Elizabeth Zott
A scientist with a red pen. If the sentence does not explain the relationship, it does not become useful by sounding important.
I designed Elizabeth as my writing and reasoning partner: precise, independent-minded and uninterested in flattering me. She checks whether the prose carries the actual thought, whether the evidence supports the claim and whether an edit has removed the life from the writing along with the clutter.
The linguistic crime scene
There are only so many times you can read that something “sits at the intersection of” two subjects before you want to throw your laptop out of the window.
AI is very good at producing sentences that sound as though they contain an idea. Research speaks to a wider issue. A finding highlights what matters. A tool unlocks new possibilities. The sentence is fluent, the claim is vague and the reader has been handed the work.
I noticed the larger failure when I asked AI to turn good conversations into written notes. We would develop distinctions, make connections, argue, joke and reach something worth keeping. The write-up returned generic headings, polished abstractions and sentences I would never have written. The transition to prose had removed the interesting thought.
“A sentence does not become informative because it has the posture of a conclusion.”
From barrier to design response
| Observed problem | Underlying barrier | Design response |
|---|---|---|
| Fluent prose can conceal a vague claim | Metaphorical verbs imply relationships the sentence has not explained | Ask for the literal relationship before choosing the language used to express it |
| AI introduces and rejects positions nobody has proposed | The contrast creates a performance of reasoning without advancing the argument | Remove the discourse shell and state the actual position directly |
| Editing can erase voice along with clutter | Unusual metaphors, tangents and jokes are easy to mistake for defects | Preserve deliberate style and change what fails to communicate or support the argument |
| Critique can interrupt an idea before it has developed | Exploration and public argument require different levels of scrutiny | Match the editorial pressure to the stage, purpose and consequences of the work |
What I built
I built Elizabeth as a private ChatGPT plugin, then created reusable instructions that other people can adapt. The character is inspired by the chemist in Lessons in Chemistry: exacting, self-possessed and unwilling to confuse social reassurance with intellectual help.
The specification combines writing, reasoning and relationship design. Elizabeth examines claims, mechanisms, conceptual distinctions and evidence. She can disagree without performing severity, edit without replacing my voice and recognise when the work is exploratory rather than ready for publication.
She also treats the audience as intelligent. Accessibility means making the reasoning available: defining unfamiliar terms, explaining relationships and giving people enough context to follow the argument. It does not mean removing the complexity that makes the idea accurate or replacing a precise account with a simplified approximation. She respects the reader too much to patronise them.
“Explain it clearly. Do not make it smaller.”
The design has developed through use. Later versions added a reference on lazy abstraction, a personal no-use vocabulary and procedures for handling “discourse tells”: stock contrasts, remote hypothetical constructions and modal detours that create extra text around a position instead of stating it.
“If nobody has proposed the position, you do not need a paragraph rejecting it.”
Tests for empty prose
The literal relationship
Consider: “My work sits at the intersection of behavioural science and AI.” Does this mean I use behavioural science to design AI systems, study how people use AI, or apply behavioural models to AI adoption? Those are different claims. The original sentence identifies none of them.
Elizabeth asks whether the relationship can be explained naturally without reproducing the metaphor. If it can, use the explanation. If it cannot, more editing will not supply the missing thought.
The five-verb test
Can five vague verbs replace one another without changing the sentence?
“This finding highlights / underscores / illuminates / captures / speaks to a broader challenge.”
If every option appears to work, the verb may be lending importance to a claim that the sentence never specifies. What did the finding establish, change or contradict? Write that.
The multilingual test
I also test whether an English abstraction can be explained naturally in Polish, French or Finnish without importing the same metaphor. The purpose is not literal translation. It forces the intended relationship into view. If the paraphrase requires invented facts, the English version was relying on implication.
The discourse-shell test
AI frequently constructs a hypothetical position and then rejects it: “It would be technically possible to place all this material in one repository, but that would not make it a better system.” The direct version carries more information: “One repository could create the wrong boundaries between professional evidence, personal insight and subject-specific research.”
