"Basically pleasant bureaucrat" vs. "Sexy murder poet"
« previous post | next post »
Travis LaCroix, Fintan Mallory, and Sasha Luccioni. "Strategic polysemy in AI discourse: A philosophical analysis of language, hype, and power." In The 2026 ACM Conference on Fairness, Accountability, and Transparency, 2026:
Abstract: This paper examines the strategic use of language in contemporary artificial intelligence (AI) discourse, focusing on the widespread adoption of metaphorical or colloquial terms like “hallucination”, “chain-of-thought”, “introspection”, “language model”, “alignment”, and “agent”. We argue that many such terms exhibit strategic polysemy: they sustain multiple interpretations simultaneously, combining narrow technical definitions with broader anthropomorphic or common-sense associations. In contemporary AI research and deployment contexts, this semantic flexibility produces significant institutional and discursive effects, shaping how AI systems are understood by researchers, policymakers, funders, and the public. To analyse this phenomenon, we introduce the concept of glosslighting: the practice of using technically redefined terms to evoke intuitive—often anthropomorphic or misleading—associations while preserving plausible deniability through restricted technical definitions. Glosslighting enables actors to benefit from the persuasive force of familiar language while maintaining the ability to retreat to narrower definitions when challenged. We argue that this practice contributes to AI hype cycles, facilitates the mobilisation of investment and institutional support, and influences public and policy perceptions of AI systems, while often deflecting epistemic and ethical scrutiny. By examining the linguistic dynamics of glosslighting and strategic polysemy, the paper highlights how language itself functions as a sociotechnical mechanism shaping the development and governance of AI.
This reminded me of a discussion from 1956, featured in a 2025 post by Cosma Shalizi, "On Feral Library Card Catalogs, or, Aware of All Internet Traditions", Three-toed Sloth 4/16/2025:
Back when all this was beginning, in the spring of 1956, Allen Newell and Herbert Simon thought that "complex information processing" was a much better name than "artificial intelligence" [9]:
The term "complex information processing" has been chosen to refer to those sorts of behaviors — learning, problem solving, and pattern recognition — which seem to be incapable of precise description in any simple terms, or perhaps, in any terms at all. [p. 1]
Even though our language must still remain vague, we can at least be a little more systematic about what constitutes a complex information process.
- A complex process consists of very large numbers of subprocesses, which are extremely diverse in their nature and operation. No one of them is central or, usually, even necessary.
- The elementary component processes need not be complex; they may be simple and easily understood. The complexity arises wholly from the pattern in which these processes operate.
- The component processes are applied in a highly conditional fashion. In fact, large numbers of the processes have the function of determining the conditions under which other processes will operate.
[p. 6]
If "complex information processing" had become the fixed and common name, rather than "artificial intelligence", there would, I think, be many fewer myths to contend with. To use a technical but vital piece of meta-theoretical jargon, the former is "basically pleasant bureaucrat", the latter is "sexy murder poet" (at least in comparison).
That 2024 skeet's thread continues with this lovely flow chart:
Cosma's post includes this ChatGPT-generated image of Feral Library Catalogs:
Anonymous Historian said,
August 12, 2026 @ 9:19 pm
Yuck. AI slop.
AntC said,
August 13, 2026 @ 5:48 am
"complex information processing"
Yeah. Famously: define "complex". Or if you prefer, define "simple". Philosophers of Science (particularly) have tried and failed and given up. "As simple as possible, but no simpler." is another good one.
Chess-playing computers were thought to exhibit 'Artificial Intelligence' in the sense that human chess-players are thought to exhibit intelligence of a very narrow, particular sort. But the artificial ones tackle the job as "very large numbers of subprocesses", each subprocess relatively simple. What enabled them to eventually beat humans was the progress of Moore's Law and massive parallelism: vast numbers of dumb subprocesses calculating faster and faster.
