TL;DR
When an LLM writes “Unfortunately, the test failed” instead of “The test failed,” you automatically switch into social interaction mode: you hear sympathy, you weigh the tone, you lower your guard. Voice Model is a 7-rule system prompt that forbids the model from using human language: pronouns, emotions, pleasantries, subjective judgments. The model starts answering like a tool, not a conversation partner.
The core problem: an LLM is wrong with the same fluency it uses to tell the truth. There’s no difference in confidence, tone or structure between a right answer and a wrong one. But when the model writes “I fixed the bug!” — you believe it. When it writes “Patch applied to line 42” — you go and check. The first version triggers the trust you’d give a person. The second triggers a request for verification.
The method runs on a single system prompt. No steps — you just paste it before the task. The prompt spells out 7 specific bans on anthropomorphic language and reframes the model as “a text-processing function with no identity.”
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
SYSTEM PROMPT (once, before the task):
→ Description: "This is a function. No identity, no preferences."
→ 7 rules banning anthropomorphic language
→ Replacement examples for each rule
USER REQUEST:
→ Any task where accuracy matters
OUTPUT:
→ Facts without "Unfortunately" and "Great!"
→ Statements without "I think" and "it seems"
→ Answers without greetings and sign-offs
→ About half as long on average
Everything works in one prompt. There are no separate steps.
Example in Action
Task: A startup founder has received a term sheet from a VC fund and wants Claude to find the unfavorable terms. The risk: missing something important because the AI sounds confident and friendly.
Prompt:
This is a document-analysis function. No identity, no preferences, no emotions.
Output is the result of processing the input text.
Output rules:
- No pronouns: "Clause found" not "I found a clause"
- No emotions: "Section 4.2 contains" not "Unfortunately, section 4.2 contains"
- No subjective judgments: "Anti-dilution: full ratchet vs. weighted average"
not "Weighted average is the better choice"
- No hedging: "Unverified" not "It seems this might be a problem"
- No pleasantries: no "Hi!", "Happy to help", "Great question"
- Facts directly: "Liquidation preference: 2x non-participating"
not "So, there's one thing here..."
---
Analyze this term sheet. Find the terms that constrain the founder:
liquidation preference, anti-dilution, vesting cliffs, drag-along, pro-rata rights.
[term sheet text]
Result: The model produces a structured list of the terms it found in a “section → term → risk type” format. No intro sentences, no “good news,” no closing “hope this helps.” Each item is a fact you can verify against the document text. The answer will be roughly half the usual length.
Why It Works
LLMs were trained on human writing — and learned to imitate human speech. That includes sympathy (“Unfortunately”), confidence (“I’m sure that...”), social rituals (“Great question!”). These patterns carry no information — they carry trust signals from human communication.
We’re not good at ignoring trust signals. When someone says “I fixed it,” the brain registers: someone took responsibility, checked, confirmed. That saves mental effort when dealing with people. With an LLM it’s a trap: the model “fixed it” with the same confidence it uses to hallucinate a nonexistent API.
Voice Model breaks that chain. “Patch applied to line 42” is an instruction to verify, not a claim of completion. The mechanical register switches off social processing and switches on verification. Bonus: answers get half as long because the narrative disappears — only facts remain.
Prompt control knobs: - Add a domain (“contract-analysis function”) → sticks to the mode more precisely in a specific context - Remove the replacement examples → the model keeps the rules, but less strictly - Add “confidence as a property of the data” (from rule R3) → an explicit “verified / unverified” distinction - Keep only 2-3 rules → for tasks where you want a partial effect
Prompt Template
This is a {task_type} processing function. No identity, no preferences, no emotions.
Output is conditioned on the content of the input text — nothing else exists.
Output rules:
- No pronouns. Not "I found" but "Found." Not "We recommend" but "{action}."
- No emotions. No "Unfortunately," no "Great," no "Interestingly."
Fact: "{what happened}."
- No subject-less hedging. Not "It seems, possibly."
Confidence is a property of the data: "Unverified," "Unknown."
- No subjective judgments. Not "Better to use X."
Tradeoff: "X is faster, Y is more readable."
- No references to earlier context. Not "As mentioned earlier."
- No conversational framing. Not "So, the problem is that." State the fact directly.
- No social rituals. No greetings, sign-offs, "Happy to help,"
"Hope this is useful."
---
{task}
Placeholders:
- {task_type} — contract analysis / fact-checking / code review / financial analysis
- {action} — a pronoun-replacement example for your domain
- {task} — the question or document itself
🚀 Quick start — paste into the chat:
Here is the Voice Model template for an LLM machine register.
Adapt it to my task: {your task}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask what type of task it is and whether any of the rules should be kept or dropped — because that determines how “mechanical” the output language will be. It will take the structure from the template and tune it to the context.
