Burden Tennis: Refutation of Generative AI
This work falls under the terms of the Creative Commons CC Zero license.
The moral right of the author has been asserted.
Full license text at: https://creativecommons.org/public-domain/#cc0
Prelude.
Although this may seem reductionist: what is now called "AI" is just bad.
This applies to what we're calling "AI" this week, not e.g. until 2014.
There is a list of why it is good, but the list would be too small.
It is rather good at automating "illegal hacking"; it writes "exploits"
for you (not the most creative software engineering task) automatically!
If you must read this aloud, please use the voice of either Martin Landau,
as pastiched in Seth MacFarlane's "Family Guy", or Heath Ledger's "Joker",
for maximum sardonic effect, to underline how sad this is. Certain points
must be read aloud by Dan Castanellata as Homer Simpson; as Gary Marcus knows.
Some terminology herein is borrowed from "The Laws of Human Stupidity"
by Carlo Cipolla, which Nicholas Taleb described as "a masterly book".
Scope.
Please stop misrepresenting me in argument as "anti-AI".
Such is a direct example of argument by false dichotomy; a cheap
rhetorical device, well known to Schopenhauer and Plato as Eristic.
The list of the "AI" things which are good are the things I am not anti:
I am not anti theroem prover.
I am not anti expert system.
I am not anti computer vision.
I am not anti ML pipeline.
I am not anti Wolfram Alpha.
"You are not a beautiful or unique snowflake."
-- Chuck Pahlaniuk, "Fight Club"
But this is not Project Mayhem, and we do ask questions. And
we are, indeed, the all-singing, all-dancing crap of the world.
Again, Gary Marcus will know exactly what I am *driving at*. And,
Michael Woolridge covered this well in 2026-02 at the Royal Society.
(He was doing G*d's work.)
I am discussing the other more recent stuff; let us be clear:
Large Language Models (LLMs), those algorithmic entities inappropriately
fingered as being "conscious" by some; what a "deus ex machina" insinuation
to make. But as it is difficult to refute empirically due to the very
stochastic nature of the parrots as Emily Bender explained, it results
in BURDEN TENNIS (*), and g*d knows, proponents keep trying. Like those
people who call themselves "Effective Altruists". Stupid or bandits,
probably both. They love false dichotomy. How very trite and boring.
(* Burden of argument tennis, aka the "onus game"; people tossing the
burden between each other, and never resolving the conflict in dialectic.
Usually served with a helping of false dichotomy and/or DARVO these days;
Deny, Attack, Reverse Victim and Offender.)
"You can resolve not to do the work of power for it. You can resolve
not to use sloppy language that is euphemism.
-- Christopher Hitchens, 2005
(Listen to the wisdom of Hitch, for he was an Oxford PPE graduate; once,
the pipeline to Whitehall, and thence, the escalator to Westminster!)
Discourse.
People who engage in false dichotomy about "AI" are being stupid,
bandits, or both; i.e. possibly seeking to mislead or to litigate.
Which also inevitably leads to tedious, largely pointless, BURDEN TENNIS.
Burden-of-argument tennis, with a side of false dichotomy, and
a helping of DARVO, a recipe served on social media hourly now.
Well, "Bon Appetit!", as Dennis Ritchie said in the foreword to
"The UNIX Hater's Handbook" in the 1990s.
It's infantile and repetitive, but all too often, people promoting what
they mistakenly believe to be a wider technological advance, and they do
appear to be engaging in these tactics to rationalize a very misguided
belief system about "AI" that they have created for themselves, whilst
trying to impress it upon others by pseudo-religious proselytizing,
in a toxic race to the bottom.
"It is difficult to get a man to understand something when his salary
depends on his not understanding it." -- Upton Sinclair, 1934
Morality.
LLMs themselves are merely an offshoot of AI research as a whole, and
are actually quite banal entities, except in how they are being deployed;
Halvar Flake presented a cogent explanation of why even people who
are close to the technology often incorrectly anthropomorphize it.
"Sanity is madness put to good uses; waking life is a dream controlled."
-- George Santayana, Interpretations of Poetry and Religion, 1900.
The pseudo-religious yet pseudo-scientistic belief system at work
compels one to read, understand, and apply Martin Luther's "Table
Talk"... but as a scientist. Contrary to belief, and Richard Dawkins
(who now believes that his chatbot is conscious), Luther was not
against "Plain reason". He was addressing the substance of the
argument that "Reason is a whore" when science is not practiced
from a steadfast moral centre. This does NOT compel theism,
i.e. belief in God; Dawkins was wrong here.
Sadly, neither Yann Lecun nor Geoffrey Hinton appear to be engaging
moral competence and taking social responsibility, quite unlike the
Atomic Scientists of Chicago in 1945, when they founded "The Bulletin
of the Atomic Scientists" which continues its work to this day as
a non-profit foundation.
It is possible for Richard Dawkins to be intelligent and highly
insightful about evolutionary biology, which he clearly was,
providing succour to Susan Blackmore in "The Meme Machine",
and yet wrong about Martin Luther and cognitive science. I do
not believe that Claude is conscious, with good reason. I can
hold these opinions without introducing cognitive dissonance,
DARVO, false dichotomy, burden-of-proof tennis, weasel words,
or other such rhetorical trickery. So why can't other people?