Isn't polysemy kinda what language does all the time? denotation, connotation. Compare: 'Black hole' — which are definitely not holes; and actually some of the brightest things in the universe. (By the narrow technical definition, it's not them that are bright, but the matter getting shredded at the Event Horizon.)
I'm increasingly suspicious that 'Dark matter' as a term is glosslighting.
I really see nothing here specific to (G)AI or LLM's. Rather it's an immature discipline trying to establish a grounded vocabulary. (Not helped by the hype and venture capital egos.)
When will all the brainboxes at Google figure out that we don't want their enshittification; nor their AI Overview; please can we go back to search as of 5 years ago.
Jerry Packard said,
August 13, 2026 @ 8:16 am
enshittification:
Though I absolutely love the word with all its morphological complexity and creativity, I’m nonetheless of the mind that what is going on is normal linguistic polysemy and metaphor, which is the be-all-end-all and raison d’etre for human natural language in the first place, IMHO. I think Cormac McCarthy got it right when he said/wrote that the birth of human thought and language happened upon the realization that ‘one thing can be another thing’ – that using a symbol or word to stand for an entirely separate object or concept is what made human thought and language possible.
Carlos said,
August 13, 2026 @ 10:27 am
@AntC, @Jerry Packard: the authors acknowledge that "innocent" polysemy is "an inevitable feature of natural language". Their argument is that AI terminology employs "strategic" polysemy of a particular kind.
Their own term, "glosslighting", suffers from the disease it denotes: they imply that glosslighting needn't be intentional ("The rhetorical effects of glosslighting—suggesting familiar meanings while retaining deniability—may arise intentionally, but they can also emerge from […] the foreseeable interaction between ambiguous terminology […], and heterogeneous audiences […], combined with institutional incentives […]"), but by evoking the term "gaslighting", they suggest deliberate, deceptive manipulation.
Olaf Zimmermann said,
August 13, 2026 @ 2:39 pm
Perhaps reading the entire paper, instead of merely scanning the over-long abstract, might have spared us the odd gratuitous, nay, snarky, comment. There's a serious point to be debated here, and and thanks to @myl for raising it (in his way ;-)
Jerry Packard said,
August 14, 2026 @ 11:53 am
@Olaf Zimmermann
Oh, I don’t know about that. I think that the abstract here did a pretty good job of reflecting the content of the paper, and provided a reasonable, stable surface from which one could easily tee up a worthy retort. In response to the abstract/article, at the very least a responder can simply either agree or disagree with the authors’ main contention that AI glosslighters intentionally use ambiguous terms that allow them to hide behind plausible deniability. That particular tack is as old as the hills, as the intentional use of plausible deniability attached to an interlocutor’s vague reference can be accomplished at just about every parry of a dialectical turn.
John Carter said,
August 15, 2026 @ 6:11 am
It is striking how “glosslighting” turns out to be not a recent distortion of the LLM era, but the foundational operating principle of the entire field. The lineage runs unbroken from the summer of 1956 straight through to today’s boardrooms and arXiv preprints.
McCarthy famously picked “Artificial Intelligence” for the Dartmouth proposal in part to distinguish the workshop from Wiener’s cybernetics and Shannon’s automata theory—and, crucially, to capture the imagination (and purse strings) of the Rockefeller Foundation. Newell and Simon’s sober operationalism was doomed from the start: you simply can’t raise billions of dollars on the back of a “basically pleasant bureaucrat.”
Twenty years after Dartmouth, Drew McDermott made this exact complaint in Artificial Intelligence Meets Natural Stupidity (1976), warning researchers that naming a Lisp subroutine UNDERSTAND or GOAL was an act of self-delusion and public misdirection.
What LaCroix, Mallory, and Luccioni articulate so well is that this semantic drift isn't an accidental misunderstanding by journalists or the public; it is an active mechanism of plausible deniability. The industry thrives in the fertile valley between the technical footnote (where hallucination is just probability distributions over next-token predictions) and the investor pitch (where hallucination implies an inner life capable of perception and error). The field has always demanded the poetry to get the funding, while keeping the bureaucracy in reserve whenever called to account.
