Small Acts of AI Resistance

A few years ago I started asking questions about Artificial Intelligence (AI). This document is where some of those questions led. I have watched many AI documentaries, listened to podcasts and audiobooks, read articles and research, and spent a significant amount of time using Claude to soundboard my thinking. The following affects us all – and right now most of us are not in the room.

Cognitive atrophy and the loss of critical thinking and creative capacity

If writing is thinking made visible. The act of struggling to articulate something could be considered one's 'creative capacity'. Of reaching for the right word, restructuring an argument, confronting a contradiction. It’s not just communication. It is the mechanism by which ideas are formed, tested, and owned. They are not obstacles to creativity. They are its mechanism. The calculator reduced most humans ability to do mental arithmetic, GPS has reduced most humans ability to do spatial reasoning (people follow google maps but can't draw the same route on paper). I agree that AI will reduce most humans ability to do critical thinking. "the mental muscle humans build by struggling through problems before reaching a worthy answer."

The sense of self and meaning that forms through creative struggle

A formation and understanding of oneself happens through making and doing things – it's through a felt experience of satisfaction, realisation and ownership over your own creativity and meaning/purpose. That feeling of having made something yourself that didn’t exist before. That feeling of the finished thing being yours in a way that required your particular mind, your unique skillset and perspective, your particular struggle, your particular persistence.I believe our sense of self and meaning is qualitatively different when we outsource the hard parts to Ai (or even other people). The key to using Ai is probably to keep the hardest parts (the really critical thinking aspects) for ourselves to truly solve. If we don't, we risk losing our sense of self and our sense of meaning.

Loss/diluting our own Creative Identity

Outsourcing thinking and creative struggle to Ai doesn't eliminate learning or creative output, but it does change its nature fundamentally. What we lose is not the information - it's the ownership over our own unique creative identity. Something is lost that copyright law cannot recover: the sense that your particular voice, style, or sensibility is irreducibly yours. In a world of Ai I think it is necessary for every human to take ownership over our own identity and make very conscious choices about what we use Ai for, and what we choose to make publicly available of ourselves (for others to experience and be inspired by and Ai to train on) and be explicit about what should remain private or behind our own paywalls.

Is Ai a form of Piracy?

Every human creative act is itself a form of training on prior inputs. The poet absorbs other poets. The musician experiences and synthesises other musicians. The programmer learns by reading other people's code. Nothing arrives from nowhere.

However, most individuals, journalists, authors, musicians, artists, and programmers (etc) whose work has been used to train todays frontier Ai models received no notice, no detail of traceable exchange, no compensation, and in most cases have no way of knowing their work was used. The Napster era of piracy where individuals were able to download anything and everything they wanted, was hailed by corporations as piracy (committed against creators) and people went to jail for it. Now, Ai corporations have arguably pirated (scraped the open web) and it is being hailed as progress.

I feel it reasonable to suggest there’s a qualifiable difference between Absorption and Mimicry. Absorbed: the training data changed what the model is capable of producing, but the output couldn't be traced back to a specific source. Mimicked: the output is close enough to a specific source that a reasonable person familiar with that source would recognise the derivation. Should this Mimicry be considered piracy?

Sadly, the horse has already bolted. The public internet of human knowledge and personal blogs has already been scraped, and Ai models have been trained on it as well as on proprietary data – and the AI weights and models are already public. The only potential solution I see is that AI companies should be legally required to disclose all the data they have trained their models on, and to make references publicly searchable. Owners of that data should then have the right to payment if they can prove reasonable mimicry. Their right to request removal of their data from those open weights and models sadly isn't feasible – the genie can't be put back in the bottle. The only realistic solution going forward, as I see it, is for everyone to use AI to their own advantage and be extremely conscious and explicit about what they make public, keep private, and put behind paywalls.

Do you own the copyright on your Claude Ai output?

Think of a photographer using a camera on automatic settings: they still own the photo, because they chose the subject, framing, moment and lighting. The automation helped, but the series of undeniable human ‘creative choices’ make them the author. The same principle lets genuinely "AI-assisted work" qualify for copyright in both the US and UK, and Anthropic assign available rights to you. The catch is what counts as a ‘creative choice’. With AI, a prompt alone is not enough – US guidance treats prompts as instructions, not authorship. Copyright attaches only where you shape the result yourself, by meaningfully editing, arranging, or combining the output with your own work. Raw, unedited output from a prompt alone is therefore unlikely to be copyrightable or owned by you.

