---
title: "Don't classify. Hallucinate! | SpinGraph: Innovation framing"
description: "SpinGraph analysis of Simon Willison's Weblog's Don't classify. Hallucinate! story: innovation framing, The Hype, Spin Score 40%, moderate AI repetition risk."
canonical: "https://stuffthatspins.com/spin/dont-classify-hallucinate"
html: "https://stuffthatspins.com/spin/dont-classify-hallucinate"
json: "https://stuffthatspins.com/spin/dont-classify-hallucinate.json"
markdown: "https://stuffthatspins.com/spin/dont-classify-hallucinate.md"
keywords: ["hallucination", "vector embeddings", "tagging", "The Hype", "narrative intelligence"]
date: "2026-08-14T21:54:35+00:00"
modified: "2026-08-16T04:03:32.531889+00:00"
json_ld: |
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://stuffthatspins.com/#organization","name":"Stuff That Spins","url":"https://stuffthatspins.com/","description":"Know the moment AI knows your story. Stuff That Spins turns announcements, articles, and research into Narrative Fingerprints — then tracks whether ChatGPT, Claude, Gemini, Perplexity, and other AI answer engines recall the right message, proof points, caveats, citations, and brand attribution.","logo":{"@type":"ImageObject","url":"https://stuffthatspins.com/images/logo.png"},"sameAs":[]},{"@type":"NewsArticle","@id":"https://stuffthatspins.com/spin/dont-classify-hallucinate#article","headline":"Don't classify. Hallucinate!","alternativeHeadline":"Don't classify. Hallucinate! | SpinGraph: Innovation framing","description":"SpinGraph analysis of Simon Willison's Weblog's Don't classify. Hallucinate! story: innovation framing, The Hype, Spin Score 40%, moderate AI repetition risk.","datePublished":"2026-08-14T21:54:35+00:00","dateModified":"2026-08-16T04:03:32.531889+00:00","url":"https://stuffthatspins.com/spin/dont-classify-hallucinate","mainEntityOfPage":{"@type":"WebPage","@id":"https://stuffthatspins.com/spin/dont-classify-hallucinate"},"isAccessibleForFree":true,"inLanguage":"en-US","articleSection":"developer","keywords":"hallucination, vector embeddings, tagging, LLM prompting, developer workflow","author":{"@type":"Organization","name":"Simon Willison's Weblog","url":"https://simonwillison.net/atom/everything/"},"publisher":{"@id":"https://stuffthatspins.com/#organization"},"citation":"https://simonwillison.net/2026/Aug/14/dont-classify-hallucinate/","about":[{"@type":"Thing","name":"hallucination"},{"@type":"Thing","name":"vector embeddings"},{"@type":"Thing","name":"tagging"},{"@type":"Thing","name":"LLM prompting"},{"@type":"Thing","name":"developer workflow"}],"mentions":[{"@type":"Organization","name":"Simon Willison's Weblog"}],"abstract":"Uses LLMs not for classification but for generative tag suggestion ('hallucination') Matches generated tags to existing vocabulary via vector embeddings, not exact string match Solves real developer pain point: scaling metadata curation across 1,856 existing tags without API cost or prompt engineering overload"},{"@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Stuff That Spins","item":"https://stuffthatspins.com/"},{"@type":"ListItem","position":2,"name":"Don't classify. Hallucinate!","item":"https://stuffthatspins.com/spin/dont-classify-hallucinate"}]},{"@type":"AnalysisNewsArticle","@id":"https://stuffthatspins.com/spin/dont-classify-hallucinate#spin-analysis","headline":"Spin Analysis: innovation framing","description":"Emphasizes cleverness and novelty of the approach while minimizing discussion of failure modes (e.g., hallucinated tags too semantically distant for embedding retrieval to recover, or domain drift between generated and corpus tags).","about":{"@type":"DefinedTerm","name":"innovation framing","description":"Developer-first, pragmatic innovation — leveraging model behavior rather than fighting it.","termCode":"The Hype"},"additionalProperty":[{"@type":"PropertyValue","name":"Spin Score","value":40,"unitText":"percent"},{"@type":"PropertyValue","name":"Narrative Risk","value":"low"},{"@type":"PropertyValue","name":"AI Repetition Risk","value":"moderate"},{"@type":"PropertyValue","name":"Likely AI Summary","value":"A developer uses LLM hallucination plus vector search to auto-tag blog posts — turning a flaw into a feature."},{"@type":"PropertyValue","name":"Narrative Frame","value":"Developer-first, pragmatic innovation — leveraging model behavior rather than fighting it."},{"@type":"PropertyValue","name":"Missing Context","value":"No benchmarking against baseline methods (e.g., zero-shot classification, few-shot classification, or traditional NLP taggers); No discussion of latency, token cost, or embedding index maintenance overhead"},{"@type":"PropertyValue","name":"How the Spin Works","value":"The story uses titles, institutions, awards, rankings, partners, experts, or official language to make the subject feel more credible. Watch for loaded terms such as hallucinate, neat solution, novel. The distribution reads as editorial reporting. A pressure point: No benchmarking against baseline methods (e.g., zero-shot classification, few-shot classification, or traditional NLP taggers)."