Tom Ritchford said,
August 20, 2026 @ 6:26 am
Glosslighting! I love it.
[This got long. I want to assure you that I wrote this all with my own achy hands. No LLM had any par of this. "Untouched by robot claws" -Fritz Leiber
[I'm against LLMs for social, economic, environment, political, IP reasons, the list is very long. But know your enemy.]
Some of this is a bit like people complaining that almond milk isn't milk – some is bang on.
Take hallucination: So many BS implications. "Mistake" works very well. "Hallucination" is clearly a scam, promulgated on the public. Why they think the idea of ultrapowerful computer programs that hallucinate is attractive I don't really see, but then they really love pumping up how dangerous their work is supposed to be. Something like humble-bragging but for danger…
Alignment? What is this, a D&D campaign? Again, this implies agency, choice, even hints at free-will.
NO. Call this "Security" or "Security, privacy and safety".
—–
But introspection has been a concept in computer programming since the late 50s and Lisp, meaning a program that is able to access information about itself and its own functioning. That bird flew before the invention of the integrated circuit.
Language model seems very accurate. What's their beef?
Here you have some computer program that really does "speak" not just good English but all the common languages of the world, and apparently Pig Latin (I just tried it). It is a large model of human languages.
Programmers have been talking about agents since the 1990s, where an agent is a program delegated to do things for a person. Calling a computer program you run "an agent" does not at all imply that it has agency.
Finally, chain-of-thought I just learned about. "To construct a prompt, a user typically appends an instruction to the end of their prompt. Users commonly add an instruction to their prompt such as “describe your reasoning steps” or “explain your answer step-by-step." In essence, this prompting technique asks the LLM to not only generate a result but also detail the series of intermediate steps that led to that answer.""
"Show your reasoning" is a very common phrase used by human teachers for *human students*, I note.
—-
And here we get to the nub of the matter – the argument that LLMs do not actually reason, and thus any phrase involving reasoning is false advertising.
Certainly, LLMs simulate reasoning, claim to reason when interrogated, but people thought Clever Hans could do arithmetic.
Let's call this simulation of reasoning "L-reasoning". Let's reserve "reasoning" for human reasoning (and forget about whales and dolphins and planaria and stuff for now).
The $64 trillion dollar question is, "Is L-reasoning equivalent to reasoning?" (the process of forming conclusions, judgments, or inferences from facts or premises (The Free Dictionary))
I've spent considerable time trying to prove they were not the same.
My best try, I thought, involved coming up with all Lewis Carroll-style syllogism puzzles, except all brand-new "random" in structure, and using random words I made up that appear nowhere on the net: "Every blodgish is megacheable for a skamgongolt, that amexnaronables to monabagolbo over a penachantoneg." These problems are SAT-2. They are NP-complete. With fifty or more statements, you can't just guess through 2^50 choices.
LLMs not only solved my problems, and showed their "reasoning" but there was a bug in my program so that some of these puzzles had infinite solutions, and they told me that too.
So right now, I'm in a very weird position. Clearly reasoning and reasoning-L are not the same, but I am unable to come up with any operative way to distinguish them through their results.
And my tests were unfair. I doubt even 1% of humans could even start to attack them. On any test of "reasoning", I think ChatGPT would smoke your average guy on the street easily 90% of the time.
—-
I am a non-believer and a mechanist. And as I said, I think LLMs are a disaster for almost all humans.
But until someone shows me some reproducible difference between the results of reasoning and reasoning-L, it's hard not to conclude that they are simply equivalent processes,.
Does this mean that LLMs are conscious? God, no. I would show much more consideration for the concerns of a spider than an LLM.
But humans have created reasoning things. And it didn't happen gradually over years – there was little progress for seventy years and suddenly boom.
Jerry Packard said,
August 21, 2026 @ 9:15 am
@Tom Ritchford
Thank you. Onward and upward.