of PhD and technician labour per structure to a matter of seconds12 (and making it all structured and searchable to boot).
by the company’s own account
timings from Google DeepMind12
the counterweight
The higher-level organisational competencies required to operate at this scale haven't been devoured by AI.
Yet.
The component units of work — the sourcing, the research, the compiling, the synthesis, the drafting, the designing — are all slowly moving from human-only, through human-guided, to completely autonomous.
If you'd told people in 2023 that by 2026 AI would match human quality on 41% of a benchmark of real-world, economically valuable freelance projects (at the time of writing), many of them would have laughed at you.
What will 2028 look like? 2030? 2050? We can't know for certain.
What we can do is focus on what we are able to control, and be intentional about protecting the competencies that humanity has earned over millennia.
understanding != outcome
People have outsourced work for centuries. From buildings to Boeings, waterways to websites, part and parcel of the human experience was knowing who to contract to get what you wanted, and finding the best value for what you could afford, even if it meant settling for less than you would originally have liked.
Doing so usually made you an owner of the outcome, but that was all. Commissioning a stadium didn't suddenly make you an architect or engineer. But you did have to contract them if you wanted something that would survive the first heavy wind.
The difference now is that getting something you want costs little beyond your token spend.
The explicit dependency between needing to understand and being able to build is eroding.
I no longer have to pay for a handful of expensive software subscriptions and several staff members to run my business; all I need is Claude, discipline and patience. That doesn't make me an accountant, a data analyst or a sysadmin, but it does mean I can add just about any role to my business that I want by waffling at Claude in broad arcs, and "own" whatever comes out of it as though I were one.(3)
An unfortunate offshoot of this is epistemic trespassing, where someone with expertise in one field mistakenly thinks that it qualifies them to speak with authority in another.
Using Claude as the engine to help me run my company might give me the confidence to advise my friends on how to organise their business finances, but the outcome would probably be humbling at best, if not downright catastrophic.
Finding some measure of success through AI in a constrained context outside of my earned expertise isn't a qualification to lead the room in that context.
Refinement and collaboration are still part of the process when creating something you want other people to use. But as the competence of AI models increases with each new release, enthusiasm is starting to become a convincing substitute for expertise.
Then and now.
Cognitive Offloading
Cognitive offloading is when you outsource the mental energy required for a task to one or more tools. We've been doing this for aeons, from hieroglyphs describing the dangers of predators to diaries reminding us of our wedding anniversaries (and the relative peril of ignoring either).
It is a normal, necessary part of being human in a complex world, and AI hasn't changed that. But it has shifted the scale.
We are now able to practise cognitive offloading to such a degree that cognitive atrophy18 is a real, measured risk: the degrading of the ability to remember, think critically and develop competence in skills that require hard work.(4)
A delightfully ironic example of this was South Africa’s Draft National AI Policy being withdrawn shortly after being gazetted.
Not because of the usual political shenanigans, or a power play by AI labs with vested interests.
Instead, the team drafting it was found to have referenced entirely fictitious sources,19 most likely generated by AI.
Ah, South Africa. When we shoot ourselves in the foot, we certainly do it with style.
As adults today, we've had the benefit of decades of being forced to think and remember for ourselves. Our children and future generations are growing up in a different world.
Why exert yourself doing homework, or wrestle with a particularly challenging topic for weeks before it finally clicks, when you can pop it into AI and have a set of answers in front of you within seconds?
The Dissolution of Effort
The hard work of developing new skills — be it developmental arithmetic or how to remove a tumour from a brain — is how we reach through our limitations, discover new horizons of potential and achieve mastery. There will be parts we dislike and parts that come naturally, but discovering the difference has always been bound to the process of doing the work.
But for anything in the realm of knowledge, intellect and digital output, that effort has started shifting from a requirement to an option.(5)
Thankfully, many parts of the human experience are still immune to this effect.
From the sweaty risk of a first date to lacing up your exercise shoes when you really don't want to, effort remains the only reliable way to achieve some things that no amount of AI will be able to help you with.
Yet 🫠.
Learning is a great example of this. The first time I used AI to explain something to me, it was a proper "I know Kung Fu20" moment. But there is a world of difference between knowing the basics because you read a good summary, and understanding the nuances that allow you to apply that knowledge in a way that is useful beyond the one example ChatGPT uses to illustrate the concept.
How well do you understand how a zip works?
Explain it, step by step. What happens inside when you pull the slider up?
Now: how well do you understand how a zip works?
You went from 6 to 3.
Most people drop. The gap has a name: the illusion of explanatory depth.21
Shortcuts
"Work smarter, not harder" is a laudable goal. But being able to work smart is almost always a product of the hard work that helped you figure out where the dumb, unnecessary work was, and how to avoid it.
