AI Transparency and Accuracy

ParliView — AI Transparency and Accuracy

Effective Date: 30 September 2026
Last Updated: 25 September 2026
Version: 2.0

---

At a Glance

---

1. About This Page

This page explains how ParliView uses AI, how you can tell what is AI-generated, what can go wrong, how to check what you read and how to report errors.

This page gives you information. It is not a contract. The conditions of taking part in the Study, and your rights, are set out in the Participant Terms and Information and the Privacy Notice. Nothing on this page limits them.

On this page:

• “ParliView” or “the Platform” means the website at www.parliview.org and the web application at app.parliview.org.
• “UCD”, “we”, “us” and “our” mean University College Dublin, National University of Ireland, Dublin, Belfield, Dublin 4, Ireland.
• “You” or “participant” means a person with a ParliView account. Every person who uses ParliView is a participant in the Study.
• “The Study” means UCD research study 025-HS-26-C-Cross, approved by the UCD Human Research Ethics Committee. An amendment describing the Platform as launched was submitted to the Committee on 25 September 2026 and is under review.
• “Answer” means text that ParliView generates in reply to your question.
• “Summary” means AI-generated content that ParliView prepares and shows to all participants, such as summaries of debates, procedures, events, adopted texts and topics, and suggested starter questions.
• “AI Output” means Answers and Summaries.
• “EP Material” means documents and data published by the European Parliament.
• “Parltrack Data” means data compiled by Parltrack (parltrack.org), mainly from European Parliament sources, used under the Open Database Licence.

---

2. Who Is Responsible

• UCD. UCD runs ParliView as a research platform, free of charge and under its own name. UCD develops the AI system, decides how it works and what ParliView shows you, including the Summaries, and is responsible for it under the law. That does not mean AI Output is checked before you see it: no person reviews it, and Section 6 explains how Summaries are made. UCD is responsible, as both provider and deployer, for the transparency rules in Article 50 of the EU Artificial Intelligence Act (Regulation (EU) 2024/1689). These rules require us to tell you when you are dealing with AI and to label or mark AI-generated content.
• Research does not exempt us. ParliView exists to conduct the Study, but it is used by people outside the research team, so the AI Act’s exclusion for research does not apply. The transparency rules apply in full.
• The Study. Every participant’s use of ParliView is part of the Study, including the questions you ask and the Answers you receive. Taking part is voluntary. For what we collect, how we use it, how study conditions work and how to withdraw, see the Privacy Notice and the Participant Terms and Information.
• AI model developers. ParliView uses AI models developed by other companies (Section 5). UCD decides how those models are used in ParliView and what ParliView shows you.
• Partners and funder. Our project partners, the University of Strathclyde and Transparency International EU, and our funder have no editorial control over AI Output.
• Not the European Parliament. ParliView is not run or endorsed by the European Parliament, and the Parliament does not check AI Output.

---

3. What ParliView Is For, and What It Is Not

What it is for. ParliView is designed to help people find, understand and check information about the work of the European Parliament. That covers its debates, votes, legislative procedures, adopted texts and written questions, and the roles of Members of the European Parliament (MEPs). We study how people use an AI system of this kind and how well it performs.

Who can use it. Access is by approval. You request access using the form on www.parliview.org, and we admit people who meet the conditions of the Study, including being 18 or over.

What it is not.

• Not voting advice. ParliView does not tell you how to vote, or which party, candidate or political group matches your views. If you ask for this, ParliView is designed to decline and to offer factual information instead.
• Not the official record. The European Parliament’s own publications are the authoritative source. AI Output is a guide to them, not a replacement.
• Not the European Parliament’s view. AI Output does not speak for the Parliament, its Members or its political groups.
• Not legal or professional advice. Do not rely on ParliView for legal, financial or other professional decisions.
• Not live news. ParliView’s information lags behind the Parliament’s own publications (Section 6).

