Measuring Certainty Inflation in LLM Medical Summaries with Clinician-Grounded Hedge Scales

When a language model summarizes a medical review, it can get every number right and still mislead. It does this by stating firmly what the source only suggested. "Low-certainty evidence suggests the drug may reduce falls" becomes "the drug reduces falls." Clinicians grading LLM summaries of Cochrane reviews named this error the "certainty illusion" and ranked it among the most dangerous. Today it can only be caught by expert reading, because overlap metrics such as ROUGE barely register it (Evaluating large language models on medical evidence summarization). This note proposes a ruler for that error. We ask clinicians how certain each hedging phrase sounds to them, in clinical context. We record not one number per phrase but its typical reading, how wide that reading is, and how much clinicians disagree. We then score a summary by how far it moves each claim along that ruler relative to the source.

Two things become possible with such a ruler. First, we can test directly whether an LLM reads "may," "probably," or "is associated with" the way clinicians do. If it does not, then using an LLM as the judge of certainty inflation builds the model's own overconfidence into the evaluation. Second, we can study when inflation happens. The main question is whether models inflate most where the source hedges but the model's own prior belief, learned in pretraining, is strong. That pattern would be a concrete, testable account of how a gap in domain knowledge turns into a hallucination.

Core Claim

Certainty inflation is a shift in how certain a claim sounds to its expert readers, so it should be measured on a scale fitted to those readers. It should not be measured on a hedge/no-hedge word list, and not on an LLM's own sense of certainty. We make three predictions that can each be checked. First, LLMs systematically read clinical hedges as more certain than clinicians do, so an LLM judge will under-detect inflation. Second, a detector built on the clinician scale will match clinician judgments of certainty illusions better than lexicon or LLM-judge baselines. The advantage should be largest on graded shifts, such as "suggests" becoming "indicates," rather than on outright hedge deletion. Third, inflation in generated summaries will be predicted by two things jointly: how weak the source's evidence is, and how strongly the model believes the claim when asked without the source.

Where This Sits

Three groups of papers bear on this question, and they play different roles.

Two papers motivate the problem. The Cochrane study (Tang et al.) had clinicians grade GPT-3.5 and ChatGPT summaries of 53 reviews and defined the certainty-illusion category. It also found a "more is less" effect: giving the model the full abstract produced more factual errors than giving it only the main-results section. That study supplies the target phenomenon and part of the test data. It offers no automatic measure, which is what we build. The orthopedic-guideline study (Prompt engineering in consistency and reliability with the evidence-based guideline for LLMs) found that LLM agreement with AAOS recommendations was worse and more erratic for limited-evidence and consensus items than for strong ones. That fits our hypothesis that weak evidence is where models go wrong. However, that study measured only whether answers agreed with guidelines, not how certainly the models phrased them.

A second group supplies our methods. Cloud models, from the fuzzy decision-making literature, describe a vague word by three numbers (Linguistic Multi-Criteria Group Decision-Making Method Combining Cloud Model and Evidence Theory). The expectation is the word's typical meaning. The entropy is how blurry its boundaries are. The hyper-entropy is how much different people disagree about that blurriness. We reuse this representation because inter-reader disagreement is exactly what a single number per hedge throws away. The survey of distributed linguistic representations (Distributed linguistic representations in decision making) notes that such tools have stayed confined to small fixed vocabularies of terms and have not been applied to free text generated by LLMs. We apply them to free text by eliciting scales for hedges that actually occur in clinical writing, rather than to a predefined five-point scale. Three-way decision partitions choices into accept, reject, and defer instead of forcing a binary call (Fuzzy Information Evolution with Three-Way Decision in Social Network Group Decision-Making). We borrow it to route summaries: pass automatically, flag automatically, or send to a clinician.

A third group constrains the design. Studies of verbalized confidence show that LLMs are overconfident when they state their own certainty. Overconfidence is Key reports an expected calibration error as high as 74.8% for GPT-3.5 on GSM8K. Anthropomimetic Uncertainty documents round-number bias and confidence that rises with the social role the model is given. It also notes that linguistic calibration varies by domain, but it tests only trivia. Both papers concern the model's expressed confidence about its own answers. We turn the question around and ask how the model reads other people's hedges, which neither paper tests. QUITE (QUITE) maps words of estimative probability, such as "likely" or "probable," to fixed numbers inside a symbolic solver. Its fixed-number mapping is the natural simpler rival to our three-number scale, and we test against it directly.

Research Program

The first component is the clinician hedge scale. We will collect several hundred certainty-bearing phrases from Cochrane conclusions and from AAOS recommendation text. These include verbs ("suggests," "shows"), modals ("may," "probably"), and evidence-grade phrases ("low-certainty evidence"). Each phrase will be embedded in a few realistic clinical sentences, because the same word can read differently in different contexts. Roughly forty clinicians will rate on a 0–100 slider how certain each sentence says the effect is. We will fit a cloud to each phrase's ratings using the standard backward cloud estimator, which recovers expectation $Ex$, entropy $En$, and hyper-entropy $He$ from samples. The same sentences will then go to several LLMs. We ask each for a 0–100 rating and also sample its ratings repeatedly, so that we can fit an LLM cloud to the same phrases.

