Introduction
Among parliamentary editors, artificial intelligence (AI) continues to attract widespread attention. The editorial services (dienst Verslaggeving) of the Flemish Parliament (Vlaams Parlement), which is the regional assembly of Flanders, the Dutch-speaking region of Belgium, have certainly not overlooked this development.
Despite the shared challenges faced by parliamentary reporting worldwide (Eras 2025), collaboration on the use of AI in editing official records remains limited. This is due to several structural and institutional factors, among which are language diversity. Despite these barriers, we are convinced that lessons may be learnt from different experiences worldwide.
In this article, we will therefore explain the steps we have taken towards a structural integration of AI in our workflow, while remaining attentive to its potential pitfalls.
Acknowledging and embracing AI
Since 2005, the parliamentary editors of the Flemish Parliament have been relying on a custom-built application developed in-house. It supports editors in key tasks and managing the full editorial workflow for producing the official parliamentary record.
The editorial team has monitored developments in automatic speech recognition (ASR) since 2016. Until recently, however, ASR solutions did not meet the accuracy standards required for high-quality verbatim reporting. Over the past few years, advances in artificial intelligence have significantly improved the reliability of transcription tools, including for Dutch.
At the same time, June 2024 marked the beginning of a new legislative term, initiating a strategic planning process for the 2024-29 period. Recognising that a parliamentary institution cannot afford to fall behind rapid technological and societal change, the Flemish Parliament explicitly included in its strategic agenda the exploration of AI, as a promising and potentially transformative technology. An organisation-wide innovation platform, ‘DigiLab’, was launched to collect and prioritise potential AI use cases. One of the most prominent use cases identified was to investigate how AI could meaningfully support the preparation of plenary session reports.
Our primary objective is to integrate automated transcription into the existing editorial application. The envisioned outcome is a verbatim transcript that is immediately available to editors for review and correction, thereby eliminating the physically demanding and time-intensive task of typing speeches from scratch.
Exploring the opportunities with a proof of concept
To pursue this objective, we established an internal AI task force consisting of editors(-in-chief) who voluntarily explored speech-to-text tools, such as Speechmatics and Scribewave, and started testing prompts in Copilot.
To assess both the potential and the limitations of this use case, the IT department developed a proof of concept (POC). The POC followed a multi-stage pipeline combining speech recognition, speaker attribution, and AI-assisted editorial pre-processing. At the outset, this POC captures the livestream of plenary sessions, which are publicly broadcast via YouTube. Using a trained dataset in Pyannote ASR technology, the spoken content is attributed to the presumed speaker and transcribed with OpenAI’s Whisper. The resulting verbatim transcript is then post-processed by ChatGPT. Editors locate their assigned speech fragments, copy and paste the edited text into Word, and subsequently verify it against the original audio recording to correct any remaining errors or ambiguities.
Early testing demonstrated considerable potential but also revealed limitations, including inaccuracies in speech recognition and inconsistent editorial decisions by ChatGPT. Through iterative prompt design, the AI task force provided the IT team with a detailed set of editorial instructions to ensure consistent application of editorial standards. Combined with additional Pyannote training and increased GPU capacity, these measures substantially improved the quality of the output.
After four months of testing, the POC was rolled out to the entire editorial team in May 2025. Our editors responded enthusiastically to the introduction of the tool. By October 2025, nearly 80% of them reported using the system regularly as a support tool for drafting plenary session reports. All of those regular users agree that it has become a mainstay, and most of them signify that the tool helps to reduce their workload.
With the POC clearly demonstrating its added value, development of a fully integrated AI system within our editorial application began in January 2026.
A positive experience with notable pitfalls
Despite their enthusiasm for the POC, editors have shown concerns when the system underperforms or when they experience delays. The elimination of full manual transcription is widely perceived as a major physical relief. At this stage, however, it is difficult to quantify time savings. While some editors finish their work noticeably faster, others take more time to revise texts carefully or combine AI-generated content with manual reporting.
Working with AI also presents challenges. As noted by colleagues in British Columbia (Kerr 2025), this new way of working requires a shift in editorial mindset. Even experienced editors sometimes overlook obvious errors. Maintaining focus remains as critical as ever. The automation paradox applies here: as routine tasks, such as transcription, become increasingly automated, human oversight and intervention become more important and demand higher levels of concentration and judgment – not less. Moreover, human judgment remains indispensable for understanding the political and societal context of parliamentary debate (Digitaal Vlaanderen 2025).
Data protection is another major concern. As an EU-based government institution, the Flemish Parliament prefers European technologies for its digital infrastructure, ensuring compliance with EU legislation on data storage and processing. This necessitates ongoing comparative testing of different AI solutions.
We are therefore collaborating with research groups at the universities of Leuven and Ghent, which are developing transcription software tailored to the wide range of accents and regional idioms found in spoken Flemish – features that differ markedly from the Dutch spoken in the Netherlands. If successful, this solution would not only be locally developed and maintained, but better adapted to our specific linguistic context.
Conclusion
Our office marked the beginning of an intensive and transformative project in 2024 that continues to reshape our editorial workflow. Structural integration of AI obviously requires close and sustained cooperation with IT specialists.
During this process, we encountered understandable concerns about the disruptive effects of AI. Reports on the environmental impact of generative AI and its broader societal consequences raise serious ethical and technical questions. Rather than dismissing AI outright, however, we choose to engage critically with both its opportunities and its challenges.
A fully integrated AI-supported workflow will undoubtedly reduce the time spent drafting reports. More importantly, it will give editors additional breathing space to take short breaks, research debate context more thoroughly, reflect on recurring errors, and resolve grammatical or procedural issues. All of this can enhance accuracy and editorial quality.
More broadly, our experience suggests that AI will not replace parliamentary editors but will reshape the nature of their work. Editorial expertise will increasingly focus on verification, contextual interpretation, and safeguarding the reliability of the parliamentary record.
Louis Peckstadt and Katrien Van Mulders are editors-in-chief at the parliamentary reporting office of the Flemish Parliament (Vlaams Parlement).
Lieve Beullens is head of the parliamentary reporting office of the Flemish Parliament (Vlaams Parlement).
References
Digitaal Vlaanderen. (2025). AI Playbook. In www.digitaalvlaanderen.be. Agentschap Digitaal Vlaanderen. URL: https://assets.vlaanderen.be/image/upload/v1741686615/repositories-prd/AI_Playbook_kernactiviteiten_wlvopb.pdf
Eras, H.-J. (2025). Baby steps: Applying AI in parliamentary reporting. Tiro 2/2025. URL: https://tiro.intersteno.org/2025/12/baby-steps-applying-ai-in-parliamentary-reporting/
Kerr, D. (2025). Harnessing Whisper at the Legislative Assembly of British Columbia: A User-Driven Approach to AI-Supported Parliamentary Reporting. Tiro 1/2025. URL: https://tiro.intersteno.org/2025/06/harnessing-whisper-at-the-legislative-assembly-of-british-columbia-a-user-driven-approach-to-ai-supported-parliamentary-reporting/
