# Hedging analyser: how it is scored and how to change it > Confidence vs hedging (confidence_vs_hedging) rates how declarative writing is on a 0 to 1 scale: 0 means heavy hedging (maybe, perhaps, I think) and 1 means hi Canonical: https://scriptgrain.com/attributes/confidence-vs-hedging # Hedging analyser *By Jack Stovell ยท published 2026-09-24* Confidence vs hedging (confidence_vs_hedging) rates how declarative writing is on a 0 to 1 scale: 0 means heavy hedging (maybe, perhaps, I think) and 1 means highly declarative. A ScriptGrain voice profile stores it as a model's judgement of your samples, not a code count, and it moves the voice match score with a weight of 1.5. ## Analyse your confidence vs hedging free Your free voice profile measures confidence vs hedging among all 45 attributes, from a sample of your own writing. No card needed. [Try the six-signal check first](https://scriptgrain.com/#sg-measure) ## At a glance - API field: **confidence_vs_hedging** - How it is measured: **Judged by an AI model** - Group: **Tone and register** (ScriptGrain voice profile, 45 attributes in 8 groups) - Scale: **0 to 1 (0 heavy hedging, 1 highly declarative)** (ScriptGrain attribute definitions) - Voice match: **Scored, weight 1.5** (ScriptGrain voice-match engine (published method)) ## What confidence and hedging measures This attribute measures the gap between stating a claim and softening it. At 0, your sentences lean on maybe, perhaps, I think. At 1, you say the thing plainly and move on. Plenty of writers sit somewhere between, and that's fine: the scale exists to describe where you land, not to tell you where you should be. It sits in the [tone and register group](https://scriptgrain.com/attributes/tone-and-register) with five other attributes, including [formality score](https://scriptgrain.com/attributes/formality-score) and [emotional expressiveness](https://scriptgrain.com/attributes/emotional-expressiveness). Confidence vs hedging is specifically about certainty markers: words and constructions that either commit to a claim or leave the door open. "This works" sits at the declarative end. "This might work, I think" sits near the hedging end. Neither is wrong. A legal disclaimer wants hedging. A product pitch wants declaratives. What trips people up: this isn't counted. No list of hedge words is tallied against your draft to produce this number. A fixed list of 16 does exist (may, might, perhaps, seems, seemed, appears, appeared, roughly, arguably, possibly, probably, fairly, somewhat, think, suspect, likely), but it drives the "Rhetoric" signal in the homepage brand check, which reports hedge words per 100 words, a different unit. For the profile attribute itself, an extraction model reads your samples and judges the value. ## How confidence and hedging is scored Confidence vs hedging enters your voice match score with a weight of 1.5. When ScriptGrain checks a draft against your profile, Claude Haiku (claude-haiku-4-5) judges the draft's value from 0 to 1, and code compares it with your profile's: similarity is 1 minus the gap divided by a tolerance of 0.35, floored at 0. Small gap, high similarity. A gap of 0.35 or more earns nothing. So if your profile says 0.7 (fairly declarative) and a generated draft comes back at 0.3 (hedgy, apologetic, full of "it could be argued"), that's a 0.4 gap against a 0.35 tolerance. The similarity floors at zero. That's a meaningful hit given the 1.5 weight. One honest caveat: these judged values vary slightly run to run. Claude Haiku is making a call each time, the way a person reading for tone would, and people don't give identical answers twice. For the full mechanics of how judged and counted features combine into one score, see the [voice measurement framework](https://scriptgrain.com/reference/voice-measurement-framework) and the [voice match](https://scriptgrain.com/glossary/voice-match) glossary entry. ## How to analyse confidence and hedging in your own writing You can do this by hand in about five minutes with any sample of your own writing. 1. Pull a sample of your writing, at least a few hundred words, ideally something typical rather than your best or worst piece. 2. Read it once for content, ignoring style entirely. 3. Read it again, this time underlining every hedge word or phrase: maybe, perhaps, I think, it seems, arguably, somewhat, probably, and similar softeners. 4. Count how many sentences make a bare claim with no qualifier attached, versus how many wrap the claim in a hedge. 