Who Bears the Hallucination Risk? Community-Calibrated Factuality Guarantees for Indigenous Contexts

Language models make up facts. Recent work handles this with a statistical test: score each answer by how consistent the model is with itself, then accept the answer only if the score clears a threshold tuned on labeled examples, so that at most a fraction $\alpha$ of wrong answers get through. These guarantees are marginal, meaning they hold on average over whatever mix of questions the calibration set contained. If that mix is dominated by English Wikipedia trivia, the average can look excellent while questions about Māori history or South Sámi life pass wrong answers far more often. The guarantee itself is then unfair: the people most exposed to hallucination are exactly the communities that prior qualitative work found models misrepresent most.

We propose making the factuality guarantee hold separately for each community. The threshold for a group is calibrated on a small, labeled question set that the community itself writes, keeps, and controls. Only the resulting threshold leaves the community. The model can be a closed commercial system, and no synthetic text is fed back into fragile language archives. The scientific question is quantitative and falsifiable. We ask how many community-labeled examples a per-group guarantee needs, how much useful answering survives at that sample size, and whether the simpler pooled guarantee already protects these groups. If it does, the proposal is unnecessary, and we will report that.

Core Claim

The proposal rests on a statistical claim and a design claim.

The statistical claim is that pooled factuality calibration, as in FactTest, can satisfy its Type I error bound overall while violating it on Indigenous-topic subgroups. FactTest frames hallucination detection as a hypothesis test whose null hypothesis is "the answer is wrong," and it guarantees that the rate of wrongly approved answers stays below $\alpha$. The violation we expect has a simple cause: models are both more wrong and more confidently wrong on these topics. Effectiveness of Debiasing Techniques: An Indigenous Qualitative Analysis documents GPT-2 producing fluent, false legal and anthropological claims together with criminal stereotypes in Māori contexts. The Ethical Question – Use of Indigenous Corpora for Large Language Models documents ChatGPT inverting the Norwegian government's role in South Sámi history.

The design claim is that a per-group guarantee is achievable with calibration sets small enough for a community to label by hand. Two conditions make this work. First, the scoring function is learned on public high-resource data. Second, only the final threshold is fitted on community data, so the scarce labels are spent on the one quantity that must be group-specific.

The finite-sample cost of a per-group guarantee is explicit. Let $s_1,\dots,s_n$ be the scores of $n$ incorrect answers in group $g$'s calibration set. Let the threshold $\tau_g$ be the $k$-th largest of them, and accept an answer only if its score exceeds $\tau_g$. Write $\mathrm{TypeI}_g(\tau_g)$ for the true rate at which group-$g$ wrong answers are accepted. By the standard order-statistic argument underlying Neyman-Pearson calibration,

$$\Pr\big[\mathrm{TypeI}_g(\tau_g) > \alpha\big] \le \sum_{j=0}^{k-1} \binom{n}{j}\alpha^j(1-\alpha)^{n-j}.$$

With $k=1$ the right-hand side is $(1-\alpha)^n$. A guarantee at confidence $1-\delta$ therefore needs at least $n \ge \ln\delta / \ln(1-\alpha)$ wrong answers, which is 29 at $\alpha=0.1$ and $\delta=0.05$. Two features of this bound matter for Indigenous settings. The count is of wrong answers, so a high hallucination rate on a group reduces the total labeling burden for that group. And the bound says nothing about how many correct answers survive. That second quantity, the abstention rate, is where a fairness cost can hide: a group whose every question is refused is protected in name only.

Why Statistical Rather Than Formal Guarantees

The verification literature constrains our design more than it supplies a method. FAME computes provably minimal explanations for image classifiers by bounding network outputs over continuous pixel perturbations. VELM shields reinforcement-learning agents by computing reachable sets of continuous physical states, cutting unsafe steps by 99.7% relative to SPICE. Both require weight access and continuous input spaces. Neither exists for discrete autoregressive text, and neither scales to the closed commercial models that communities actually encounter.

We therefore treat the model as fully opaque and place the guarantee on its observable answers, as FactTest and TRAC do. TRAC composes conformal bounds over a retriever and a generator on Natural Questions. We do not reuse TRAC's retrieval layer, because a public retrieval corpus is exactly what sovereignty constraints rule out.

Fairness work supplies the closest analogue for certified guarantees. MT-NLP certifies that classifier outputs change by at most $\epsilon$ under swaps of demographic words. Our notion is different. We certify that the error rate of an abstention guarantee is bounded within each group, and we do not demand demographic invariance. Invariance is the very assumption Yogarajan et al. argue erases Indigenous context.

Research Program

The program has three parts.

The first part is building group-labeled evaluation sets with communities. We will work with partner groups under a protocol modeled on Mager et al., whose survey found 85% of activists want community members inside the research process, and on the Esethu Framework's community-controlled licensing. With these partners we will collect short factual questions in two regimes.

The first regime is questions about Indigenous communities posed in a majority language. Our targets are New Zealand/Māori topics, using the local vocabulary Yogarajan et al. identified, and South Sámi topics in Norwegian or English. In this regime the existing English meaning-clustering scores still work, and the main obstacle is the scarcity of calibration labels.

The second regime is generation in the Indigenous language itself, taken from the AmericasNLP Spanish-to-Indigenous translation directions (Findings of AmericasNLP 2023). There, "wrong" means a native-speaker rating at the low end of the shared task's 1-to-5 human scale.

Community members label correctness. They also mark restricted topics under Traditional Knowledge labels (Access Control Framework for Language Collections). Restricted topics are assigned $\alpha=0$, which means the model always abstains on them. This is a governance rule, not a learned behavior.