---

4. How You Can Tell What Is AI-Generated

4.1 On Screen

• Before your first question. A notice on the empty chat screen tells you that ParliView’s Answers are written by AI, that AI Output can be wrong and that ParliView does not give voting advice. It is always visible before you ask your first question.
• On every Answer. Each Answer carries the label “AI-generated” at the top. The label covers everything in the Answer, including summaries of sources, notes on the research steps and suggested follow-up questions.
• On every Summary. Each Summary carries the label “AI-generated” at the start.
• Near the question box. A short reminder that Answers are AI-generated and can contain errors stays visible while you use the chat. It links to this page.
• The same in every study condition. The notice, the labels and the reminder are the same whichever study condition you are in.
• Readable labels. Labels are written in text, not shown only by colour or an icon, so that screen readers can announce them.
• What is not AI-generated. Vote tables, MEP profile cards and links to documents show data taken from EP Material and Parltrack Data. The AI does not write them, but they can still contain errors from the source data or from how we process it.

4.2 Machine-Readable Marking

The same rules also require AI-generated text to carry a machine-readable mark, so that software can detect it as AI-generated.

At launch, the AI model we use does not add a machine-readable mark to the text it writes. Marking text reliably is not yet technically settled, and we will update this page if that changes. The on-screen labels are the way to identify ParliView’s AI Output.

Code of Practice. UCD has not signed the European Commission’s Code of Practice on Transparency of AI-Generated Content.

4.3 Copying and Sharing

ParliView does not currently offer buttons to copy or download Answers. If you copy text from ParliView yourself, say clearly that it was generated by AI and use the attribution in the Participant Terms and Information, Section 4.2. If you publish AI Output to inform the public on a matter of public interest, the law may require you to say that it was generated by AI.

---

5. How Answers Are Produced

When you ask a question, ParliView carries out several steps. Each step uses an AI model, so answering one question involves several calls to the models.

1. Understanding your question. A model turns your question into search queries and plans what to look for. It takes account of the earlier questions and Answers in the same conversation.
2. Searching ParliView’s own index. ParliView searches an index that UCD builds and stores. The index contains:
   • EP Material, such as verbatim reports of plenary debates, adopted texts, legislative procedures, press releases, fact sheets, the Rules of Procedure and written questions;
   • Parltrack Data, such as MEPs’ roles, party and political group membership, and roll-call votes;
   • some ParliView Summaries; and
   • information about ParliView itself.
   ParliView does not search the live web to answer your question.
3. Checking and widening the search. A model assesses whether the sources found are relevant, and ParliView may search again more broadly.
4. Writing the Answer. A model writes the Answer from the sources found, adds numbered citations and suggests follow-up or alternative questions.

Models and where they run

• Answers are written by Anthropic’s Claude models. We use them through Amazon Bedrock, a service of Amazon Web Services (AWS), and only in AWS data centres in the EU.
• A Cohere embedding model (embed-multilingual-v3), also used through Amazon Bedrock in the EU, converts questions and documents into numbers so that relevant passages can be found.
• We do not send your name, email address or user identifier to the models. AWS states that model providers cannot access the prompts sent to Amazon Bedrock or the text it returns, and that this content is not used to train models.
• Summaries are written by the same models through the same service. See Section 6.
• For how we handle your personal data, see the Privacy Notice, “How AI processes your information”.

Study conditions

• ParliView is a research platform. To study how design choices affect the usefulness and accuracy of an AI system, we may use different AI models, instructions or settings for different groups of participants. We call these study conditions.
• Any participant may be placed in a study condition. Placement is part of the Study you agreed to take part in, and it does not affect your access to ParliView.
• We do not tell you which condition you are in while the Study runs, because knowing could change how you use ParliView and affect the results. After the Study ends, we will tell you if you ask.
• The AI labels, the notices and the rule that ParliView does not give voting advice apply in every condition.
• For the data we collect about your use and how we analyse it, see the Privacy Notice, “The Study and what taking part means for your data”.

Citations

• Most Answers include numbered citations that link to the sources used. A citation shows where information came from. It does not prove that the Answer reflects the source accurately.
• Some statements in an Answer may have no citation, and some Answers have no citations at all. Take particular care with anything uncited.
• Some cited sources are ParliView Summaries, which are themselves AI-generated. These are marked as ParliView summaries in the list of sources.

