Schools, universities and peer-review venues now face a specific question: did a writer use AI beyond what the rules allow, for example generating paragraphs when only grammar fixing was permitted? Current detectors answer badly in two ways. Their scores barely move between 1% AI polishing and 75% machine generation, and their false alarms fall unevenly across groups of writers. One audit found a commercial detector flagged 2.1% of chemistry abstracts but 26.6% of theology abstracts (Why AI Detection Fails for Academic Integrity). Non-native writers are flagged disproportionately as well (Watermark in the Classroom). The usual repair is calibration: pick each population's flagging threshold from that population's own permitted writing. That repair requires local data. Each department or school holds only a few hundred such essays and cannot legally share them.
The obvious way to combine federated learning (training a shared model across institutions that exchange parameter updates instead of raw data) with AI-text detection is to federate the training of the detector itself. We argue that this targets the wrong part of the system. The harm in deployment is an uncontrolled false-accusation rate, and that rate is set by the threshold, not by the detector's weights. We propose to federate the calibration step instead. Institutions keep a detector trained centrally on public data. Each one shares only a privacy-noised summary of how that detector scores its permitted writing. A server combines these summaries so that every institution, including small ones, receives a threshold with a provable bound on its own false-positive rate. This costs one round of communication and a few hundred numbers per client, compared with thousands of gradient rounds.
Core Claim
Our claim has two parts. First, for false-positive control on permitted assistance, federated calibration of a fixed detector matches pooled centralized calibration, and it beats both a single global threshold and purely local calibration. It also achieves something no current method does: a bound on each institution's own false-positive rate. Second, once calibration is done properly, federated fine-tuning of the detector adds little or nothing on the safety side, measured as per-client false-positive rate. Its value, if any, appears only in detection power on clients whose writing is far from the public training data.
If the second part fails, so that federated fine-tuning substantially improves power at a matched false-positive rate, that is itself a useful result. It would tell practitioners when the much more expensive protocol is worth running.
Prior Work and Its Role
Three papers motivate the gap by documenting where false positives fall. Almost AI, Almost Human evaluated 12 detectors and showed they flag extremely minor Llama 2-7B edits at high rates; GPTZero flagged 64.71%. Why AI Detection Fails for Academic Integrity traced the disciplinary skew to surface features such as academic vocabulary density. Watermark in the Classroom showed that a fixed watermark threshold at $\alpha=0.05$ produced false-positive rates above 25% on essays with permitted grammar help.
That last paper also supplies our method. It applies conformal prediction, a way of setting a threshold from held-out scores so that the false-positive rate is guaranteed at a chosen level, in standard, hierarchical and weighted variants. It does this on native (LOCNESS) and non-native (TOEFL11) essays. We differ from it in three ways. It pools all calibration essays on one server. It is tied to watermark p-values. And it does not address clients too small to calibrate alone. We keep its weighted and hierarchical calibration ideas but run them over privacy-protected summaries. Our method also works with any detector score, so it applies to open-source models that carry no watermark.
The detectors we calibrate come from the hybrid-text line of work:
- EditLens regresses how much AI editing a text received. Its continuous score is the natural quantity on which to set a "permitted versus not" threshold.
- DivEye extracts nine features of how surprise rises and falls across a text, read off a frozen model, and classifies them with XGBoost.
- SeqXGPT labels individual sentences in mixed documents.
We use these detectors as they are; we do not propose a new one.
FedNLP constrains the design. It showed federated Transformer fine-tuning on 20Newsgroup falling from about 86% (centralized) to about 55% under skewed client data. That warns that federating the detector itself may hurt it. FedEx-LoRA supplies the stronger version of the obvious alternative: exact aggregation of low-rank adapter updates, small add-on weight matrices fine-tuned in place of the full model. FedHealth showed that personalizing a shared model to each client raised accuracy by 5.3 points on sensor data. Our calibration step is personalization in its cheapest form: only the threshold is personalized.
