Ádám EszterOn June 30th, 2026 the ICON-S Annual Conference, hosted by the University College Dublin, explored the core theme, "Reimagining Public Law for a Fractured World: Technology, Identity & Truth."
Fitting precisely within this focus, the panel – "From Stenographers to Algorithms: Mapping the Implementation and Governance Gaps of AI-Based Automated-Speech Recognition Systems in Parliaments" – addressed how automated speech recognition (ASR) tools are reshaping legislative workflows amid a global decline in human stenographers.
Chaired by Nora Chronowski (ELTE CSS Institute for Legal Studies, Hungary), the session featured insights from Kevin Wayne Settles (Institute for Parliamentary Research, Germany), Boldizsár Szentgáli-Tóth (ELTE Centre for Social Sciences, Hungary), Klára Szalay (ELTE Centre for Social Sciences, Hungary), and Rudolf Berkes (ELTE Doctoral School of Sociology, Hungary). Together, these participants balanced the efficiency of parliamentary AI against critical risks to procedural integrity, data sovereignty, accuracy, and institutional trust.
Kevin Wayne Settles frames this digital transformation as an internal overhaul of parliamentary administrative routines. He argues that the core normative stake of this technological pivot is procedural integrity. In his view, democratic accountability, institutional transparency, and the traceability of evidence depend heavily on how algorithmic systems define relevance, project authority, and implicitly bound the perceived space of legal options available to drafters. He observes that this computational reconfiguration occurs across a multi-layered process that spans institutional actors, administrative relationships, and the automated translation mechanisms that turn historical archives into actionable modern legal claims. Over time, Settles warns, these AI-mediated workflows risk swift normalization, becoming an unquestioned, invisible routine within everyday parliamentary administration that hides algorithmic design choices from political oversight.
To support his thesis, Settles reconstructs the design choices and intended uses of two pioneering international systems. In the United Kingdom, the Parlex system operates as an advanced internal conversational retrieval tool trained on a massive historical corpus encompassing all digitally available parliamentary records from the year 1267 onward, alongside historical bill drafts and summaries. Settles highlights that it functions as a mega-brain utilized almost exclusively by civil servants in Whitehall to prepare the structural paperwork that underpins legislative drafting, while Members of Parliament rarely interact with the system directly. Conversely, he points to the Austrian Parliament, which developed a localized, speech-to-text automated system engineered specifically for its internal stenographic department. This specialized tool automatically transcribes live plenary speeches and identifies specific speakers, feeding the raw data into an editorial pipeline where human editors verify and correct the output before it is published as a permanent, legally binding official record.
Through these divergent structural approaches, Settles identifies critical institutional tensions. The first major tension lies in the delegation of relevance judgments to algorithmic retrieval and ranking mechanisms. When an administrative staffer submits a query to a system like Parlex, the ranked algorithm implicitly decides what information is shown first and what is effectively ignored, functioning before a human sees a single word. Settles notes that accompanying system disclaimers stating that the text is merely a "starting point" do not solve this problem; they merely shift legal liability without mitigating the underlying automation bias. Second, he argues that interfaces construct false hierarchies of authority through automated synthesis and source signalling, wherein small, invisible algorithmic parameters determine what information is marked as accurate. Third, Settles demonstrates that the capacity to contest these records is strictly bounded by interface design. Current speech-to-text systems struggle with semantic and contextual nuance; in the Austrian example, the AI routinely misinterprets acoustic data, tagging the collective sound of applause as "laughter" or an "interjection," and it remains fundamentally incapable of detecting sarcasm. Because these incorrect emotional tags are integrated directly into the preliminary draft, they risk becoming fossilized within the permanent, unexamined official transcript.
To protect the integrity of these essential public institutions, Settles states that legislatures must abandon the flawed "human-in-the-loop" doctrine, which often reduces human oversight to a single box to tick at the end of a process where a person simply verifies whether an AI agrees with its own output. True oversight requires a robust safeguards framework with checkpoints integrated directly at the initial data entry points. Settles proposes a clear safeguards checklist: parliaments must adopt clear disclosure protocols where all AI-generated content is visibly marked, implement contestability-by-design features that allow staff to easily challenge and correct algorithmic outputs, enforce role-sensitive access governance, and mandate comprehensive technical logging to maintain public accountability.
Boldizsár Szentgáli-Tóth shifts the discussion toward the technical typology and global distribution of these Automated Speech Recognition (ASR) systems. He explains that these tools serve as the primary interface between human legislators and machine infrastructure, and notes that their rapid adoption is driven by an acute institutional vulnerability: the structural decline and suspension of human shorthand and stenographer training in nations such as Estonia and Japan. To maintain the continuous generation of official records, parliaments are forced to turn to automation. However, Szentgáli-Tóth cautions that not all systems are suited for legislative environments. While speaker-dependent systems can be trained on a narrow, closed group of politicians' voices and faces to yield high accuracy, they lack generalizability. Parliaments require continuous speech recognition capable of processing flowing, natural debate without artificial pauses, alongside advanced transcription AI that can handle complex conversational shifts rather than rigid, grammar-based rules.
