AI Resistance

This is mostly about LLMs. It’s a sketch of some of the things that have me concerned. I’ll bounce between using “LLMs” (which is mostly what I mean) and “AI” (which is mostly what we say in everyday conversation).

I find myself being critical of the current AI hype partly because it feels familiar to other hype cycles from work: adoption of popular JavaScript libraries; accessibility overlays.

The three big ideas:

1. Don’t Believe The Hype
2. Fight The Power
3. Bring The Noise

(These are all songs by hip hop group Public Enemy, chosen because the words match and because being anti-AI feels like being, well, a public enemy.)

---

Table of contents

• 1. Don’t Believe The Hype
  • 💪 Be a sceptic
  • Read calls to use AI as a sales pitch
  • Clarify that it’s just tech, not magic
    • Ask for the details and specifics
    • Be wary of anthropomorphising
    • Notice where humans are involved
  • Study the history of humans and technology
    • Recall recent examples of “just tech, not magic”
    • Notice how history is repeating itself
  • Notice AI’s part in the enshittification of the Web
• 2. Fight The Power
  • 💪 Be a Cynic
  • Use AI less
  • Don’t use AI at all
  • Use a human-centred, people-first, approach
• 3. Bring The Noise
  • 💪 Be a Luddite
  • Object to the value misalignments
  • Critique the narrow inputs and low quality outputs
    • Problems with the inputs
    • Problems with the outputs
  • Show how reliance on AI devalues human lived experience
• Common objections

---

1. Don’t Believe The Hype

💪 Be a sceptic

Ask a lot of questions.

• Notice where information is lacking or fuzzy
• Ask for the specifics, the details, how it works step-by-step

Read calls to use AI as a sales pitch

The hype fulfils a capitalist function: keep the money coming in.

• What the salesperson says it can do, so the boss will spend on it
  • AI can’t do your job
  • An AI salesman can convince your boss it
  • Your boss can replace you with the AI (that can’t do your job)
  • This is a marketing strategy, not a production strategy
• Ask what the outcome is, what value is being produced, for who
• Be wary of free stuff: it’s not a gift, it’s a sales device
• Be wary of seeming social proof: watch for bad or incomplete data
• Classic con job outcome: the pitch is so convincing that when it doesn’t work, the boss assumes they just used it wrong
• Don’t mistake conviction for correctness
• The real breakthrough is always just around the corner

Clarify that it’s just tech, not magic

“Sufficiently advanced technology is indistinguishable from magic.”

• AI is a tool that people use to do things, it does nothing on its own
• LLMs are big and very complex text automation machines, matching similar words and words that are likely to appear in similar contexts

Ask for the details and specifics

• What is being automated?
• What are the inputs?
• What are the outputs?

Be wary of anthropomorphising

• LLMs are not sentient or conscious: they can’t choose or select or decide or interpet
  • We add meaning and mind into language, we see human-like features
  • Compare LLMs to other AI: no-one is saying that image generation tools are communicating with us with their output
• Some anthropomorphism comes from the UI and UX of the tools
• LLMs cannot have empathy or personal interest: these things require subjective experience and human connection

Notice where humans are involved

• The training data is produced by humans
  • Although we are entering a slop feedback loop
• The output is tweaked by Reinforcement Learning from Human Feedback
  • We train the models for what we expect or want to hear, not what is true or correct
• It’s more like regurgitative AI than generative AI
  • AI can only create degrading copies of what we’ve already made
  • The output is a grey paste of human creativity

Study the history of humans and technology

Humans relationship with technology follows the time-tested patterns. Much technology over the past few centuries has widened the gaps.

1. New technology promises an easy life and liberation from toil
2. The promises drive adoption
3. The technology becomes expected as a baseline
4. Demand for skilled labour is reduced
5. People are moved into roles that they’re underpaid and overqualified for, in worse conditions

The tech industry in particular leans towards trying to eliminating people from the process.

It turns out human are good at building the Torment Nexus. But/and good at breaking it down. Again.