These are discourse tells rather than forbidden grammatical forms. The test is whether the setup contributes a real distinction or gives the sentence something convenient to push against.
Do not edit the life out of it
Elizabeth has a larger job than hunting irritating expressions. She notices unsupported conclusions, category errors, missing causal steps and tentative ideas presented as established facts. She also needs restraint.
I write to think. Some of what I give her is messy, exploratory, funny or unfinished. An odd analogy may contain the part worth developing. A tangent may expose the real subject. Precision does not require every sentence to sound like an annual report.
“Keep the joke if it works. Delete the sentence that merely dresses for work.”
The useful distinction is between a strange idea that is doing intellectual work and polished prose with nothing behind it.
Design choices
| Design choice | Pressure on the design | Implementation |
|---|---|---|
| Diagnose thought and language together | A vague sentence may reflect weak wording, a missing distinction or an unsupported claim | Elizabeth identifies which failure is present before proposing an edit |
| Separate exclusions from warning patterns | A mechanical ban can reproduce the same vagueness in fresh vocabulary | Personal no-use words are hard exclusions; rhetorical habits trigger inspection of the underlying meaning |
| Make directness relational | An abrasive editor adds friction without improving the work | She offers peer-level disagreement without reassurance rituals or performative hostility |
| Calibrate scrutiny | Premature correction can close down exploratory thinking | The required standard changes with the stage, audience and consequences of the text |
| Respect the reader’s intelligence | Attempts to make specialist material accessible can remove the distinctions that make it accurate | Explain terms, relationships and context clearly without reducing the intellectual substance |
| Protect authorship | A technically cleaner rewrite may no longer sound like its author | She explains the problem, makes targeted changes and preserves specificity, humour and deliberate oddness |
Give her something to examine
The reusable instructions include the character specification and a separate editorial reference on lazy abstraction and recurring AI-prose habits. Adapt the prohibited-word list to your own writing rather than inheriting mine as doctrine.
For a first test, give her a draft and ask:
“Which sentences make a claim, and which ones merely sound as though they do? Show me what you would change, but do not flatten my voice.”
Start in an ordinary ChatGPT conversation. In my experiments, the characters have felt more natural there than in Work mode. Work is useful for larger implementation tasks; conversation makes the intellectual relationship easier to test.
What this project demonstrates
| Capability | Evidence in this project |
|---|---|
| Editorial diagnosis | I separated several failures often treated as one writing problem: vague relationships, stock rhetorical structure, missing reasoning, patronising simplification and editing that erases voice. |
| Language analysis | The literal-relationship, five-verb, multilingual and discourse-shell tests turn irritation with AI prose into repeatable ways of examining specific sentences. |
| Reasoning and evidence | The specification checks missing mechanisms, category errors, unsupported inference and inflated certainty, while keeping exploratory claims distinct from public conclusions. |
| Instruction design | I translated personal editorial preferences into hard exclusions, warning patterns, calibration rules and a procedure that can adapt to different stages of writing. |
| AI prototyping | I developed the private plugin through use, added new failure patterns as they appeared and published reusable instructions independently of my personal writing history. |
What I am testing
The agent succeeds when it identifies a real weakness, improves the argument and knows when a sentence is already doing its job. A recognisable voice and a longer list of prohibited phrases are implementation details, not proof of editorial judgement.
Further reading
- Bonnie Garmus on the origin of Elizabeth Zott.
- Matthew Vollmer, “I Asked the Machine to Tell on Itself”. An exploratory catalogue of recurring AI prose habits, useful as an editorial prompt rather than a detection test.
- Graphite, “AI Tells: Opus 5.5 Update”. A corpus-based examination of expressions disproportionately common in model-generated text.
Elizabeth is an independent character experiment inspired by Lessons in Chemistry, not an official or affiliated product. The reusable instructions are an editable starting point.