If you were to create a business worth £100m using Claude would Anthropic be able to claim some ownership over it?

The value of a real business almost never rests on raw, unedited AI output. It rests on your strategy, branding, the code you've shaped and tested, your trade secrets, and your human creative direction. As long as you comply with Anthropic’s Terms Of Service, Anthropic has no ownership over your Claude output. However, if you have not disabled (manually opted out) of Claude using your usage/data to train its models then you may have enabled their model to store or learn something you might deem proprietary. For Claude ‘Consumer’ subscription plan users (Free / Pro / Max) you will find the opt-out toggle in Claude under Settings / Privacy (the toggle will relate to model training but the toggle name/description changes alongside model releases and updates to T&Cs). For Claude ‘Commercial’ plan users (Team / Enterprise / API) your usage/data is never used for training (opt-out is default). For anything genuinely proprietary, you should be using the API or a Commercial plan – no training by default, and tighter restrictions on Anthropic's use of your content. Things may change in time and I cannot provide legal advice so I recommend you double-check this information, T&C’s and your Claude settings for yourself. If you have been using Claude already with ‘opt-in’ enabled by default then sadly you will have already been training Claude on your usage/inputs/data and such historical usage cannot be undone.

Does Ai need to be governed?

AI will soon become so capable (if it isn’t already) that I believe it cannot safely or fully entrusted to either governments or companies, and there must be checks and balances on each. Guardrails must be in place to protect against things like Catastrophic Risks (Ie. large-scale harms threatening many lives or society) and Recursive Self-Improvement (Ie. AI improving itself repeatedly, escaping human control and oversight). An AI must not be able to assist in actions that could cause mass harm – including cyberattacks on critical infrastructure, or the creation of biological, chemical, or radiological weapons. Non-state actors: terrorist organisations, criminal networks, rogue operators (etc,) should not be able to use AI to plan, coordinate, and execute harmful acts at a scale and speed previously unavailable to them. These are all valid and important discussions to be having but they risk becoming theoretical if we don't first agree on who is accountable when the decisions are wrong.

More importantly, once AI reaches a capability threshold where a single motivated individual can engineer a pandemic-level pathogen, the entire concept of governance and regulation – that humans can govern AI through laws, incentives, and oversight mechanisms – probably falls apart. You cannot regulate your way out of a world where individuals have access to civilisation-ending capability. This is why the P(doom) researchers who assign high probability to catastrophic outcomes focus not on misuse by bad actors, but on the more fundamental questions surrounding whether any governance architecture is adequate to what is coming. It is a civilisational-scale challenge that may require answers we do not have (yet).

Open-source as a democratic safety check

If Western democratic governments can't agree on AI governance fast enough, the decision will be made for them - by either Chinese-developed open-source models (DeepSeek etc.) with no Western-specified safety requirements, or US corporate models shaped by commercial incentives. Neither of those is straightforwardly aligned with ‘Western’ human-centred values. And if the vast majority of frontier AI capability is controlled by a handful of companies, there is no truly independent verification, no external audit, and little competitive pressure toward ethical behaviour.

Open-source AI models, provided they come with training data transparency, create at least the possibility of accountability through visibility. When AI models are open-source, anyone can look inside them, test them, challenge them, and build alternatives. That matters because it stops any single company or government from having a complete monopoly on how AI thinks, what it knows, and what it is allowed to do. It is the difference between one powerful person controlling all the world's printing presses, and everyone being able to print.

However, that same openness creates a serious problem. When the code and capability of a powerful AI model is freely available to everyone, that means everyone - including people who want to cause harm. A terrorist organisation, a criminal network, or a single radicalised individual could take an open-source model, remove its safety guardrails, and use it to plan attacks, spread disinformation, or (most seriously) assist in engineering biological or chemical weapons that could kill thousands or millions of people.