To paraphrase Shane Parrish, "you can't shortcut your way to shortcuts."22
That is why it is worth embracing friction.
It isn't about making things unnecessarily hard, or engaging in busywork that can and should be automated.
If you want quick results in a realm you already know well, shortcuts are great.
But if the ideal outcome depends on earned expertise and due process, shortcuts are likely to work against you.
Doing the hard parts is sometimes the only way to get to the good parts.
Embracing Friction
It is tempting to wring our hands, bemoan the new devil of technology and wish for an earlier age. AI isn't inherently good or bad. It is simply here, and as humanity has proven over and again, we can and will adapt.
Looking for ways to add friction feels counterintuitive because it means being disciplined enough to reject the path of least resistance. And minimising effort — especially where it doesn't appear to have any benefit23 — is how most of us approach life.(6)
It's no coincidence that your house, desk or computer has at least one corner that is just an absolute bloody mess.
It's not that you couldn't keep it tidy if you wanted to, but the effort you'd have to expend to do so usually feels like more than it is worth.
The optimist's promise about AI — echoing through history about each new technology that automates labour — is that we will get to do less work.
But if you've spent any time ranting at your LLM and fixing the same mistake over and over, you know how many ways that isn't true.
AI doesn't reduce the amount of work we have to do.24 It simply changes the shape of it.
Attention Demand Disorder
A wealth of information creates a poverty of attention.25
Herbert Simon, 1971
Once I discovered how much of my work I could automate using Claude, I was hooked. Overnight I went from focusing on one thing at a time to working on as many projects concurrently as my token cap allowed.
But after a couple of brutal days working at a relentless pace without getting anywhere, I hit a wall.26
The more workstreams I had open, the less attention I could give to each, which meant I was both context switching at an unhealthy rate, and farming out more and more decision-making responsibility to Claude.
The bottleneck for putting out the calibre of work I was happy with wasn't the technology. It was my attention.
Doing "just one more task" doesn't come without cost.
It took dropping a few important balls, realising I was barely spending quality time with my wife and kids, and hitting "brain fry"27 — that mental fugue one gets when tweaking on LLMs — several days in a row to realise I needed a different approach.
Mental fatigue from using or overseeing AI beyond your capacity.28
Blocky
So I built a hook into Claude (which I affectionately call Blocky) with two modes. I was allowed no more than
- three workstreams on a build day, and
- five workstreams on an admin day.
If I exceeded that limit — even accidentally — Claude literally downed tools and said "Nope, not doing another thing until you drop your workstreams back within cap".
As frustrating as it was at times, Blocky provided two invaluable types of friction:
- I had to constantly evaluate whether the work I was engaged in was the best way I could direct my focus, especially on days with multiple deliverables for different clients.
- Knowing I could turn it off at any time forced me to reflect on what mattered most: sneaking an extra workstream or two (at the cost of splitting my focus even further), or staying in the process to deliver my best work, and having the energy to be human at the end of the day.(7)
you wantfights what
you want SOFTa reminder17:00 wind-downworksa warningdecays HARDa blockoverkillBlocky,
the scoping gateholds
Friction can be soft when it fits what you already want, but needs to be harder when it has to fight your natural tendencies.
Claude warns me every turn from 16:00 to 17:00 that I need to start wrapping up, and that is enough because it reminds me of something I am striving for every day anyway.
Blocky was harder because it stopped me from defaulting into patterns that I know aren't healthy but tend to fall into anyway.
The alluring call of "But AI can do it while you're doing other stuff" is relentless, but I postponed plenty of work I would have liked to do — and probably could have done — now because of Blocky.
It became the mechanism that helped me say, "No, I want to do this one thing properly, and be present for myself and my family".
I haven't hit brain fry once since I started using it.
Suspending Blocky
Around the middle of October I added a "suspend" mode. With the pile of end-of-year deadlines on top of the usual day-to-day stuff, a cap of five workstreams had started creating more anxiety than it resolved.
Blocky still defaults back to admin mode at 03:00 each morning, and I have to manually switch to suspended every time I want to activate it.
I was a little worried about what would happen, but after a few fried-brain-free weeks — working in suspend mode most of every day — I started wondering whether I should drop Blocky completely.
I decided against it.
Retaining the friction of having to intentionally switch is a valuable reminder that my attention is the most important resource I have.
The hard limit lifted, but the awareness and behaviour it helped me develop remain.
Rebuilt from my time archive.