On a macro legal level, Szentgáli-Tóth finds that global implementation remains highly sporadic and strategically uncoordinated. He points out that the European Union's AI Act does not explicitly classify parliamentary AI applications as high-risk, leaving a wide margin of movement for developer teams but stripping democratic institutions of uniform, legally binding safeguards. Soft law guidelines from the Inter-Parliamentary Union champion abstract ideas of accountability and democratic integrity but fail to provide explicit operational instructions. Consequently, global use cases vary wildly, ranging from the simple construction of massive historical databases for future algorithm training in countries like Norway, Lithuania, Latvia, Greece, Argentina, South Africa, and Canada, to incremental automation models under human supervision in Estonia and Japan.
More advanced "hybrid archival optimization" models, such as Italy’s Gestore Emendamenti, leverage AI to actively optimize parliamentary procedures. Szentgáli-Tóth explains that the Italian system is designed to identify and group thousands of nearly identical, structurally repetitive amendment bills submitted by opposition parties as a filibustering tactic, allowing the parliament to dispose of them efficiently in a single, joint procedure. It also features semantic search capabilities that allow users to scan forty years of archives for thematic concepts, recognizing relevant discussions even when the exact phrasing and legal contexts differ completely. In Portugal, the STAr system expanded automated archiving to include localized committee meetings, eventually demonstrating an ability to auto correct its own transcription errors, which prompted the parliament to drastically reduce human supervisory loops. Furthermore, Szentgáli-Tóth details how "multilingual equalization" models are breaking down severe communication barriers. Spain uses automation to bridge regional languages like Basque and Spanish, Bahrain leverages it to adapt diverse regional Arabic dialects for public broadcast, and India uses it to navigate an environment with over two hundred officially recognized languages, making debates accessible to all citizens. In Estonia, this drive for inclusivity extends to generating automated sign language adaptations from live plenary feeds.
Despite these clear achievements in accessibility, Rudolf Berkes argues that the systemic integration of these tools introduces severe technical, security, and constitutional vulnerabilities. Technically, Berkes finds that high word error rates under real-world parliamentary acoustic conditions create a systemic linguistic bias that marginalizes minority languages and regional dialects due to a lack of clean, localized training data. From a security perspective, he warns that these systems introduce unprecedented attack vectors. Malicious actors can compromise legislative integrity by deploying unhearable or adversarial acoustic frequencies via mobile devices during live sessions, exploiting algorithmic weaknesses to corrupt internal databases or open unauthorized developer backdoors. This risk becomes critical during closed, confidential parliamentary committee sessions, where an undetected software backdoor could expose sensitive state secrets to foreign cyber-adversaries.
Constitutionally, Berkes emphasizes that these technical and security failures threaten the authenticity of official state records. As parliaments embrace low-cost, convenient AI alternatives and human staff is displaced, legislatures face a dangerous over-reliance on a black-box system where enforcing legal liability for data omissions is structurally impossible. This vulnerability is deeply tied to the erosion of institutional data sovereignty. Berkes notes that the widespread use of convenient third-party software means that sensitive legislative data is routinely transmitted across continents to be processed on cloud servers owned by foreign multinational corporations (such as Amazon infrastructure in America). This compromises the state's sovereign control over its own data, a reality that Berkes connects to a critical, ongoing push within the European Union to encourage and mandate localized server parks located physically within state territory.
Ultimately, Klára Szalay outlines how these technological shifts depend heavily on institutional decision-making and compliance frameworks. She asserts that parliaments must transform from passive consumers of market software into active developers capable of engineering localized solutions that comply fully with risk management laws, national strategies, and strict data protection regulations. Szalay observes that legislatures face critical structural choices in an audiovisual age; a parliament might legitimately decide that traditional written record-keeping is obsolete, choosing instead to upload raw, verified video streams as the official public record. However, if a parliament maintains a written system, it must confront the reality that AI does not currently save operational time. While automation successfully spares staff from repetitive physical typing, constant errors in numbers, dates, and names require intensive human editing. Szalay clarifies that human shorthand writers do not disappear; their roles must actively shift into validators and editors who operate within strict time windows (such as a ninety-minute post-shift window) to verify algorithmic drafts against live audio.
The panel proceedings conclude with historical and social framing from Nora Chronowski, who integrates these findings into a broader civic warning. Reflecting on contemporary philosophical perspectives and the risks raised by Berkes, Chronowski emphasizes that society must approach AI with active critical thought and lifelong learning rather than blind compliance. To protect a resilient democracy and counter the consolidation of power by a small group of wealthy technology companies, she concludes that public institutions and educated citizens must assert rigorous legal, physical, and sovereign control over these platforms, ensuring that the critical data of governance remains transparent, secure, and entirely within public hands.
_________________________________________________________
This report was prepared with the support of the ELTE CSS Algorithmic Constitutionalism Lendület/Momentum Research Group (LP2024-20/2024), funded by the Hungarian Academy of Sciences.
_________________________________________________________
The views expressed above belong to the speakers and do not necessarily represent the views of the Centre for Social Sciences.