When a manager claims they said something different in a meeting, or denies making a promise you remember clearly, AI can search your records and surface earlier communications where they said the opposite thing. The contradictions become visible as a pattern, not an isolated he-said-she-said dispute.
Semantic similarity is the measure of how much two pieces of text mean the same thing, even if they use completely different words. AI systems use embeddings—mathematical representations of meaning—to assess similarity. In workplace gaslighting contexts, semantic similarity becomes a detective tool for finding contradictions that aren't immediately obvious because the vocabulary changed.
A toxic manager says in July: "You need to improve your execution discipline." In September, they claim: "I've never expressed concerns about your work quality." Technically, these use different words. But semantically, they're contradictory. Execution discipline IS a concern about work quality. A human might connect these dots, but when you're overwhelmed and gaslit, semantic connections sometimes blur. AI can surface them systematically.
This is particularly dangerous because gaslighters often shift vocabulary intentionally. They deny saying X while claiming Y, betting you won't recognize that X and Y are semantically equivalent. "I never said you were disorganized" (denies) vs. their earlier comment: "Your project structure is chaotic" (same meaning, different words).
Modern AI systems like Claude use embedding models that convert text into high-dimensional vectors (essentially lists of numbers that represent meaning). Two pieces of text with similar embeddings are semantically similar, even if they look different on the surface.
The technical process: Manager Statement A gets converted to an embedding. Manager Statement B gets converted to an embedding. The system calculates the similarity between these vectors (cosine similarity is common). High similarity = semantic equivalence despite word choice differences.
Create a system where you list all the claims a manager has made about you: "You don't take feedback well", "You're resistant to change", "You struggle with adaptability", "You need to be more open-minded". Semantically, these are very similar claims expressed in different ways.
Feed this list to Claude: "Compare these statements for semantic similarity. Group them by core meaning. Show me where the same fundamental claim is repeated with different wording." The AI will identify that all five statements essentially mean "You're resistant to feedback", even though the wording varies.
Then when the manager later denies they said you were resistant—"I said you should be more adaptable, not that you're resistant"—you have the semantic analysis showing these are the same claim. Your documentation becomes irrefutable.
Use semantic similarity to track how a manager's claims about you evolve contradictorily. They might say in early performance reviews: "Your strength is independent problem-solving." Six months later, in a performance improvement plan: "You need to collaborate more instead of working independently." These aren't just different—they're contradictory.
Ask Claude to analyze statements semantically: "Do these two statements about my working style support each other, contradict each other, or address different aspects?" If they're contradictory, the AI surfaces that explicitly, and you have documented evidence that the manager's narrative about you shifted inconsistently.
Semantic similarity isn't perfect. Sometimes statements are genuinely different despite sounding similar. "You should collaborate more" and "You sometimes over-rely on others' input" might both involve collaboration but mean opposite things. This is why AI semantic analysis should inform your thinking, not replace human judgment.
The best use case: AI identifies potential contradictions through semantic similarity, then you manually review whether they're real contradictions or just surface-level similarities. This prevents both false negatives (missing contradictions) and false positives (over-claiming contradictions).
Gaslighting often involves contradictions across multiple statements. Manager claims: "I've always believed in your potential" (Statement 1) + "Your performance has never met our standards" (Statement 2) + "I give all my employees equal opportunities" (Statement 3). Semantically analyzing all three together shows they're mutually contradictory.
Use Claude with semantic analysis to identify these patterns: "Show me all the core claims this manager has made about me. Group semantically similar claims. Then identify where different groups contradict each other." The system becomes a contradiction-detection engine.
The most powerful use case combines semantic similarity with temporal analysis. A manager makes a claim, you counter it, they make a semantically similar claim in different words, claiming it's a new point. The documentation system catches that claim-counter-rephrase pattern, proving they're arguing the same point repeatedly rather than engaging with your counter.
For example: "You don't follow instructions" (Claim 1) → You provide evidence of following instructions → "You sometimes miss nuances in direction" (Claim 2, semantically similar to Claim 1) → You respond → "You need better attention to detail" (Claim 3, semantically equivalent to 1 and 2). The pattern is recursive gaslighting, and semantic analysis exposes it.
Try this: Think of a workplace conflict where a manager made similar criticisms in different words. Write down 4-5 variations of the criticism as they expressed it. Paste these into Claude and ask: "Which of these statements are semantically similar? Do they mean roughly the same thing despite different wording? Group them by core meaning." You'll immediately see how AI surfaces the pattern you might not consciously recognize. This is the foundation of using semantic similarity for contradiction detection—making obvious what gaslighting obscures.
Peri can explain this concept, give practical examples, help you decide whether it applies to your situation, or recommend a journey if appropriate.
Explore related journeys or tell Peri what you're working through.