Why Answers vary

• The same question can produce different Answers. Language models do not give identical output each time, and our index and models change over time.
• Answers can also differ between study conditions, as explained above, and when we roll out an update.

Sources and licences

EP Material: © European Union, [year(s)] – Source: European Parliament. Parltrack Data is available under the Open Database Licence 1.0.

---

6. How Summaries Are Produced

• What they are. Summaries include summaries of debates (with a headline, a summary and key points for each speaker), legislative procedures, events, adopted texts and topics, including yearly topic overviews. Suggested starter questions, generated from European Parliament press releases, are also Summaries. Nobody checks them either, and you can report them in the same way (Section 9).
• How they are made. An AI model writes each Summary from EP Material and Parltrack Data. A Summary is generated the first time a participant opens the item. It is then stored and shown to every participant who opens the same item. Summaries are produced in the same way for every participant, whatever study group they are in. We do not generate Summaries in bulk in advance.
• Where they are made. Summaries are written by the same models as Answers, through Amazon Bedrock in AWS data centres in the EU. Only EP Material and Parltrack Data are used for this purpose. No participant data, questions or conversations are used.
• No human review. No person reviews a Summary before it is shown. Each Summary is labelled as AI-generated (Section 4.1).
• Timing. A Summary can be produced only after the Parliament has published its record and we have added it to our index, usually within a day or two, so Summaries lag behind the Parliament’s own publications. A Summary may not reflect corrections the Parliament makes to its record later.
• Quotations. Summaries do not contain quotations. Text under the heading “Key points (AI paraphrase, not a quotation)” is the AI system’s own description of what a speaker said. It is not a quotation, even where it resembles the record, and you must not republish it as one. UCD does not present AI paraphrase as quotation. A translation made by the AI is also a paraphrase. If a Summary contains quotation marks around text that does not match the Parliament’s record, report it (Section 9) and we will correct it. To see what a speaker said, follow the link to the verbatim report.
• What Summaries say about named people. A Summary that describes how an MEP voted or what they said is written by us from the public record. The law allows this for two reasons. The person made the vote or speech public (Article 9(2)(e) GDPR), and ParliView is a research platform (Article 9(2)(j) GDPR with sections 42 and 54 of the Data Protection Act 2018). If a Summary goes beyond what the record shows, it is wrong, and we correct or remove it when told (Section 9).
• Record of corrections. We keep the current version of each Summary. When we correct or withdraw a Summary, we record the date and the reason, and keep that record with the error report (Section 9). We do not keep every earlier version of a Summary as it was shown.

---

7. Known Limitations and Error Types

AI models write fluent text by predicting likely words. They do not understand information the way a person does. Basing Answers and Summaries on retrieved sources reduces errors but does not remove them, and an error can read as confidently as a correct statement.

Errors to watch for:

• Plausible but wrong statements. A sentence can sound right and still be false.
• The wrong MEP. A statement or vote can be attributed to the wrong MEP, for example where MEPs have similar names, belong to the same political group or spoke in the same debate.
• Misquotation. AI Output can put words in someone’s mouth, or present a paraphrase as if it were a quotation.
• Mixed-up proceedings. Similar debates, amendments and votes can be merged, including across committees, plenary sessions and parliamentary terms.
• Vote counts and results. Numbers, amendment references and results can be wrong in AI text. Vote tables come from Parltrack roll-call data, which can be incomplete. Where the final vote on a file is missing from our data, ParliView may show the most recent amendment vote instead, with a warning. The adopted or rejected label is worked out by comparing votes for and against. It can be wrong where a vote needed a special majority, so check the Parliament’s own vote results.
• Out-of-date information. ParliView lags behind the Parliament. MEPs change roles and political groups, and the Parliament corrects records after publishing them.
• Language coverage. English is currently the only language ParliView supports and has tested. It uses the English-language versions of European Parliament documents, and Summaries are written in English. You can ask questions in other languages, and ParliView may answer in them, but we have not tested those Answers and we cannot and do not vouch for their accuracy. Treat them as unchecked, and verify anything that matters against the sources. Speeches given in other languages may appear in the original language or in translation, and AI Output may represent them less faithfully.
• Political framing. Word choice, emphasis and the choice of which speakers or sources to include can make AI Output seem to favour a view, even when each statement is accurate. Contested positions can be simplified or presented as settled.
• Gaps. If ParliView finds nothing on a subject, that does not mean nothing exists. Our index does not cover every document or meeting. For example, it does not cover committee meetings.
• Errors passed on. An Answer that draws on a Summary can repeat an error in that Summary.
• Differences between conditions. Because study conditions use different models or settings, an error you see may not appear for another participant, and the other way round. Reports of errors help us compare conditions, so please send them.