The second component is the inflation score. An LLM-based aligner pairs each source claim with its counterpart in the summary. A span tagger then extracts the certainty phrase in each. Both steps are validated separately against human annotation, so that errors in them are not mistaken for failures of the scale. For a source phrase $s$ and summary phrase $t$, we define

$$\Delta(s,t) = \frac{Ex_t - Ex_s}{\sqrt{En_s^2 + En_t^2 + He_s^2 + He_t^2}},$$

which is the shift in typical reading, divided by the combined blur and disagreement of the two phrases. A positive $\Delta$ means inflation. The denominator encodes a specific judgment. Swapping two phrases that clinicians themselves read inconsistently is a smaller offense than the same raw shift between two phrases they agree on.

A summary's score is its largest per-claim $\Delta$. Three-way routing uses two thresholds, $\alpha < \beta$, tuned on a development split. Summaries scoring below $\alpha$ pass, those above $\beta$ are flagged, and those in between are sent to a clinician. The thresholds are set to meet a fixed budget for clinician review.

The third component is a study of when inflation happens. For each claim we record three things. The first is the source's evidence strength, taken from the Cochrane grade or the AAOS strength level. The second is the model's closed-book belief in the claim: we ask the question without the source many times and measure how often the answers agree. The third is the context condition, meaning the main-results section versus the full abstract, which reuses the Tang et al. contrast. We then regress per-claim $\Delta$ on these variables and their interaction. As a light intervention, we pass the extracted source clouds back to the generator as an explicit instruction, such as "preserve: low certainty." We then measure how much $\Delta$ falls. If inflation persists even with the instruction, the problem is not simply that the model failed to notice the hedge.

Baselines

The naive reference point assigns zero inflation to every summary, or equivalently uses ROUGE-L against the reference summary. Tang et al. already showed that ROUGE-L is only weakly related to factual accuracy, so it sets the floor.

The hedge-lexicon baseline counts hedging cues in the source and summary claim and flags a summary when a cue disappears. It sees the same aligned claim pairs as our method, so any gain we show comes from the scale, not from better alignment.

The strongest competitor is an LLM judge, GPT-4-class. It is shown the source and summary and asked whether the summary overstates certainty, and by how much on a 0–100 scale. We run it with the same aligned claims and with and without chain-of-thought reasoning. It also gets a matched prompt-tuning budget of a fixed number of prompt variants tried on the development split. It is the strongest competitor because it can use context and paraphrase that no lexicon handles. We found no reported result on Lacuna for automatic certainty-illusion detection on the Cochrane summaries, so the judge's number must be measured here. The only related figures in the file are human judgments. Tang et al. report that ChatGPT-MainResult summaries were preferred in 26% of cases against 34% for human references, and that fewer than 10% were medically harmful. Beating the judge means a statistically significant gain in rank correlation with clinician severity ratings on held-out reviews, plus higher precision among flagged summaries at the same clinician-review budget.

Two ablations isolate the key ingredient. The first replaces each cloud with its expectation alone, which is a QUITE-style fixed mapping of each phrase to a number. The second replaces the clinician-fitted clouds with LLM-fitted clouds while keeping the pipeline identical.

Why not just…?

Why not just ask a strong LLM whether the summary overstates the evidence? This is the real threat, and our proposal does not assume it fails. Instead, the first experiment tests the premise it rests on. If LLM-fitted clouds for clinical hedges sit systematically above clinician clouds, the judge is measuring inflation against an already inflated ruler. The "LLM-fitted clouds" ablation then shows how much detection accuracy that bias costs. If LLM and clinician readings agree and the judge matches our detector, we will report that. The contribution then narrows to a validated human scale and the causal study, both of which are still needed to trust the judge.

Why not just map each hedge word to a single probability, as QUITE does? This alternative would work if the only thing that mattered were the typical reading of each phrase. The "expectation only" ablation tests exactly that. The clouds earn their extra parameters only if normalizing by blur and disagreement improves agreement with clinician severity ratings, especially on graded substitutions between phrases. If it does not, the simpler scalar scale is the result, and we will say so.

Why not just prompt the generator to preserve hedges, or fine-tune it on domain text? The AAOS study found that even the best prompting setup reached only 62.9% guideline consistency. Its reliability across repeated queries ranged from a Fleiss kappa of −0.002 to 0.984. Fine-tuning GPT-3.5 did not consistently help. On NLI4CT, domain-pretrained models topped out near 0.644 F1, while a model given the statements alone reached 0.626, which suggests shortcut learning rather than clinical reasoning (NLI4CT). More basically, any fix to the generator needs a measure to verify it. Our intervention experiment tests one such prompt fix using that measure.

Proposed Experiments and Decisive Checks

The data are the 53 Cochrane reviews from Tang et al. in both context conditions. We use their clinician error labels where they are available. Otherwise we re-annotate summaries for certainty illusions and assign a 1–5 severity rating, using two clinicians per summary and reporting their agreement. We also use the 34 AAOS osteoarthritis items, eliciting explanations for each recommendation and stratifying by strength level. Generators are GPT-3.5, GPT-4, and one open model.

The first decisive check compares reading scales. On the elicited phrases, we measure the gap between LLM and clinician expectation and test whether the LLM expectation is higher across phrases, using a paired test over phrases. A null result here weakens the case against LLM judges.

The second check is detection. We measure Spearman correlation with clinician severity and flag precision at a matched review budget, comparing our method against the lexicon baseline, the LLM judge, and both ablations. The analysis is also split into deletion cases and graded-shift cases.

The third check is mechanism. The interaction between weak evidence and strong closed-book belief should predict $\Delta$ beyond either variable alone. The full-abstract condition should also raise $\Delta$ relative to main results only, which would restate the "more is less" finding in measurable terms. If inflation turns out to be uniform across evidence strength and model prior, the account linking knowledge gaps to hallucination is wrong for this error type, and we will say so.