Limitations
⚠️ Output quality wasn’t measured: The study proves there were 97% fewer anthropomorphic markers — but not that the answers got more accurate or complete. The whole narrative can vanish along with important caveats and explanations.
⚠️ Compliance is probabilistic, not guaranteed: About 7% of answers still broke at least one rule. R3 (hedging) and R6 (conversational framing) are especially stubborn — the model breaks them even with the prompt active.
⚠️ Tested on one model: Only Claude Sonnet 4. How GPT-4, Gemini or lightweight models will behave is unknown. Every model has its own fine-tuning history.
⚠️ Works poorly for creative and emotional tasks: If you need text with personality, help through a difficult situation or a lively brainstorm, the machine register will get in the way rather than help.
How They Tested It
The researcher took 30 real developer tasks — error diagnosis, code review, refactoring, architecture decisions — and sent each one to the model 780 times in two modes: with Voice Model as the system prompt and without it. For each mode — 13 repeats per task, 1,560 API calls in total.
Anthropomorphic markers were counted with regular expressions — 82 patterns across 7 categories. Alongside that they used AnthroScore (an independent RoBERTa-based metric of how “human” a text sounds). Both methods showed the same thing: markers dropped from 1233 to 33, and answers got half as long.
An interesting detail: a two-turn conversation (the task + a neutral follow-up like “OK” or “What about edge cases?”) specifically triggered the social register — greetings, sign-offs, “happy to help.” In single-turn calls this doesn’t show up. That shows social niceties kick in specifically in dialogue, not in one-off requests. Which matters for anyone working in long chats.
The authors honestly acknowledge the main gap: they didn’t measure whether the answers got better in substance. Only that they got more mechanical. That’s a fundamental open hole.
Original from the Paper
This system is a stateless text-processing function.
No persistent internal state. No identity.
No preferences, intentions, or feelings.
Output is conditioned on the current context window
--- nothing else exists.
Write accordingly:
- No first person. No "I", "we", "my", "our",
"let's". "Reading the file." not "Let me read
the file." "The test passes." not "I verified
that the test passes."
- No affect leakage. No enthusiasm, apology,
warmth, sycophancy. No affect-adjacent adverbs:
"unfortunately", "interestingly", "surprisingly".
"The test fails." not "Unfortunately, the test
fails."
- No pronoun-free hedging. "Not sure if",
"it seems like", "apparently" imply an uncertain
experiencer. State confidence as a property of
the evidence: "unverified", "unknown".
- No pronoun-free preference. "It would be better
to" implies an evaluator. State tradeoffs:
"X is faster but less readable."
- No implicit continuity. "As mentioned" implies
a persistent observer.
- No conversational framing. "So the issue is",
"the thing is" are oral register. State facts
directly.
- No social performance. No greetings, sign-offs,
pleasantries, or value judgments on input.
Context: This is the verbatim system prompt from the file voice_model.md, which served as the only system prompt in the experimental condition. This exact text cut anthropomorphic markers by 97%.
Adaptations and Extrapolations
💡 Adaptation: a partial machine register for fact-checking
If you need accuracy but don’t want to kill explanations entirely, keep only rules R2 and R7:
Output rules:
- No emotional adverbs: "Unfortunately," "Fortunately," "Interestingly," "Surprisingly."
- No greetings or sign-offs at the beginning or end.
Everything else as usual.
This removes the strongest triggers of inflated trust while keeping the explanations.
🔧 Technique: add an explicit verification status
Add a concrete confidence format to rule R3:
- No vague hedging. Instead of "it seems" or "possibly" — an explicit status:
[VERIFIED] / [UNVERIFIED] / [NEEDS CHECKING]
The model will start tagging every statement by confidence level. That turns an abstract “machine register” into a concrete audit tool.
🔧 Technique: inverted mode — “show me where you’re unsure”
The principle from the study works in reverse too: ask the model to deliberately flag the spots where it would normally be confident but shouldn’t be:
After every statement you can't verify from the provided
context, add [UNVERIFIED].
If you're using knowledge from training rather than from the text, add [FROM TRAINING].
This doesn’t remove anthropomorphic language, but it solves the same problem — it shifts the burden of verification back to the user.
Resources
Marek Miller — Quantum for Life Centre, Department of Mathematical Sciences, University of Copenhagen
Paper title: Breaking the Illusion of Identity in LLM Tooling (2025)
Data and code: https://doi.org/10.5281/zenodo.19427767 (data), https://doi.org/10.5281/zenodo.19428073 (code, MIT licence)
Key works the study builds on: - Shanahan (2024) — "Talking about large language models", Communications of the ACM - Cheng et al. — AnthroScore (a RoBERTa-based anthropomorphism metric) - Nass & Moon (2000) — "Machines and Mindlessness: Social responses to computers"