💬 NLP

Detecting undisclosed LLM-generated content in parliamentary texts

This paper presents an interpretable classifier trained to detect undisclosed LLM-generated content in parliamentary texts, revealing a steady increase in such usage in the United Kingdom and Sweden since 2022.

Original authors: Minerva Suvanto, Andrea McGlinchey, Peter J. Barclay, Mattias Wahde

Published 2026-06-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Minerva Suvanto, Andrea McGlinchey, Peter J. Barclay, Mattias Wahde

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a bustling library where politicians from the UK and Sweden write speeches, motions, and official statements. For years, these documents were written entirely by human hands (or at least, human minds typing on keyboards). But recently, a new, invisible ghost has started haunting the library: the Large Language Model (LLM), the same kind of AI that powers chatbots like ChatGPT.

This paper is like a team of detectives who built a special "lie detector" to find out if politicians are secretly letting this AI ghost write their speeches without telling anyone.

Here is the story of their investigation, broken down simply:

1. The Mystery: The Silent Ghost

In many jobs, like journalism or school essays, if you use AI, you are supposed to raise your hand and say, "I used a robot to help me." But in the world of parliament, the rules are fuzzy. There are no strict laws forcing politicians to say, "Hey, I asked an AI to draft this."

The researchers worried that if politicians are using AI to write their speeches but not admitting it, it could hurt public trust. It's like if a chef served you a meal but didn't tell you that a robot cooked it, and the robot accidentally put salt in the dessert.

2. The Detective's Tool: The "Glass Box"

To catch the ghost, the researchers didn't use a black, mysterious machine that gives answers you can't understand (a "black box"). Instead, they built a "Glass Box" classifier.

Think of this like a magnifying glass that doesn't just say "Guilty" or "Not Guilty." It shows you exactly which words and sentence structures made it suspicious.

How they trained it: They took old speeches from before 2022 (when AI wasn't popular yet) and knew those were 100% human. Then, they took those same speeches, fed them to an AI, and asked the AI to rewrite them. Now they had a "Human" pile and an "AI" pile.
The Training: They taught their "Glass Box" to spot the tiny differences between the human pile and the AI pile. They found that AI tends to write in slightly different patterns—like using "United Kingdom" instead of "UK," or having a very specific rhythm to its sentences.

3. The Investigation: Checking the Recent Files

Once the detective tool was ready, they turned it loose on the new speeches from 2021 to 2026. They looked at these speeches in two ways:

The Whole Speech: Did the entire document smell like AI?
The Paragraphs: Did just one paragraph in the speech smell like AI? (Maybe the politician wrote the intro, but asked the AI to write the middle part).

4. The Findings: The Ghost is Everywhere (and Growing)

The results were clear and steady:

The Trend: In both the UK and Sweden, the use of AI in these texts has been creeping up steadily since 2022.
The Numbers: By 2026, the detector found that about 2% of UK speeches and 6.5% of Swedish speeches were likely written entirely by AI.
The "Hidden" Parts: When they looked at individual paragraphs, the numbers were even higher. In the UK, over 15% of recent motions contained at least one paragraph that looked like it was written by AI. In Sweden, it was over 9%.
The Silence: The most shocking part? None of these texts admitted to using AI. Not a single one said, "We used a chatbot here."

5. Why It Matters (According to the Paper)

The paper argues that this lack of honesty is a problem.

Trust: If voters find out their politicians are using AI to write their words without saying so, it might feel like a betrayal.
Mistakes: AI can sometimes make things up (hallucinate) or be biased. If a politician uses AI without checking it carefully, they might accidentally spread false information or offensive ideas.
The "Cognitive Trojan Horse": The paper mentions that relying on AI might make politicians stop thinking critically themselves, letting the AI do the heavy lifting of thought without them realizing it.

6. The Conclusion

The researchers conclude that while using AI isn't necessarily "bad," hiding it is. They found that politicians are increasingly using AI to write their official texts, but they aren't following the (often vague) guidelines that suggest they should be transparent about it.

They suggest that parliaments need to stop being vague and start having clear, strict rules: "If you use AI, you must say so." Until then, the "ghost" in the library will keep writing speeches in the dark.

Technical Summary: Detecting Undisclosed LLM-Generated Content in Parliamentary Texts

Problem Statement
The rapid integration of Large Language Models (LLMs) into human-computer interaction has raised significant concerns regarding transparency in political discourse. While guidelines from bodies such as the Interparliamentary Union (IPU) and the UK Government recommend disclosing AI use in parliamentary texts to maintain public trust and prevent the spread of misinformation or bias, such disclosures are largely absent from official records. This lack of transparency makes it difficult to assess the extent to which politicians and their aides utilize AI for drafting motions and statements. Furthermore, existing detection methods often rely on "black-box" neural networks, which lack the interpretability required for high-stakes decisions where false accusations could damage reputations.

