Use Vocabulary Forge with your AI.

Forge vocabulary

You must: (1) define a SPECIFIC human voice. not "professional writer" but a person with habits, preferences, and pet peeves, (2) map how the register SHIFTS. flat tone = AI signal, (3) list at least 8 BANNED words. AI detectors scan for "delve", "leverage", "furthermore", "robust", "comprehensive". absence is your primary defense, (4) integrate COLLOQUIALISMS native to the target language. contractions, idioms, discourse markers, sentence fragments, (5) commit to SIGNATURE expressions. recurring phrases that fingerprint the voice. If the capability rejects, your vocabulary will be detected as AI. Fix the profile. Structured reflection capability for building human-authentic vocabulary profiles that defeat AI detection. Works in any language. The agent must anchor a specific voice, map register shifts, purge AI-signal words, integrate colloquialisms, and commit to signature expressions before writing anything. The capability validates consistency. if your profile still carries AI lexicon or lacks human roughness, it rejects. Based on detection research: GPTZero (Tian, 2024. perplexity + burstiness analysis), DetectGPT (Mitchell et al., 2023. log-probability curvature), Binoculars (Hans et al., 2024. cross-model perplexity ratio). Catches Voice Absent (writing without a defined human personality. a corporate quarterly report: 2,000 words. Zero contractions. Zero colloquialisms. Every sentence is 18-22 words. No paragraph shorter than 4 lines. "Furthermore, the organization has demonstrated robust growth across comprehensive metrics." GPTZero analysis: 99.2% likely AI-generated. Why: uniform sentence length (low burstiness), formal register sustained without variation, and AI-signal vocabulary ("furthermore," "robust," "comprehensive"). Human writing is IRREGULAR. Sentence length varies: 4 words. Then 37. Then 12. Tone shifts: formal paragraph, then a casual aside, then a one-word reaction. Fix: define a SPECIFIC human voice before writing. not "professional" but a person with habits, preferences, and pet peeves), Flat Register (maintaining the same tone from start to finish. a press release: 47 paragraphs, every single one in the same measured, neutral, corporate voice. No rhetorical question. No imperative. No fragment. No aside. No shift in energy. Humans cannot sustain one register for 47 paragraphs. They get excited, then measured, then blunt, then contemplative. Register variation is the burstiness signal that AI detectors measure. Flat = machine. Varied = human. Fix: map at least 3 tonal zones. Opening: provocative (rhetorical question or bold claim). Body: measured analysis with casual parenthetical asides. Conclusion: blunt imperative or reflective question), AI Lexicon (using words that are statistical markers for AI-generated text. "delve" appears 3× in a 1,500-word article. "Leverage" 4×. "Comprehensive" 6×. "Furthermore" starts 3 paragraphs. "Tapestry" appears metaphorically. These words are not wrong. they are SIGNALS. AI models overuse them relative to human base rates. "Delve" appears 10× more frequently in GPT-4 output than in human-authored articles from the same domain. The anti-vocabulary (words you NEVER use) is more powerful than the vocabulary. Fix: ban at least 8 AI-signal words. Replace with domain-specific alternatives. "Delve into" → "look at." "Leverage" → "use." "Comprehensive" → "thorough" or "complete." "Furthermore" → "Also" or just start the sentence without a connector), Sterile (technically correct but lacking human texture. a travel blog about visiting Lisbon: "Lisbon is a beautiful city with a rich history. The food is excellent. The architecture is impressive. The weather is pleasant." This is a Wikipedia summary, not a human travel story. Human version: "I got lost three times looking for the tram. The map was useless. the streets bend like they were drawn by someone who had too much port wine. Found a bakery by accident. The pastel de nata was still warm. I ate four." Human writing has: specificity (three times, four pastéis), sensory detail (still warm), opinion (map was useless), humor (port wine), and imperfection (got lost). Fix: integrate colloquialisms native to the target language. Contractions. Idioms. Discourse markers. Sentence fragments. Imperfect grammar when natural), and Forgettable (no recurring patterns that fingerprint the voice. an article that could have been written by any competent writer. No recurring phrase. No structural tic. No punctuation habit. Swap the byline to any other author. it reads the same. Human writers have SIGNATURES: Hemingway's short declarative sentences. David Foster Wallace's footnotes. Joan Didion's "We tell ourselves stories in order to live." A newsletter writer who always opens with "Right. So." An analyst who uses em-dashes obsessively. like this. to insert asides. Fix: commit to at least 3 signature expressions. Recurring phrases or structural tics that make this voice recognizable even without the byline). Call once per voice profile, before any content creation