LLM output sentiment analyzer
Generated copy can drift more positive, negative, or flat than you intended. This tool scores the sentiment of LLM output against a bundled AFINN-style lexicon, all in the browser, and shows which words are pushing the tone each way — a fast way to sanity-check the emotional register of a response before you use it.
When tone checking matters
LLMs are not neutral tone generators. Their default register depends on the instruction, the model, and subtle statistical properties of the training data. Common drift patterns:
- Customer-facing copy is overly positive. Models default toward upbeat, enthusiastic language even when the brief called for measured, balanced copy. High net positive scores on a product description are often a sign of hollow superlatives worth editing.
- Rejection or refusal emails read as neutral but feel cold. Negative-leaning words are absent, but so is warmth — the net score near zero can mask an emotionally flat response.
- Risk or compliance content reads too alarming. Prompts that mention regulatory risk sometimes produce output heavy with negative-valence words even when a balanced treatment was intended.
How it works
The text is tokenised into words, and each word is looked up in an AFINN-style lexicon that assigns integer valence scores from −5 (very negative) to +5 (very positive). A simple negation rule flips a word’s sign when a negator like “not” or “never” precedes it within a few words. Positive and negative scores are summed separately, combined into a net score, and normalised by the count of scored words to give an intensity figure and an overall label.
Reading the scores
The tool shows three numbers: a raw positive sum, a raw negative sum, and a net. An intensity figure normalises the net by the number of scored words so short and long texts are comparable. A rough guide:
- Net strongly positive (intensity above 0.5): copy is enthusiastic or encouraging — check whether that fits the context.
- Net near zero: balanced, cautious, or emotionally flat — may lack warmth for supportive contexts.
- Net negative: alarming or critical — appropriate for warnings, may be excessive for nuanced topics.
The word list below the score shows the top contributors in each direction — the most informative part of the output, since a single high-valence word can swing a short passage.
Tips and notes
- Use enough text. A few words give noisy results; a paragraph or more produces a steadier reading.
- Sarcasm and irony fool it. Lexicon methods read “oh great, another bug” as positive — verify edge cases by eye.
- Domain matters. Words neutral in one field can be loaded in another; treat the score as relative across your own outputs rather than absolute.
- Compare versions. Running the same prompt’s outputs through the analyzer is a cheap way to see whether a prompt edit shifted the tone.