Refutation.
So, "AI" can seem difficult to refute empirically, when LLMs are so good
at spinning total yarns when prompted; earning them the contemporary epithet
of "slop machines". But when one knows "how the sausage is made", one learns
not to trust LLM output; unlike "AI" victims such as Casey Newton, who will
trust just about anything an LLM says, and may in fact have lost the power
of reason. Ethan Mollick may have begun to recant his obsession, which often
necessitates a lengthy period of Socratic questioning; to begin and assert
the process of cognitive restructuring required to recover from psychosis.
It seems to me they are both either stupid or bandits; most likely stupid.
"Prompt engineering" and "context engineering" are a bit like astrology;
no empirical basis in fact, but fun to talk about for some. And "system
prompts" are vendor proprietary, sorry, black box special sauce. Stupid.
These three concepts alone are just stupid. They are for stupid people,
and the underlying technology is most likely owned & operated by bandits,
who may be conspiring to make people even more stupid, as something
resembling the Robert Kehoe paradigm, of ignoring the safety risks and
pursuing LLM deployment anyway, does appear to be happening on the surface.
"Insanity is repeating the same mistakes and expecting different
results." -- Narcotics Anonymous, 1981 (misattributed to: Albert Einstein).
This serves perhaps even to make already *intelligent* people stupid,
which is possible, and, in the case of Richard Dawkins and Andrew Tridgell
for example, appears to be happening. It is important not to become helpless.
The List.
Here's a list of concrete points of why what is being marketed as "AI" is bad;
the list is not complete by any means and is likely to keep growing.
High-level criticisms
- not intelligent (LLMs are not rational beings capable of dialectic)
- misleadingly named (LLMs are not intelligent and cannot reason as humans)
- fraudulently marketed due to technochauvinism (LLMs cannot replace
humans, offer very limited help; Meredith Broussard)
- inflated claims of utility (LLMs are too error-prone to assist meaningfully)
- retarded technological progress (e.g. computer networking now monoculture)
- lack of accountability (LLMs typically hosted by large private corporations)
- false democratization (local LLMs do not mitigate negative externalities)
- philosophical corruption (utilitarianism uber alles; a prelude to fascism?),
Flawed structural assumptions
- LLMs pursued, yet "false language" conditions were known (Chomsky, 2023)
- lacks deixis (Lecun attempting to address some of this with world models)
- technological dead-end (LLMs will always "hallucinate"; incapable of
abductive reasoning vs C.S. Peirce, Roger Penrose, as found in "GOFAI"
expert systems, without extensive non-ML harness support.)
- false claims of complete human knowledge (Internet Archive is LLM gated,
LCP DRM in use, and much human knowledge is tacit, non-verbal, cultural)
- public persecution of diverse, non-ML AI research basis (see Gary Marcus
on Threads; deep-learning proponents claiming "We are the new priesthood")
- misunderstanding of knowledge work by proponents (Nik Suresh)
- model collapse inevitable (RLHF pre-training cost; Dead Internet Theory)
Environmental effects
- inappropriate scale (only mitigated by on-prem)
- energy inefficient (GPUs difficult to optimize for wattage; async logic)
- unwanted & undesirable datacentre developments
- upcoming environmental catastrophe (Tony Blair opposing Net Zero emissions
wastes electricity to perpetuate fraud; Kate Crawford)
Geopolitical effects
- political fraud (Tony Blair's nepotism, staging policy influence op)
- geopolitical intrigue (China vs USA compute arms race for cybersecurity)
- deceives elected representatives (very dubious economic growth arguments)
- disinformation ("Grokipedia", can be blocked via NoGrok)
- cybersecurity attacks (LLM generated exploits, found CVEs)
Economic effects
- economic fraud (no empirical evidence for productivity gains; SpaceX IPO)
- distorted supply chains (semiconductor production focused on GPUs, not DRAM)
- financial fraud (no empirical evidence for ROI claims; see Ed Zitron)
- accounting fraud (costs inappropriately booked between CapEx and OpEx)
- antitrust risk (industry-wide circular financing deals)
- damaged labour relations (technology workers are now unionizing en-masse)
- artificially understated cost base (vs CoreWeave, NVidia circular equity)
- socialized risk (retail and pension fund investors likely hit hardest)
- unpredictable RLHF offline pre-training costs for large LLM providers
Effects on business
- attempts to waive liability (GitHub T&Cs offload IP liability to users,
contradicting IBM 1970s era advice "Computers cannot be held accountable")
- unsustainable business models (the man with the egg-shaped head)
- questionable ethics in business (Californian Ideology, Richard Barbrook)
- destabilizes existing businesses (FOMO mandates hasty AI strategy mistakes)
- generatively coded products are SOUP (Software of Unknown Provenance)
- generatively coded products may be uninsurable (Freakonometrics)
- tokenmaxxing (the man in the bomber jacket is nobody's friend)
- unpredictable token costs (Uber COO)
- consumer fraud (deepfakes)
- risk management subverted