5. Ask yourself, honestly, whether each hedge earns its place: is the underlying claim actually uncertain, or are you just softening out of habit? That last step matters more than the count. A hedge attached to genuine uncertainty is doing its job. A hedge attached to something you're actually sure of is just noise. If you want a number to track, count hedge words per 100 words: the homepage brand check does exactly that with its list of 16, and reads more than 1.2 per 100 as "You qualify claims before making them". That is a different unit from the profile's 0 to 1 value. ScriptGrain's free voice profile reads confidence vs hedging itself from a writing sample among all 45 attributes. ## Examples of confidence and hedging in real writing Example: "This approach works." That's declarative, no qualifier, sits near 1 on the scale. Example: "This might work, though I'm not entirely sure." That's hedged twice over, sits near 0. Example: "This usually works, except when the data's incomplete." That's a middle case: a plain claim with a stated condition attached, rather than a vague qualifier. These placements are illustrations, not typical values from any corpus. ## How to make AI writing more confident Getting an AI assistant to state things plainly instead of hedging is mostly about what you ask it to cut. 1. Ask it to state the claim first, then justify it, rather than building up to a soft conclusion. 2. Tell it to cut qualifiers you wouldn't personally use: perhaps, arguably, somewhat. 3. Replace constructions like "it could be argued that" with the argument itself, stated directly. 4. Keep a hedge only where the evidence is genuinely uncertain, and say what that uncertainty is, rather than hedging by reflex. When ScriptGrain writes from your profile, confidence vs hedging goes into the voice profile JSON block of the system prompt as a hard constraint. The generation prompt also bans the symmetrical "perhaps X, perhaps Y" hedge and hedged abstract closes such as "something approaching" or "perhaps enough after all". The profile reaches ChatGPT, Claude and other tools through the API and MCP server, so they can read the same value. ## How to make AI writing more careful and qualified Sometimes you want the opposite: an assistant that stops sounding falsely certain and starts qualifying claims the way you actually do. 1. Ask it to qualify claims using your own hedge words, not a generic list, so "I reckon" or "probably" shows up if that's genuinely your habit. 2. Attach a source or a condition to each strong claim, rather than leaving it bare. 3. Allow "I think" and "probably" specifically where your own writing samples use them, and nowhere else. A ScriptGrain profile applies a low value the same way it applies a high one: as a hard constraint in the voice profile block, not a tone to approximate. Fair enough if that sounds strict; a number is easier to hold a draft to than a vibe. ## Questions ### What is a hedging analyser? A hedging analyser is a tool or method for reading how much a piece of writing qualifies its claims versus stating them plainly. In ScriptGrain's system, this shows up as the confidence_vs_hedging attribute, judged by an extraction model reading your writing samples, scaled 0 (heavy hedging) to 1 (highly declarative), stored inside the tone and register group of a voice profile. ### How is hedging scored? It's judged, not counted. Claude Haiku reads a draft and assigns a confidence vs hedging value from 0 to 1, and code compares it with your stored profile value: similarity is 1 minus the gap divided by 0.35, floored at 0. That similarity feeds your overall voice match at a weight of 1.5. Judged values vary slightly from run to run. ### What are hedge words in writing? Hedge words are terms that soften a claim rather than stating it outright: maybe, perhaps, might, seems, arguably, possibly, probably, somewhat, think, suspect, likely, roughly, fairly. A related but separate ScriptGrain signal, the homepage "Rhetoric" check, counts hedge words from a fixed list of 16 per 100 words; the stored profile attribute itself is judged, not tallied. ### How do I make AI writing more confident? Ask the assistant to state its claim first, cut reflexive qualifiers like perhaps and arguably, and replace "it could be argued" with the argument itself. Keep a hedge only where there's real uncertainty, and say what that uncertainty is. When ScriptGrain writes, it sends your confidence_vs_hedging value as a hard constraint in the system prompt, and the prompt bans hedged abstract closes and symmetrical hedging pairs. ### How do I make ChatGPT stop hedging or sound less assertive? For less hedging: ask it to lead with the claim and drop soft qualifiers unless the evidence warrants them. For less assertive, more qualified writing: ask it to attach a source or condition to strong claims and to use your own natural hedge words rather than a generic list. Formality score and emotional expressiveness shape how a claim lands too. ## Related - [Tone and register attributes](https://scriptgrain.com/attributes/tone-and-register) - [Formality score](https://scriptgrain.com/attributes/formality-score) - [Humour register](https://scriptgrain.com/attributes/humour-register) - [Emotional expressiveness](https://scriptgrain.com/attributes/emotional-expressiveness) - [Tone and register, measured: the attributes behind 'that sounds like us'](https://scriptgrain.com/reference/tone-and-register) - [Homepage brand check signal](https://scriptgrain.com/) ## Sources - [How ScriptGrain scores voice match](https://scriptgrain.com/reference/voice-measurement-framework)