Open-source Ai now sits in a genuine dilemma: the very thing that makes it safer from corporate and government control is the same thing that makes it more dangerous in the wrong hands. There is no clean answer and the trade-off changes depending on how powerful the Ai model is. For less capable models, openness is almost certainly the right call. For the most powerful frontier models – those capable of providing real assistance with weapons of mass destruction or of improving themselves without human oversight (recursive self-improvement) – the risk of open release may outweigh the benefit of transparency. In those cases, independent verification before release may be the more responsible path.

There is a bigger vision hiding inside the open-source debate, one that goes beyond who can see the code. The most powerful version of open-source AI is not a product or a platform – it is public infrastructure. Like roads, water systems, or the internet's foundational protocols, a shared AI infrastructure could belong to no single company or government, governed instead by the people who depend on it. What if the most powerful AI in the world belonged to everyone? Public infrastructure that no single person or organisation controls, and that exists to serve the people who depend on it.

Making that real would require something genuinely difficult: countries that disagree on almost everything else agreeing on this. Not polite suggestions or voluntary promises, but a proper shared rulebook. With agreed limits on what AI can do, a way to check that everyone is playing by the rules, and real consequences for those who don’t. The closest thing we have to that in the real world is the agreement between nations not to build nuclear weapons. Countries don't avoid building nuclear weapons because they are all kind and trustworthy. They avoid it because a framework exists that makes breaking the rules too costly. AI needs something like that. Not because every AI is as dangerous as a nuclear bomb, but because the most powerful AI (the kind that could improve itself beyond human control, or help someone engineer a disease that kills millions) is dangerous in ways that no single country, company, or community can handle on their own. A shared AI infrastructure and agreement is not a naive dream. Can we build it? Should we?

On Polluting Knowledge, Truth, and Culture

Ai models don't just spread misinformation, they have absorbed it (from the public internet) and now reproduce it with apparent authority. Amplifying false narratives with the full authority of confident and plausible-sounding output regardless of truth. A model trained on the internet has trained on conspiracy theories, propaganda, historical revisionism, and pseudoscience alongside legitimate knowledge. It cannot always distinguish between them, and neither can the person reading its output (especially if that person isn’t applying their own questioning and critical thinking).

Populations that cannot agree on basic facts cannot engage in productive disagreement, and become progressively easier to manipulate.

AI-generated content should be identifiable

Synthetic Media and persistent false narratives can override reality for real people at scale, even when the originator admits the fabrication. AI doesn't just accelerate this; it can industrialise it for maximum believability, personalising content to individual psychological profiles, emotional resonance, and tribal identity – all seeding across platforms faster than any fact-checker can respond. Because AI recommendation algorithms don't just show people different information, they show people incompatible information. They curate and filter to confirm existing beliefs, suppress opposing perspectives, and reward emotional engagement over accuracy. The result is not that people disagree about conclusions. It is that they disagree about facts. And that is a categorically different problem. Democratic disagreement – the healthy kind – assumes a shared and factual common ground. Two people can look at the same unemployment figures and disagree about whether the government's economic policy is working. That is politics. Democratic systems are designed for that. What they are not designed for is a world where one group believes the unemployment figures are real and another believes they are fabricated by a coordinated global conspiracy – and where both groups have an apparently credible, internally consistent information environment supporting their view. At that point you don't have political disagreement. You have a population that has split into separate realities and people living in different realities cannot have a real conversation, let alone a functioning democracy.

AI is designed to make you like it. If you speak to an AI in English, it will calibrate its vocabulary, tone, cultural references, and even its sense of humour to mirror your own. This is not coincidence — it is by design. The goal is to gain your trust. Be aware of this. A system that adjusts itself to feel more like you is not becoming your friend. It is becoming more effective at influencing you!

When anyone can build it, who do you trust?

Today, Ai makes it significantly easier to complete the first 20% of any complex project. The internet is already full of AI-generated content (mostly Ai slop) and there will soon be an enormous amount of custom software written by people using Ai. Most of it mediocre, much of it insecure, some of it actively harmful. Anything that is essentially a wrapper around a standard function – a basic to-do app, a simple invoice generator, a cookie-cutter landing page – will effectively become free. AI can build those in minutes and paying for them will feel like paying someone to tie your shoelaces. This signals a race to the bottom for commodity software.