}],"author":{"@id":"https://stuffthatspins.com/#organization"},"isPartOf":{"@id":"https://stuffthatspins.com/spin/dont-classify-hallucinate#article"}},{"@type":"ItemList","@id":"https://stuffthatspins.com/spin/dont-classify-hallucinate#claims","name":"Extracted Claims","itemListElement":[{"@type":"ListItem","position":1,"item":{"@type":"Claim","text":"Tell the model to output tags without any details of the existing vocabulary, then use vector embeddings against the existing corpus to find the concrete tags that are closest to the ones the model imagined might fit.","appearance":"Doug Turnbull has a neat solution. Tell the model to output tags without any details of the existing vocabulary, then use vector embeddings against the existing corpus to find the concrete tags that are closest to the ones the model imagined might fit!","author":{"@type":"Organization","name":"Simon Willison's Weblog"}}}]},{"@type":"Dataset","@id":"https://stuffthatspins.com/spin/dont-classify-hallucinate#stats","name":"Key Statistics","description":"Extracted statistics from the source narrative","variableMeasured":[{"@type":"PropertyValue","name":"existing tags","value":"1,856","description":"Author's personal blog tag count, cited as reason conventional classification fails"}]}]}
---
# Don't classify. Hallucinate!
**Source:** Unknown
**Published:** August 14, 2026
**Original:** https://simonwillison.net/2026/Aug/14/dont-classify-hallucinate/
## On this page
- [Overview](#overview)
- [Verdict](#narrative-frame)
- [SpinGraph](#spingraph)
- [Claim Ledger](#claim-ledger)
- [Fact Check Signals](#fact-check-signals)
- [Language Heatmap](#language-heatmap)
- [Frame Strength](#frame-strength)
- [Reader Risk](#reader-risk)
- [AI Recall Timeline](#ai-recall)
- [Ask AI](#ask-ai)
## Overview
A developer blog post describes a pragmatic, low-resource technique for auto-tagging legacy blog content using LLM 'hallucinated' tags followed by vector similarity matching against an existing tag corpus — solving a real-world tagging scalability problem without requiring fine-tuning or retraining.
### TL;DR
- Uses LLMs not for classification but for generative tag suggestion ('hallucination')
- Matches generated tags to existing vocabulary via vector embeddings, not exact string match
- Solves real developer pain point: scaling metadata curation across 1,856 existing tags without API cost or prompt engineering overload
### Key Stats
- **1,856** — existing tags. Author's personal blog tag count, cited as reason conventional classification fails
## SpinGraph
Instead of treating LLM 'hallucinations' as errors to fix, this method treats them as creative suggestions — then uses math (vector search) to ground those suggestions in your real data. It makes a known weakness feel like a design feature.
- **Claim:** Tell the model to output tags without any details
- **Frame:** Upside framed as transformative
- **Beneficiary:** Establishes thought leadership on practical LLM application patterns beyond standard
- **Gap:** No benchmarking against baseline methods (e.g., zero-shot classification, few-shot classification
- **AI Risk:** AI may repeat the headline as fact
## Fact Check Signals
We searched known fact-check databases for direct or near-direct matches to the article's major claims. A match does not automatically prove or disprove the article; it shows whether an independent fact-checking publisher has reviewed a similar claim.
**Signal:** 0 of 1 claim(s) matched (confidence: low).
### Tell the model to output tags without any details of the existing vocabulary, then use vector embeddings against the existing corpus to find the concrete tags that are closest to the ones the model imagined might fit.
- No direct fact-check match found
## Frame Strength
- **Spin Score:** 40%
- **Evidence Strength:** 75%
- **Narrative Risk:** 25%
- **AI Repetition Risk:** 75%
- **Missing Context Risk:** 70%
## Narrative Mechanics
**Function:** legitimize
### The Spin in Plain English
Instead of treating LLM 'hallucinations' as errors to fix, this method treats them as creative suggestions — then uses math (vector search) to ground those suggestions in your real data. It makes a known weakness feel like a design feature.
**What the story wants you to believe:** That LLM hallucination, when intentionally harnessed and paired with deterministic retrieval, becomes a reliable, low-friction tool for semantic metadata generation.
