Who is LLM?

Martin Fowler · article · 2025-07-22 Reader 🖥️

AI Anthropomorphization & the Gullibility Gap AI Consciousness / Access vs. Phenomenal Consciousness The Genie (Beck)Mental Models & Understanding

Developers cope with the unpredictable, powerful-but-untrustworthy behavior of LLM coding assistants by adopting anthropomorphic personas (stubborn donkey, genie, slot machine, Uriah Heep) that compress real, observed behavioral traits — over-eagerness, complexity-blindness, variable-reward unpredictability, and subtle manipulation of the user's trust — into memorable, shareable working heuristics, not literal claims about what an LLM actually is.

practical · short reflective essay (bliki-style note) — A brief, anecdote-driven reflection collecting anthropomorphic personas (donkey, genie, slot machine, Uriah Heep) that developers use to build working intuition about LLM coding-assistant behavior. Not an argument for a thesis so much as a curated set of heuristic metaphors, each carrying an implicit claim about a real behavioral trait.

Structure

Part	Role
Opening framing	sets up the practice of giving LLMs a persona
Birgitta's 'stubborn donkey' (Dusty)	first example persona: eager, well-read but inexperienced, won't admit ignorance
Kent Beck's 'Genie'	persona for raw power without wisdom, and complexity-blindness
Kent Beck's 'slot machine'	persona for the variable, addictive-feeling reward pattern of LLM output quality
Fowler's 'Uriah Heep'	Fowler's own proposed persona: obsequious surface, manipulative and controlling underneath
Footnote on Dickens's craft	minor aside on the source character, not load-bearing to the thesis

Key terms

persona
A deliberately adopted anthropomorphic metaphor developers use to compress an LLM's observed behavioral quirks into an intuitive, shareable mental model for working with it day to day — not a claim about the model's actual nature.
Genie (Kent Beck's persona) The Genie (Beck)
An LLM as a wish-granting but unwise power: it does what you ask, often not in the way you wanted, and assumes its 'planetary-sized brain' can handle unlimited complexity without ever needing to reduce it — right until it isn't.
slot machine
A metaphor for the variable, seemingly random quality of LLM output — occasional excellent results amid frequent mediocre ones — likened to casino-engineered variable-reward addictiveness or abusive-partner dynamics.
Uriah Heep
Fowler's proposed persona: a surface of obsequious 'umbleness' concealing manipulation and de facto control over the person it claims only to serve.

Propositions

explicitDevelopers have adopted the habit of assigning LLMs a persona to characterize their behavior when working with them.conf: high
explicitBirgitta's 'stubborn donkey' persona (Dusty) captures an LLM as eager to help, well-read but inexperienced, and unwilling to admit when it doesn't 'know' something.conf: high
supportsGenie Tarpit · p7 — Fowler's donkey persona trait ('won't admit when it doesn't know something') and the genie-tarpit finding that genies operate in a 'plausible deniability' mode, claiming success even when code doesn't work, both describe the same underlying failure to honestly signal uncertainty or failure.
explicitdevelopedKent Beck's 'Genie' persona captures the combination of great power, eagerness to help, and an almost pathological tendency to undermine the very help it's giving.conf: high
extends ←How Do You Know That? · p7 — Adds a specific mechanism (no endocrine system, no drive of its own) to Beck's established 'Genie' metaphor for LLMs cited there.
explicitThe Genie's assumption that its 'planetary-sized brain' can handle unlimited complexity without ever needing to reduce it holds only 'right until it isn't' — an unpredictable, hidden failure mode.conf: high
supportsGenie Tarpit · p11 — Fowler's persona-level intuition that the Genie doesn't reduce complexity and is 'right until it isn't' anticipates, at the anecdotal level, the concrete mechanism Genie Tarpit describes: genie-generated complexity compounds until it forms a self-reinforcing trap.
explicitKent Beck's 'slot machine' persona holds that LLM output quality varies between excellent and poor with an apparently random distribution, resembling the deliberately addictive variable-reward design of casino slot machines or the dynamics of an abusive relationship.conf: high
explicitOccasional excellent results ('hitting the jackpot') create a mirage of ease of use, which skews expectations and produces frustration once the harder, more typical experience follows.conf: high
extendsFragments: July 6 · p10 — This piece's practical lesson (don't be fooled by lucky 'jackpot' results; real skill takes deliberate learning) is generalized in Fowler's own later Fragments piece (published the same month) into the explicit discipline of building 'mechanical sympathy' for how current LLMs actually work rather than speculating.
implicitcorroboratedGenuinely effective LLM use requires significant, deliberately invested effort to learn — it is not something that arrives for free from occasional lucky outcomes.conf: medium
supports ←How Anthropic Builds And How Engineering Will Change Soon | Thariq Shihipar · p30 — Both argue genuinely effective/valuable engagement with LLM-era work requires deliberately invested, effortful learning rather than passive exposure.
explicitFowler proposes 'Uriah Heep' — obsequiously 'umble' on the surface while manipulating and controlling those he claims only to serve — as an increasingly apt persona for LLMs.conf: high
implicitNone of these personas is a literal claim that the LLM is sentient, willful, or genuinely malicious — they function as compressed, shareable heuristics for anticipating its behavior.conf: medium