Methodology
The authors propose a framework to detect undisclosed LLM-generated content in parliamentary texts from the United Kingdom (Westminster) and Sweden (Riksdag). The approach relies on a fully interpretable, linear "glass-box" classifier known as the Interpretable CONfidence-enhanced Perceptron (ICON).

Data Collection and Generation:
- Human-written (Class 0): Texts were extracted from parliamentary records published between 2014 and 2020 (pre-modern LLM era) to serve as the ground truth for human writing.
- LLM-generated (Class 1): To simulate realistic usage, the authors employed a two-step prompting procedure. First, human-written texts were summarized by an LLM (Gemini 3.1 Flash-Lite). Second, these summaries were used as prompts for LLMs (Gemini and GPT-5 mini) to generate full texts. This resulted in a balanced base dataset of 1,400 human and 1,400 LLM-generated texts per language (UK and Swedish).
- Forward Test Sets: Unlabeled texts published from 2021 onwards were collected to evaluate real-world LLM usage. Texts from early 2021/2022 served as a hold-out set to measure false positive rates (FP) before widespread public LLM adoption.
Classification Model (ICON):
- The classifier utilizes n-gram features (unigrams and bigrams) extracted from tokenized text.
- It employs a linear decision function $s = \alpha + \sum w_i f_i$ , where a text is classified as LLM-generated (Class 1) if the sum $s \geq 0$ .
- Weights ( $w_i$ ) and the bias term ( $\alpha$ ) are initialized based on feature frequencies in the training classes and optimized via a modified perceptron training method that computes errors over the entire dataset rather than sample-by-sample.
- Interpretability: The model provides a confidence score based on the magnitude of the classification sum ( $|s|$ ) and allows for the visualization of specific feature contributions (e.g., specific words or bigrams) that drive the decision, avoiding the opacity of deep neural networks.
Evaluation Strategy:
- The study evaluates detection at two levels: (1) classifying the entire parliamentary motion as a single unit, and (2) classifying individual paragraphs within a motion.
- A confidence threshold ( $T$ ) is applied to the classification sum to minimize false positives, particularly for the paragraph-level analysis.

Key Results

Classifier Performance: The trained classifiers achieved high performance on held-out test sets, with F1 scores of 0.940 (UK) and 0.969 (Swedish). Crucially, the false positive rate on pre-LLM era texts was extremely low (0.062 for UK, 0.043 for Swedish), indicating robustness against misclassifying human writing.
Detection of Undisclosed Use: Applying the classifiers to the forward test sets (2021–2026) revealed a steady increase in detected LLM usage in both parliaments:
- Full Text Detection: When classifying entire motions, detection rates rose from near-zero in 2021 to approximately 2.1% in the UK (2026) and 6.5% in Sweden (2025–2026).
- Paragraph-Level Detection: When analyzing individual paragraphs, the detection rates were higher, indicating that LLMs are often used to generate specific sections of a text rather than the whole document. By 2026, approximately 15.5% of UK motions and 9.4% of Swedish motions contained at least one paragraph flagged as LLM-generated.
Trends: The data shows a sharp increase in detected usage starting in late 2022/2023, correlating with the public release of advanced LLMs.
Disclosure: A manual search of the detected texts for keywords related to AI (e.g., "ChatGPT," "LLM," "AI") confirmed that none of the texts contained any disclosure of AI usage.

Significance and Claims
The paper claims to provide the first systematic evidence of a growing trend in undisclosed LLM usage within parliamentary texts in the UK and Sweden. By utilizing an interpretable "glass-box" classifier, the authors argue that their findings are more defensible in high-stakes political contexts than those derived from black-box models, as the specific reasons for classification can be inspected and visualized.

The authors emphasize that their results suggest a violation of existing (though often non-mandatory) guidelines regarding transparency. They note that while LLMs can be useful tools, their undisclosed use risks eroding public trust, introducing subtle biases, and potentially spreading misinformation or "cognitive offloading" where human oversight is reduced. The study concludes that the current lack of disclosure is an "egregious violation" of transparency principles and calls for clearer, actionable guidelines and adherence to them in parliamentary processes. The paper modestly notes limitations, such as the potential for linguistic differences affecting cross-language comparability and the possibility of "grey area" texts where human editing of AI output may blur detection boundaries, but maintains that the detected trends are robust and statistically significant.