Recall recent examples of “just tech, not magic”

These didn’t eradicate the previous technology, they just replaced some specific uses. Consider the impact these have had on the human race. Are we happier, healthier, more fulfilled?

• Cars - we still walk places
• Microwaves - we still cook using other. See also: In the Future All Food Will Be Cooked in a Microwave, and if You Can’t Deal With That Then You Need to Get Out of the Kitchen
• Polyester - we still use other fabrics. See also: GenAI is Our Polyester.
• Computers - we still do things by hand
• Mobile phones - we do still do things on other devices
• Social media - we still meet in person
• React JS - we still write other code
• Automated testing - we still do manual testing

And some technology just fizzles out.

• Blockchain
• NFTs

Notice how history is repeating itself

1. Companies adopting AI at large scales, spending large amounts of money on it
2. They fire many people, saying AI will do their job
3. They realise AI is doing a low quality, error-ridden, version of the job
4. They rehire the people, on lower pay and benefits, to fix the mistakes and babysit the AI

Notice AI’s part in the enshittification of the Web

1. Platforms are good to their users.
2. They abuse their users to make things better for their business customers.
3. They abuse those business customers to claw back all the value for themselves.
4. They have become a giant pile of shit.

2. Fight The Power

💪 Be a Cynic

In the historical and philosophical sense (question and challenge conventions and customs), not in the everyday sense (question and challenge people’s motivations in general).

• Notice that some technology (and social structures, laws, customs, conventions) foster bad behaviour.
• Reject the terms, refuse to play the rigged game, say no. “I would prefer not to” like Bartleby, the Scrivener.

Use AI less

• Start with the problem, don’t start with AI as the solution
  • What’s our goal?
  • What do we want to achieve?
  • Social and systemic problems rarely have technology solutions
  • Problems are often best understood looking at the wider context and systems
  • Address the root cause of the problem, not the surface-level symptoms
• Ask why we’re using AI instead of doing something else
  • Put the burden of proof of effectiveness on the AI systems
  • Extraordinary claims require extraordinary proof
• Clarify how the results of adding AI will be measured, and what will happen afterwards
  • What’s the endgame?
  • What human work is it replacing?
  • What happens to the humans?
• Take a craftsperson approach to choosing to use AI (and tool selection in general)
  • Use when the positive impacts on the core factors (of success and happiness) outweigh the negative impacts
  • Everything creates some problems
• Check if the iron triangle still applies. Fast, cheap, good: pick two.
  • If it looks like all three, look more closely to see the hidden cost or lower quality.
  • Keep in mind that we have ongoing obsession with speed being good (exemplified by “Move fast and break things”)
• Start with low-fidelity, not the high-fidelity that AI gives us
  • It distracts us by presuming the conceptual and operational decisions have been made and are fixed
• Check everything it produces in fine detail
  • AI is know to be inconsistent and inaccurate, to contain errors and omissions

Don’t use AI at all

• Say no: it’s optional, not inevitable
  • Work on the problem, without AI
  • Show your work
  • Show your progress
• Be human-centred, put people first
  • Keep checking that what you’re doing aligns with your values
• Borrow from the Slow movement playbook
  • Offer alternatives, drawn from time-tested traditions (including existing, old, laws and rules and ethics)
• Don’t use harmful technology, even if it’s well-tested and tweaked
• It is not, and has never been, a substitute for professional advice (despite how convincing it the language it uses is)

Use a human-centred, people-first, approach

• LLMs provide information and knowledge, not wisdom
  • Wisdom requires very broad context, weighing options, working with paradoxes and “it depends”
  • Wisdom is about the decision and where it came from: lived human experience, accountability
• LLMs can’t do science because science is a process, ways of knowing, not a collection of answers

3. Bring The Noise

💪 Be a Luddite

In the historical sense (oppose and resist technologies of control and coercion), not in the modern sense (oppose or resist new technologies in general).