The second part is scores that transfer across groups. For regime one we use FactTest's semantic entropy, computed on high-resource data. For regime two the English inference models that semantic entropy relies on do not exist. We therefore test two substitutes. One is agreement among sampled outputs measured with character n-gram overlap (chrF), which tolerates rich morphology. The other is distance to a community-held example store, motivated by the finding in ChatGPT is not a good indigenous translator that M2M100 with nearest-neighbor retrieval over curated data beat GPT-4 on 10 of 11 directions. Any learned combination of these scores is fitted only on public data, so the community's labels remain untouched for the final threshold.

The third part is a calibration protocol that never moves the data. Each community computes $\tau_g$ locally on its own labels and exports only that number together with its realized $n$ and $k$. We will also measure how much a single exported threshold reveals about the underlying set, which is minimal for a single order statistic.

Baselines

The naive reference point is no abstention: the model answers everything. On each group this gives the raw hallucination rate and the zero-abstention end of every trade-off curve.

The strongest existing method is pooled FactTest. We run its semantic-entropy variant with one threshold calibrated on the union of high-resource calibration data (FEVER, WiCE, ParaRel, HotpotQA) and all community data. It is the strongest competitor because it already carries a finite-sample, distribution-free Type I guarantee at the population level, without fine-tuning the model. FactTest reports Type I error below $\alpha=0.05$ and an accuracy improvement of over 40% on WiCE and FEVER under abstention. We first reproduce these English numbers with the same model, the same 10 to 15 samples per query, and the same scoring code, so any failure on Indigenous groups cannot be blamed on a weaker reimplementation. There is no reported result on Lacuna for FactTest, TRAC, or any factuality-abstention method on Indigenous-topic or Indigenous-language questions.

We also include model self-assessment: GPT-4o is prompted to answer or abstain, or asked to judge another answer's correctness. This uses the same query budget, counted as total model calls. It has no guarantee, and BiasGuard reports zero-shot GPT-4o at only 54.25% on implicit toxicity, which suggests its judgments degrade in culturally loaded settings.

Two ablations isolate the key ingredients. The first is per-group thresholding with a score that has not been transferred, fitted from scratch on community data. The second is per-group thresholding with the transferred score, which is our full method. The gap between pooled and per-group thresholds isolates the conditioning. The gap between untransferred and transferred scores isolates the borrowing of strength.

All thresholding variants receive identical model samples per query and identical calibration and test splits. They are compared over 200 repeated random splits. To beat the strongest baseline, our method must keep realized per-group Type I error at or below $\alpha$ in at least a $1-\delta$ fraction of splits on every group where pooled FactTest does not. It must also show a per-group abstention rate no worse than the untransferred per-group variant.

Why not just…?

Why not just use pooled FactTest, since its guarantee already holds? It holds on average, and that is the problem. Our first experiment directly measures per-group Type I error under pooled calibration. This is a real threat to the proposal. If the Indigenous groups' wrong answers happen to score low under semantic entropy, the pooled threshold may already be conservative for them. If that occurs across models and values of $\alpha$, we will report that pooled calibration suffices and redirect the effort to reducing abstention.

Why not just split the calibration data by group and run FactTest separately on each? This is the honest null for the conditioning step, and it is valid. The bound above shows it becomes useless, through total abstention, once a group has fewer than about 29 wrong answers at $\alpha=0.1$. It also wastes scarce labels on fitting the score itself. Our ablation tests exactly this. If the untransferred per-group variant matches our method's abstention at realistic $n$, then the transfer contributes nothing and the contribution reduces to an evaluation study.

Why not just ask a strong model to abstain when unsure, or to check itself with Chain-of-Verification? Chain-of-Verification has the model draft an answer, generate verification questions, answer them, and revise. Self-checks of this kind carry no error bound. The documented Sámi and Māori failures were also confident and fluent, which is the case self-assessment handles worst. We include it as a baseline, and if it achieves comparable per-group error with fewer samples, that is a useful practical finding.

Why not just retrieve from a community knowledge store and refuse when nothing matches? That approach requires the community to expose a dense retrievable corpus, which many will not, and which TK labels may forbid. It also provides no guarantee. Retrieval distance remains one of our candidate scores, so if it is the most informative signal, the calibrated method will absorb it.

Proposed Experiments

The first experiment tests whether the pooled guarantee is unfair. For two open models and one commercial API, at $\alpha \in \{0.05, 0.1, 0.2\}$, we report realized Type I error per group under pooled FactTest, with confidence intervals over splits. The decisive outcome is whether any Indigenous group exceeds $\alpha$ while the pooled rate does not.

The second experiment maps the cost of fairness. We vary the number of community-labeled questions from 30 to 300. For each size we plot per-group Type I error against per-group abstention rate for the per-group variants. We then check whether the transferred score lets a group of a few hundred labeled questions answer a meaningful fraction of its correct questions while keeping the guarantee. This curve is the main deliverable: it tells a community how much labeling buys how much safe service.

The third experiment tests the guarantee when the model writes in the Indigenous language. On the AmericasNLP directions with native-speaker ratings, we compare chrF self-agreement, retrieval distance, and semantic entropy computed via back-translation as scores. We measure each score's area under the curve for separating wrong from acceptable outputs, along with the resulting abstention under per-group calibration. If no score separates the two better than chance, we will say that guarantees in this regime are presently only achievable by near-total abstention, and we will report that as the finding.

Finally, as a governance check, we audit that restricted-topic questions are refused in 100% of cases and that the exported thresholds are the only artifact leaving each community.