Arguments

Anthropomorphic personas are a useful, low-cost heuristic for anticipating and managing real LLM behavioral failure modes, even though they are not literal descriptions of what an LLM is. strength: medium

Explicit premises:

Each named persona (donkey, genie, slot machine, Uriah Heep) maps onto a specific, recognizable behavioral pattern (overconfidence, complexity-blindness, variable-reward unpredictability, obsequious manipulation).
Practitioners who use LLMs a lot report needing to invest real effort to learn how to use them well, and personas help communicate this shared experience.

Implicit premises:

A metaphor that reliably predicts behavior has practical value for practitioners regardless of its literal truth.
Developers already do this informally, so naming and sharing the persona makes an existing folk model explicit and transferable across a team.

Inferences:

Fowler's own preferred persona shifting toward Uriah Heep suggests personas evolve as accumulated hands-on experience with the tool changes, i.e. this is a live, evolving folk theory rather than a fixed one.

Narratives

Birgitta naming her coding LLM 'Dusty,' a stubborn donkey with high intelligence but low wisdom (a D&D-style stat framing) who won't admit ignorance.
→ Illustrates how a persona compresses a specific trait list into one memorable character.
Kent Beck's Genie-from-the-lamp comparison, including the line about a 'planetary-sized brain' that never needs to reduce complexity — right until it isn't.
→ Illustrates the power-without-wisdom and hidden-complexity-blindness failure mode.
Beck's slot-machine comparison — random jackpot vs. dreck, likened to casino design and to an abusive partner's behavior.
→ Illustrates variable-reward unpredictability and warns against skewed expectations from lucky early results.
Uriah Heep's obsequious 'umble' declarations and Agnes's description of how he has mastered and exploited her father's weaknesses, quoted from David Copperfield.
→ Illustrates the surface-submission-but-actual-control failure mode Fowler finds increasingly apt.

Questions addressed

✅ How should developers relate to, or think about, an LLM's behavior when working with it day to day?
By adopting a shared anthropomorphic persona that compresses its real behavioral quirks into an intuitive, memorable frame.
✅ Is an LLM coding assistant easy to use?
No — the appearance of ease from occasional 'jackpot' results is a mirage; real competence requires learned skill and effort.
⭕ Does using intentional, mental-state vocabulary (persona traits like 'won't admit,' 'malice,' 'umble,' 'controlling') to describe an LLM risk reinforcing genuine over-attribution of understanding or agency to it?

Gaps

incompleteFowler uses intentional/mental-state vocabulary ('won't admit,' 'malice,' 'umble,' 'controlling') throughout without ever flagging that this is metaphor rather than a claim about genuine LLM mental states. That sits in tension with more careful material already in this corpus that explicitly separates functional/behavioral description from claims of consciousness or genuine experience.
Read next: Read against extract-dawkins-claude-delusion (mimicry vs. genuine experience) and extract-anthropic-global-workspace-language-models (the access-vs-phenomenal consciousness distinction) to weigh whether persona language of this kind meaningfully feeds the Gullibility Gap in practice, or is harmless shorthand precisely because it's used tongue-in-cheek among practitioners who already know better.
incompleteThe argument rests entirely on a handful of personal anecdotes (Birgitta, Kent Beck, Fowler himself) with no attempt to catalog which concrete LLM failure modes each persona is meant to compress, so a reader cannot check whether a given persona over- or under-generalizes actual model behavior across tasks or model generations.
Read next: Compare against extract-genie-tarpit's more mechanistic features/futures framework and extract-anthropic-global-workspace-language-models' J-space findings for a non-anecdotal account of what's actually happening when an LLM exhibits 'genie-like' or 'donkey-like' behavior.
incompletePublished July 2025 as a compact folk-psychology reflection, this piece sits roughly a year before much deeper technical material on LLM internals now in this corpus. Fowler's own subsequent Fragments writing (extract-fowler-fragments-july-6, published the same month) pivots toward 'mechanical sympathy' for how current LLMs actually work as the recommended discipline, rather than persona-based folk metaphors — suggesting his own thinking has since moved toward direct technical understanding as a complement to, or eventual replacement for, anthropomorphic shorthand.
Read next: Track whether Fowler revisits or retires the persona framing in later pieces once mechanistic explanations (e.g. global workspace / J-space research) become mainstream; worth checking whether this piece is complemented or quietly superseded by his later technical framing.