Research Program
Let $s(x)$ be a detector's score for text $x$. Each client $k$ holds $n_k$ permitted texts with scores $S_k=\{s(x_i)\}$. Permitted texts are human-written, or human-written with the allowed assistance such as grammar fixing. The goal is a client threshold $\tau_k$ satisfying
$$\Pr_{x\sim P_k}\big[s(x)>\tau_k\big]\le\alpha,$$
where $P_k$ is client $k$'s distribution of permitted text and $\alpha$ is the target false-positive rate. Purely local split conformal calibration meets this bound exactly. However, it needs $n_k\ge 1/\alpha-1$ samples just to give a finite threshold, which is 99 essays at $\alpha=0.01$. Below a few hundred essays its realized false-positive rate also fluctuates widely from one calibration draw to the next.
Under our protocol, each client releases a differentially private histogram of $S_k$ over fixed score bins. Differential privacy here means adding calibrated noise so that the released counts reveal almost nothing about any single essay. The server clusters these histograms and returns to each client a pooled reference built from weights $w_{kj}$ over the other clients $j$. Each client then sets $\tau_k$ by weighted conformal calibration, with its own scores weighted most heavily.
Standard weighted-conformal reasoning under distribution shift gives the resulting false-positive rate a bound of the form
$$\mathrm{FPR}_k\le\alpha+\mathrm{TV}\Big(P_k,\textstyle\sum_j w_{kj}P_j\Big)+\epsilon_{\text{DP}}+O\big(1/n_{\text{eff}}\big).$$
Here TV is total-variation distance, the largest difference in probability the two distributions assign to any event. $\epsilon_{\text{DP}}$ is the error introduced by the privacy noise, and $n_{\text{eff}}$ is the effective calibration size after weighting. The weights are chosen to trade the distance term against the sample-size term. Each client estimates that distance term locally, on its own held-out scores, without revealing them.
A second component addresses a different problem. Unpermitted generation is caught at the rate the detector's power allows. As a stretch goal, if histograms show a client's score distribution has collapsed toward the threshold, we add one federated step: a federated XGBoost update of the DivEye feature head, which is small, rather than any adapter on the backbone model.
Baselines
The naive reference is a single global threshold. We set it at the $\alpha$-quantile of permitted-text scores pooled from public data, which mirrors how commercial tools are deployed. The relevant reported figures are the more than 25% false-positive rate on permitted grammar help at $\alpha=0.05$ and the 4% rate on unassisted human essays (Watermark in the Classroom). Beating this baseline means bringing every client's false-positive rate on permitted text to within $\alpha$ plus 1 point.
The privacy-violating upper bound is pooled centralized hierarchical conformal calibration in the style of Watermark in the Classroom, run on the same client data with all calibration essays placed on one server. Our method succeeds if its worst-client false-positive rate and its detection power fall within noise of this oracle. No reported result found on Lacuna for a federated or decentralized version of this calibration.
Local-only conformal calibration is the strongest private competitor, since it already has an exact guarantee. We run it with each client's own data and no communication. We beat it if our method reduces the spread of realized false-positive rates across calibration draws and raises power for clients with $n_k<300$, while keeping the same worst-case false-positive rate.
The strongest existing modeling method is federated low-rank adapter fine-tuning of the detector. We use FedEx-LoRA aggregation (FedEx-LoRA) on an EditLens-style regressor and follow it with our calibration. The fairness constraints are:
- The same client data serves as training data for the federated method and as calibration data for ours.
- The base detector is the same.
- Hyperparameter tuning gets an equal budget of 20 configurations.
- Communication cost is reported for both.
For reference, the centralized EditLens regressor reports 94.7% F1 for human versus any AI and 90.4% ternary F1 (EditLens). No reported result found on Lacuna for any federated training of a hybrid-text detector.
The key ablation isolates the weighting step. We remove the server's cluster-based weights and pool all clients uniformly, keeping the privacy noise. If uniform pooling performs as well, the similarity weighting is unnecessary. We expect uniform pooling to fail precisely on the theology-like and non-native clients.
Why not just…?
Why not just train one detector on large public and synthetic data and use one threshold? This is the status quo, and the evidence against it is direct. A single threshold produced 2.1% versus 26.6% flag rates across disciplines (Why AI Detection Fails for Academic Integrity). It also produced above 25% false positives on permitted grammar help (Watermark in the Classroom). Our method keeps the centrally trained detector, so it keeps whatever that detector does well, and changes only the threshold. The global-threshold baseline tests this directly.