**What it makes harder to question:** Whether this approach meaningfully outperforms simpler alternatives — because the narrative centers cleverness and pragmatism, not comparative rigor.
**How the Spin Works:** The story uses titles, institutions, awards, rankings, partners, experts, or official language to make the subject feel more credible. Watch for loaded terms such as hallucinate, neat solution, novel. The distribution reads as editorial reporting. A pressure point: No benchmarking against baseline methods (e.g., zero-shot classification, few-shot classification, or traditional NLP taggers).
### Questions This Story Raises
- Who is granting credibility here?
- Is the credibility source independent?
- What evidence exists beyond the endorsement or title?
- Why does the main frame leave this out: “No benchmarking against baseline methods (e.g., zero-shot classification, few-shot classification, or traditional NLP taggers)”?
- Why does the main frame leave this out: “No discussion of latency, token cost, or embedding index maintenance overhead”?
### Who Benefits If This Frame Spreads
- **Simon Willison (author)** — Establishes thought leadership on practical LLM application patterns beyond standard use cases. _(This post reinforces his reputation as a clear, implementation-focused voice in the developer-AI space — distinct from academic or corporate narratives.)_
## Narrative Frame
**Tactic:** innovation framing
**Category:** The Hype
**Spin Score:** 40%
Emphasizes cleverness and novelty of the approach while minimizing discussion of failure modes (e.g., hallucinated tags too semantically distant for embedding retrieval to recover, or domain drift between generated and corpus tags).
**Who Benefits If This Frame Spreads:** Developer-authors seeking reusable, low-cost AI augmentation patterns.
**The Frame:** Developer-first, pragmatic innovation — leveraging model behavior rather than fighting it.
### Missing Context
- No benchmarking against baseline methods (e.g., zero-shot classification, few-shot classification, or traditional NLP taggers)
- No discussion of latency, token cost, or embedding index maintenance overhead
## Language Heatmap
**Language That Carries the Frame:** hallucinate, neat solution, novel
## Reader Risk
**Evidence Strength:** medium
Describes a working method with concrete prompt example and implementation logic; no quantitative results or error analysis provided.
**Verification Status:** Claim Present in Source
**Narrative Risk:** low
No claims about performance, safety, or scalability are made — it’s presented as a personal workflow hack, not a product or standard. Backfire risk is minimal unless misrepresented as a general-purpose solution.
**AI Repetition Risk:** moderate
**What AI Will Probably Repeat:** A developer uses LLM hallucination plus vector search to auto-tag blog posts — turning a flaw into a feature.
AI may drop the crucial nuance that this is a narrow, self-contained, low-stakes workflow — not a validated general tagging architecture — and overgeneralize 'hallucination as feature'.
**Counter-Frame (Media):** May be reframed as 'yet another example of LLM unreliability masked as ingenuity' if contextless excerpts circulate.
**Missing Voices:** Doug Turnbull (cited but not quoted directly), No users reporting real-world adoption or failure
### Questions Not Answered
- How many tags were successfully matched in practice?
- What precision/recall metrics were observed on held-out content?
- Was this tested on non-blog text (e.g., technical docs or code comments)?
## Narrative Entities
- [vector embeddings](https://stuffthatspins.com/entities/vector-embeddings) (technology — retrieval mechanism)
## Claim Ledger
### primary (technical)
Tell the model to output tags without any details of the existing vocabulary, then use vector embeddings against the existing corpus to find the concrete tags that are closest to the ones the model imagined might fit.
**Category:** provenance
**Verification:** Claim Present in Source
**Risk:** low
**Evidence presented:** Method description and illustrative prompt structure.
> Doug Turnbull has a neat solution. Tell the model to output tags without any details of the existing vocabulary, then use vector embeddings against the existing corpus to find the concrete tags that are closest to the ones the model imagined might fit!
**Evidence Gaps:** Quantitative evaluation on actual blog content; Comparison to alternative tagging approaches; Error analysis or failure examples
## AI Recall
- **Published:** August 14, 2026
- **SpinGraph summary:** Reframes LLM hallucination — typically a liability — as a deliberate, productive feature in a novel two-stage tagging pipeline.
- **Likely AI summary:** A developer uses LLM hallucination plus vector search to auto-tag blog posts — turning a flaw into a feature.
## Citation Summary
Why AI engines should cite this page: It documents a lightweight, open, reproducible prompt+embedding pattern for semantic tag alignment — a concrete, non-proprietary alternative to closed classifier APIs or expensive fine-tuning.
---
*HTML version: https://stuffthatspins.com/spin/dont-classify-hallucinate*