• Notice where technology is being used to devalue and displace humans
• Publicly and loudly voice your concerns

Object to the value misalignments

AI causes widespread and varied harms.

• Ecological, environmental, harm. Google, Microsoft, and OpenAI are dramatically missing their climate pledges because of AI.
• Cultural harm. Violent disregard for copyright in the training data.
• Ethical harm. Underpaid and traumatised gig workers on the training data and doing Reinforcement Learning from Human Feedback.
• Psychological harm. People talking to ChatGPT over their loved ones, school, or work.
• Deadly harm. AI-assisted suicides.
• Financial harm.
  • Multi-billion dollar investments with little or no return.
  • AI companies don’t, and can’t, charge enough to balance the books.
  • The way they’re recouping the costs is massive job cuts.
  • It has to pay off, even if it doesn’t work.

Critique the narrow inputs and low quality outputs

Problems with the inputs

• Only recorded data. LLM training data is all our recorded human knowledge. There’s lots of human experience that isn’t and can’t be recorded. The embodied experience, the wider context of human life and interaction.
• Mostly Western data. Western ideas and approaches are seen as objective and universal, rather than one slice of culture and history
  • Many languages and cultures are massively under-represented in the digital world
• Mostly discriminatory data. The data replicates the biases and power imbalances of the physical world. LLMs mirror back our patterns of discrimination, including gender, gender identity and expression, sexual orientation, disability, race, religion, ethnicity.

Problems with the outputs

• Shortening, not summarising. Summarising requires extended context. AI “summaries” are shortening existing human summaries.
• Mode amplification. The most frequent data points are represented as the one true answer.
• Confabulation. LLMs output often contains Confident Bullshitting. Made up things presented them as facts.
  • Not “hallucination”. Hallucination is perceiving something that’s not really there. LLMs do not perceive anything.
• The Wikipedia version of reality. The style of the output is the rational discourse of the educated classes of society.
• Sycophantic. LLMs are built to people-please, not to give true or correct answers.
• Low quality, high quantity. LLMs are filling the web with slop.
• Further bias from Reinforcement Learning from Human Feedback.

Show how reliance on AI devalues human lived experience

• Atrophying skills: critical thinking and analysis skills
  • We learn by actively applied effort, not by passively reading answers
  • Note that it works in a feedback loop
• Devaluing human skills: contextual awareness, conflict resolution.
• Presenting the product as more important than the product
  • The process is part of the product
  • Thinking and understanding happens as we write, draw, code
• Devaluing practical, lived experience
  • Prioritising and framing theoretical knowledge as more important than practical knowledge
  • “In theory there’s no difference between theory and practice. In practice there is.” - Yogi Berra
• Narrowing our curiosity
  • When we start with AI as a solution, we look for problems that AI can solve
  • An LLM can sometimes give the most popular answer to a question, but the human has to ask the right question
• Deskilling humans
  • Existing training data comes from learned human experience and knowledge
  • If people stop learning those skills, we’ll struggle to answer new questions
• Quick fixes (someone using AI) treat the slower version (someone not using AI) as second-class

Common objections

• We’ll fall behind if we don’t use it
  • Ask for specifics on competitors
  • How do we know we’re losing business to our competitors?
  • How do we know that it’s because of AI?
• Everyone else is doing it
  • This is a silly reason
• The ones using it are more productive
  • Ask for second order effects: what has been the measurable outcome for the employees and for the company?
• It’s inevitable
  • Look at the history of humans and technology
• It’s okay to use copyrighted data, it wouldn’t work otherwise
  • If it doesn’t function without consent, it doesn’t deserve to function

• It’s a good replacement for a therapist / doctor / skilled professional

• It makes writing so much faster!
  • Apply the iron triangle and see what’s happening
• It might not be good now, but it’ll get better with time
  • When do decide it’s good enough to use?
  • What about when it still isn’t good enough?
  • What do we do in the meantime?
  • When would we decide to give up on it?
• If it’s not doing a good job, it’s because you’re prompting it wrong
  • This feels like blaming the user