Why not just calibrate locally at each institution? This is a real threat, because local split conformal calibration is already exactly valid. Its weaknesses appear at small sizes. With tens to low hundreds of essays, the realized false-positive rate varies widely across draws, and the threshold becomes conservative enough to lose power. If our results show local calibration matching ours at $n_k\ge 200$, we will report that as the size at which federation stops mattering. The protocol would then be justified only for smaller units such as individual courses or small venues.
Why not just federate the detector's training, the obvious combination? This is the idea that would naturally be funded first. FedNLP's drop from about 86% to 55% under skewed clients suggests it is costly and fragile (FedNLP). More importantly, a better detector does not by itself guarantee any client's false-positive rate. It still needs a per-client threshold, which our method supplies at far lower cost. The federated-adapter baseline settles the question empirically. If it gains more than 5 points of power at matched false-positive rates on shifted clients, we will say so and recommend combining the two.
Why not just rely on watermarks? Watermarks survive only partially under mixed editing (Optimal Watermark Localization in Mixed-Source Large Language Model Texts), and open-source models running locally carry none. Even watermark scores need calibration to avoid flagging permitted help (Watermark in the Classroom). Our method calibrates any score, including watermark p-values, so the two are complementary.
Proposed Experiments
We simulate institutions from corpora whose natural groupings match real deployment:
- TOEFL11 split by the writer's native language (11 clients), plus LOCNESS as a native-English client (Watermark in the Classroom).
- The OpenAlex abstract corpus split by discipline (Why AI Detection Fails for Academic Integrity), where chemistry and theology serve as the known extreme clients.
- MIXSET operations and APT-Eval polishing levels, used to generate permitted variants (extremely minor to minor polishing) and unpermitted variants (major edits, generation) of each client's human text (LLM-as-a-Coauthor; Almost AI, Almost Human).
Client sizes are subsampled to $n_k\in\{30,100,300,1000\}$. Calibration uses Llama, GPT-family, Claude and DeepSeek editors. We then test on editors held out from calibration, which probes whether the guarantee survives a change of generator.
Three detectors supply scores: an EditLens-style regressor trained on the public EditLens data, DivEye features with XGBoost, and SeqXGPT sentence scores aggregated to the document. We report the following metrics:
- Worst-client false-positive rate on permitted text.
- The across-draw spread of each client's false-positive rate.
- Detection rate on unpermitted text at $\alpha\in\{0.01,0.05\}$.
- Privacy budget $\varepsilon\in\{1,4,\infty\}$.
- Bytes communicated.
Three checks decide the project:
- Under the global threshold, the theology and low-proficiency TOEFL clients must exceed $\alpha$ by a wide margin, while our method brings them within $\alpha$ plus 1 point at $\varepsilon=4$ and $n_k=100$.
- Our method must match the pooled oracle's power to within 2 points.
- The federated-adapter comparison must show whether detector federation buys any additional power once calibration is fixed.
A negative result on the second check at small $n_k$ would mean privacy noise dominates. In that case we would publish the minimal per-client size and privacy budget at which federated calibration becomes worthwhile.
References
- Watermark in the Classroom: A Conformal Framework for Adaptive AI Usage Detection
- Almost AI, Almost Human: The Challenge of Detecting AI-Polished Writing
- EditLens: Quantifying the Extent of AI Editing in Text
- Diversity Boosts AI-Generated Text Detection
- SeqXGPT: Sentence-Level AI-Generated Text Detection
- FedNLP: Benchmarking Federated Learning Methods for Natural Language Processing Tasks
- FedEx-LoRA: Exact Aggregation for Federated and Efficient Fine-Tuning of Large Language Models
- FedHealth: A Federated Transfer Learning Framework for Wearable Healthcare
- Optimal Watermark Localization in Mixed-Source Large Language Model Texts
- LLM-as-a-Coauthor: Can Mixed Human-Written and Machine-Generated Text Be Detected?