The Luxury of Friction
I'm aware that I'm in a fortunate position. My newsfeed is full of impassioned pleas to senior leadership about the expectation that just because AI can, it obviously should:
- Senior engineers working 13-hour days to achieve the AI-fuelled ultra-productivity that is the vogue, only instead of writing code, they're just pressing enter.29
- Teams faced with delivering on hyper-polished AI boardroom pitches30 that never had to answer, "Is this realistic?"
- Teams expected to deliver immediate ROI on AI investments31 at the expense of the necessary rigour.
And the list goes on.
Maybe you don't feel like you have the luxury of friction in order to keep your job, or survive in a competitive market. Or maybe you've gotten so used to AI doing your work for you that you can no longer imagine your life without it.
I get it. Truly.
I've gone from being stressed and anxious about when I'd next drop a ball or miss some critical bit of admin, to delivering more work at a higher standard and being able to spend more time with my kids, not less, thanks to how I've leveraged AI in my business.
But.
Build the World You Want to Live In
The world I want to live in is one where I own every single pixel, word and character that comes out of my business, and with each passing month the quality of everything I do continues to rise.
Is one where I meet and exceed my own standards for living, while becoming a more engaged husband and father.
Is one where I am constantly exercising curiosity, practising compassion and discovering new ways to live a richer, fuller life.
AI is enabling that by automating away hours of admin and repetitive drudgery, which I can now invest in work I love doing and people I love being with.
Embracing friction is not about avoiding AI or wasting time on tedious, repetitive tasks that can and should be automated.
It's about slowing down, taking stock, and reflecting on what matters to you in the long term.
Then developing the disciplines and processes that will help you build the world you want to live in.
Hard parts and all.
Postscript
This piece is an apt synecdoche for what I'm hoping to convey.
It took me the best part of three months to finish, and every time I sat down to write I had to:
- commit to the process instead of doing other client work at the same time, and
- honour my commitment to not letting AI do any of the writing for me,32 despite how much faster it would have been.33
Facing down the blank pages, condensing the chaos in my head into a coherent structure, and chopping out bits I really liked but didn't fit was a real challenge.
But seeing this live is deeply rewarding. And as a bonus, I developed two awesome new tools in the process that will make the next time even better.
If this resonated with you and you'd like to find out more about building friction into your work, or about making the most of AI in your work environment, please get in touch. I'd love to hear from you. Book a free consultation here.
Notes
- 1Current Biology, 2011: Woollett and Maguire, on how learning London’s streets changed taxi drivers’ brains. doi.org ↩
- 2BBC News, on a virus genome designed with AI. bbc.com ↩
- 3Design Science (Cambridge): the hidden costs of AI-assisted ideation for novice designers. cambridge.orgAmerican Council on Science and Health, 1 April 2025: “When AI takes over: the hidden cost of technological progress”. acsh.org“AI assistance reduces persistence”, research project site. ai-project-website.github.io ↩
- 4The wider stakes, one source per cost. Each points strongly one way; none of them settles the question.aEmployment: Brynjolfsson, Chandar and Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence”, Stanford Digital Economy Lab, 2025. A working paper built on payroll data. Over the same period, the Yale Budget Lab found no discernible disruption across the wider US labour market.bEconomic inequality: Cazzaniga et al., “Gen-AI: Artificial Intelligence and the Future of Work”, IMF Staff Discussion Note, 2024. A staff note that models exposure rather than measuring outcomes, and it isn’t peer-reviewed. A large call-centre study found AI helped the least experienced workers most, which would narrow some gaps (Brynjolfsson, Li and Raymond, Quarterly Journal of Economics, 2025).cEcological damage: Alex de Vries-Gao, “The carbon and water footprints of data centers and what this could mean for artificial intelligence”, Patterns, 2026. Peer-reviewed, with wide ranges because AI companies don’t disclose their figures. The IEA expects data centres to account for around a tenth of global electricity demand growth to 2030.dLabour exploitation: Muldoon, Graham and Cant, Feeding the Machine: The Hidden Human Labour Powering AI, 2024. Built on more than a decade of interviews and fieldwork rather than a survey. Conditions vary by company: in Fairwork’s own ratings, one data-work firm scored 6 out of 10 after making changes.eAutomated discrimination: Hofmann, Kalluri, Jurafsky and King, “AI generates covertly racist decisions about people based on their dialect”, Nature, 2024. Peer-reviewed, and tested in controlled prompts rather than deployed systems. The same study found the models’ overt stereotypes are far more positive than their covert ones.fCritical judgement: Lee et al., “The Impact of Generative AI on Critical Thinking”, CHI 2025, Microsoft Research and Carnegie Mellon. Peer-reviewed, but self-reported by 319 knowledge workers, so it measures how people judge their own effort. The same paper found critical thinking shifting toward checking and integrating AI output.gHuman connection: Fang et al., “How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use”, MIT Media Lab and OpenAI, 2025. A four-week controlled study of 981 people, released as a preprint. The finding is a correlation, so it can’t tell whether heavy use causes loneliness or lonely people use chatbots more.hShared knowledge: del Rio-Chanona, Laurentsyeva and Wachs, “Large language models reduce public knowledge sharing on online Q&A platforms”, PNAS Nexus, 2024. Peer-reviewed, and about one platform: Stack Overflow posts fell by about a quarter within six months of ChatGPT’s release, compared with similar sites. The authors note the questions may have moved into private chats with AI.iTrust in information: Spitale, Biller-Andorno and Germani, “AI model GPT-3 (dis)informs us better than humans”, Science Advances, 2023. A preregistered study of 697 people, who couldn’t tell AI-written tweets from human ones. Simon, Altay and Mercier argue the fears are overblown, because most people see little misinformation in the first place.jCreative diversity: Doshi and Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content”, Science Advances, 2024. Peer-reviewed, from an online short-story experiment. The same study found AI ideas made individual stories better, most of all for less creative writers. ↩
- 5The Christian Science Monitor, 9 May 1986, on calculators in the classroom. csmonitor.com ↩
- 6Nextgov, March 2015: “Remember computerphobia? When people feared PCs”. nextgov.com ↩
- 7Nicholas Carr, “Is Google Making Us Stupid?”, The Atlantic, 2008. theatlantic.com ↩
- 8Plato, Phaedrus, c. 370 BC: writing will bring “forgetfulness in the learners’ souls”. ↩
- 9Allison Futterman, “These Five Doctors Experimented on Themselves — And Made Big Breakthroughs”, Discover, 14 October 2021. discovermagazine.com ↩
- 10The timing comes from ENEOS Materials’ account, when applying ChatGPT to the Hungarian-language sources. The published case study leaves open how it was measured and whether it included checking the results. Here's the OpenAI case study for reference. ↩
- 11The Nobel Prize in Chemistry 2024, press release. nobelprize.org ↩
- 12Google DeepMind, “AlphaFold reveals the structure of the protein universe”, 28 July 2022: determining a protein’s 3D structure “used to take many months or years, it now takes seconds”. deepmind.google ↩
- 13OpenAI, the ENEOS customer story (the company’s own account). openai.com ↩
- 14Chen et al., 2024: Microsoft and Pacific Northwest National Laboratory, peer-reviewed. ↩
- 15Bastani et al., PNAS, 2025: Turkish high-school pupils given GPT-4. ↩
- 16The Lancet Gastroenterology & Hepatology, 2025: adenoma detection at four Polish endoscopy centres. ↩
- 17Strömberg, Lei and Wu, 2026: about 27,000 Chinese secondary pupils. ↩
- 18Polytechnique Insights: “Generative AI: the risk of cognitive atrophy”. polytechnique-insights.com ↩
- 19SAnews, 26 April 2026: the Minister withdraws the Draft National AI Policy. sanews.gov.za ↩
- 20“I know Kung Fu”, The Matrix (1999). ↩
I know Kung Fu - 21Rozenblit and Keil, 2002, Cognitive Science: the illusion of explanatory depth. ↩
- 22Shane Parrish, Brain Food, 28 December 2025, Farnam Street: "You have to put in the hours before you can see the shortcuts." ↩
- 23Neuroscience News: people avoid effort that looks wasted. neurosciencenews.com ↩
- 24Ranganathan and Ye, AI Doesn't Reduce Work—It Intensifies It, Harvard Business Review, 2026. ↩
- 25Herbert Simon, 1971. ↩
- 26Lisanne Bainbridge, "Ironies of Automation", Automatica, 1983. ↩
- 27The American Journal of Medicine, 2026, on “brain fry”. amjmed.com ↩
- 28BCG survey of 1,488 US workers, reported in Harvard Business Review, March 2026. ↩
- 29A post on X (@v0xium). x.com ↩
- 30Harvard Business Review, September 2026: “Research: Gen AI is collapsing creative processes”. hbr.org ↩
- 31Forbes, 26 January 2026: 71% of tech chiefs say leadership has unrealistic AI expectations. forbes.com ↩
- 32“Keep Your Dead Ends”, an earlier essay on this site. trustdavid.co.za ↩
- 33Claude did help me with some of the research, the final copy edit and the heavy lifting when it came to developing the graphics. But that's where it excels, and without that I might well have spent another three months muddling about instead of finally getting this into the world. ↩