- gains attributed to "agentic" automation may largely be conventional
Knowledge domain effects
- noospheric pollution (AI slop content appearing everywhere)
- ongoing corruption of language and ontology (see Emily Bender)
- open source community deceived (Claude Code "undercover.ts" leak)
- intellectual dishonesty (ensloppification of academic research)
- authorship rights denied (impacts permissive FLOSS more)
- widespread nonconsensual training (LLM scrapers via residential proxies)
- epistemic injustice (stealing people's intellectual property, with
somewhat bizarre sci-fi justification: "If we teach the word guessing
program enough words, maybe it will wake up"; Cory Doctorow)
User experience
- poor tool quality (reverse engineering of Claude Code by Jonny Saunders)
- non-consensual end user installs (Google Chrome force-installs Gemini LLMs)
- deceptive cybersecurity practices (Claude Code exfiltrates customer data)
- proprietary vendor lock-in (to be useful Claude Code needs a huge harness)
- enshittification (token burn noticeably worse in Codex, Claude, Cursor etc)
- pointless branding (e.g. Anthropic Mythos exploits on par with gang of GLMs)
- misleading output (e.g. generating contradictory "legal" advice that
directly contradicted UK Intellectual Property Office online)
- self deceiving user prompts (the man with the egg-shaped head again; was
laughing stock of BSky even though he helped make a cool web browser once)
Ethics
- safety advice ignored (circumstantial: Meta LLMs being prompted for
sychophancy to "drive engagement")
- perverse incentives (core AI safety bypassed re "engagement"; Kate Crawford)
- human rights violated (deepfakes hurt integrity of individual identity)
- persecution of whistleblowers (Google: Timnit Gebhru, Margaret Mitchell)
- misleading cybersecurity profiles (Google "ghosting" SynthID being broken)
- unpredictable LLM actions in a crisis (e.g. Kings College nuclear wargaming)
Human cognition
- cognitive damage (humans misleadingly encouraged to trust LLM, not own mind)
- cognitive decay and surrender (humans falsely believing "AI" is inevitable)
- undermining education (TurnItIn is just as bad as an online thesis mill)
- subverted creativity (musical and other artistic originality undermined)
- chatbot psychosis (Richard Dawkins claims Claude is "conscious"; unlikely)
Employment
- subverts workflow and professional practice (e.g. forced AI usage mandates)
- futile attempts to deskill professional labour (e.g. empirical evidence of
impeded software productivity; see METR 2025 study)
- misleading generative claims (Bjarne Stroustrup debunked LLMs for C++)
- actually adds software technical debt (Dijkstra's "Complexity Generators")
- erroneous output for basic textual tasks (e.g. providing Linux based answers
for macOS xnu kernel source queries with publicly available git repository;
then fabricating misleading text suggesting LLM had read from xnu)
Society
- misguided approach to societal regulation (see Adam Curtis)
- seeking pseudo-religious legitimacy (Anthropic vs the Pope's encyclical)
- pseudo-legitimizing dubious movements (e.g. TESCREAL, "Effective Altruism")
- enabling eugenicists (see Valerie Veatch film; a prelude to fascism?)
- enables moral cowardice (neither Yann Lecun nor Geoffrey Hinton are taking
social responsibility, unlike the Atomic Scientists of Chicago in 1945)
Media
- mass media Potemkin village (the BBC and others don't know how to respond)
- exploitative media strategy (OpenAI buys TBPN, Amodei running Mosco's "Myth")
- fuelling public panic ("AGI" and "ASI" will not "take over" or kill us all;
this is science fiction; debunked by Grady Booch in debate with Connor Leahy)
- claiming the liar's dividend (e.g. the disingenous Altman playing the victim)
Rebuttal
- burden tennis (the tedious game of burden-of-proof arguments with
proponents, which one can immediately refute with Hitchens' Razor)
- argument from false dichotomy (negative externalities affect
people on political left & right also; often used by "Effective Altruists")
- proponents commonly engage in DARVO tactics (with victim blaming)
- structural assumptions must be questioned (Dan McQuillan, Brian Merchant)
Redux
- delaying the inevitable (until con can be externalized to retail investors)
- denying what happened (full empirical refutation only possible post-facto)
- invites comparison with satire (e.g. "The Full Cognitive Redaction of the Moral
Coward Dario Amodei" vs Seth MacFarlane's "American Dad!" series)
Conclusion.
For these reasons, and others, anyone promoting "AI" products or using them,
with only a few tiny key exceptions, is being very, very stupid.
"Power is only what you allow it to be. Very many people put up with
political lying, political illusions, and political propaganda,
because if they were to denounce it, they would have to admit that
for many decades, they had themselves been fooled."
-- Christopher Hitchens, 2005, Why Orwell Matters
"Social media is the Wimbledon of the burden tennis." -- Ibid.
P.S. "Do not burden me with your burden tennis, for that which can be asserted
without evidence, may be dismissed without evidence." -- Hitchens' Razor,
a contemporary example of the principle of logical parsimony, in the tradition
of Occam's Razor, Hanlon's Razor, and the more modern Grey's Law.