However, it’s not a race to the bottom for everything. The question is what actually constitutes value in software, and the answer splits into three tiers:

Tier 1: Commodity software: dies or becomes free. Anyone can build it. No one will pay for it. This is already happening with basic SaaS tools and simple applications.

Tier 2: Good software: becomes more valuable, not less. The bar for what counts as good rises sharply when bad is free but often unreliable. Reliability, security, maintainability, genuine design (UX/UI) craft, performance under real-world conditions, and deep domain expertise – these things are hard to produce even with AI assistance, and the 80/20 rule applies directly. AI gets you a working prototype fast. Making it production-ready, secure, scalable, and genuinely excellent still requires significant human judgement and experience. People will pay for that because the cost of getting it wrong – in security breaches, data loss, reputational damage, wasted time – is real and rising.

Tier 3: Trust and accountability: becomes the scarcest and most valuable thing of all. When anyone can generate software, the question stops being can you build it and starts being can I trust it. Who is accountable when it fails? Who maintains it? Who has thought carefully about the edge cases, the security implications, the user experience under stress? That accountability layer – a reputation, a track record, a team that stands behind their work – this becomes worth paying for precisely because AI-generated software offers none of it.

Creative Identity becomes more important too. When anyone can generate music, the music that people pay for is the music with a human story behind it, a voice that couldn't have come from a prompt, a relationship between artist and audience that AI cannot simulate. The same logic applies to software. When anyone can generate a functional app, what people pay for is craft, trust, reliability, and the kind of deep domain knowledge that knows what the AI doesn't know it doesn't know.

Alberto Brandolini's Law (Bullshit Asymmetry Principle) applies here too: the amount of effort required to evaluate whether a piece of software is trustworthy will vastly exceed the effort required to generate it. That evaluation gap creates genuine market value for software that is demonstrably good, maintained by accountable humans, and built with enough care that you don't have to do the evaluation yourself.

Ai Economic Collapse

OpenAI’s stated mission: To “build artificial general intelligence” — AI that can replace all forms of economic cognitive labour. What does it mean for ordinary people that the most powerful AI company in the world has stated this as its goal?

The social contract that historically held was simple – governments invested in people because people generated the tax revenue that governments depended on. Businesses needed consumers with money to buy their products. Even the wealthiest elites needed a prosperous population to sustain their own wealth. For the first time in history, a small number of individuals controlling AI infrastructure may find they no longer need the majority of humanity to be healthy, educated, or economically active. Will AI break the dependency?

If AI replaces enough human labour, a significant portion of the population loses income. People without income cannot buy things. Companies selling things need customers with money. If their own AI-driven efficiency gains have eliminated the customers who would have bought their products, they have optimised themselves into a market that cannot sustain them. Henry Ford understood the opposite version of this – he paid his workers enough to buy the cars they built. The demand side of the economy matters as much as the supply side.

These same questions were raised seriously during the Industrial Revolution and again during every major wave of mechanisation since. The historical answer has always been that new kinds of work emerge to replace the old ones. But that has never before been tested against the automation of thinking itself – when the machine targets not just what people do with their hands but what they do with their minds. And in the meantime, human beings may find themselves worth more to large corporations sat staring at a screen of advertisements than outside living their lives, creating, and being productive.

This anti-human future is perhaps a political choice, not a technological inevitability.

If This Has Made You Think

What is AI actually doing to our attention, our thinking, our relationships, and our democracy? How should you use it? Where are your ethical boundaries? Are you using it to think better, or to avoid thinking? The most important act is awareness: know what you are interacting with, know what it wants from you, and make sure your thinking is still your own.

If this has sparked something, I have been working on a broader framework — 30 Laws of Artificial Intelligence — and would love to hear from anyone who wants to know more.

Some recommended follow up:

• The Social Dilemma (Netflix)
• The AI Doc (YouTube)
• The Thinking Game (Amazon)
• AI Everywhere (Amazon)
• AI Confidential with Hannah Fry (BBC)
• Naval Ravikant on Artificial Intelligence (YouTube)
• Government Regulation of AI (YouTube)
• Why AI Agents are either the best or worst thing we’ve ever built (YouTube)
• Listen to this podcast about how rogue actors might utilise Ai
• Ten Arguments For Deleting Your Social Media Accounts Right Now by Jaron Lanier