---

8. How to Check What You Read

• Open the sources. Check that the cited document says what the Answer or Summary says.
• Check quotations and attributions. Before you attribute words or a vote to an MEP, find them in the Parliament’s verbatim report of the debate or its roll-call vote results.
• Check numbers, dates and results against the Parliament’s minutes, vote results and Legislative Observatory.
• Take care with uncited statements and with Answers that have no sources.
• Ask in another way. Specific questions tend to give more focused Answers. If an Answer matters to you, rephrase your question and compare the Answers.
• Look at the date. Check the dates of the sources and whether the Parliament’s record may have changed since.
• Keep the context. When you share AI Output, keep the question, the label and the sources with it (Section 4.3).

---

9. Reporting Errors and Requesting Corrections

Who can report. Anyone can report an error, including people who do not use ParliView, such as MEPs and their staff.

How to report. Email corrections@parliview.org.

What to include.

• Where the error is: for a Summary, its title, web address and when you saw it; for an Answer, your question, the Answer and when you received it.
• What is wrong and, if you know, what is correct, with a source.
• Whether the error is about you or someone you represent.
• How we can contact you.

Please do not send more personal information than we need.

What happens next.

1. We acknowledge your report within 2 working days.
2. We assess it against the sources.
   • Reports about a named person, such as a false quotation, a vote attributed to the wrong MEP or an inaccurate suggestion of wrongdoing: we assess the report within 2 working days. Where the report is credible, we hide the Summary from view while it is under review.
   • Other reports: we respond within 10 working days.
3. If a Summary is wrong, we correct it, replace it with a new version or remove it, and we record the date and the reason (Section 6). Statements of correction. If a Summary about a named person was wrong, we offer the person a short statement of correction and publish it on request. We keep the statement with the correction record.
4. If an Answer is wrong, we cannot recall copies a participant has already made. Where the Answer stored in a participant’s conversation contains inaccurate personal data about a named person, we correct or delete that stored text on request. We review the cause and may change instructions, sources or data to reduce the risk of the error happening again. Because an Answer may come from one study condition, we check whether the same error occurs in the others.
5. We tell you the outcome, including why we have not made a change. The times above are commitments, not aims.
6. We keep a record of the report and our decision for 3 years.

Text attributed to ParliView. You may see text that is said to come from ParliView and think it has been altered, taken out of context or did not come from ParliView at all. If so, tell us at corrections@parliview.org and send the text, with the question if you have it. We cannot confirm whether a text is an unaltered Answer. We can say whether it carries ParliView’s label and citations, whether the cited sources support it, and whether ParliView is designed to produce output of that kind. We share our assessment with the people affected. If the text is an Answer that contains an error, we also treat your message as an error report.

Your data protection rights. If AI Output contains inaccurate personal data about you, you have the right to have it corrected and, in some cases, to object to its processing. We also treat a report to corrections@parliview.org about inaccurate information about you as a request under those rights. For more information, see the Privacy Notice, “Withdrawing, and your rights”, or email privacy@parliview.org. Reporting an error does not affect any of your other legal rights.

---

10. What We Do to Reduce Errors

• Grounding in sources. Answers are written from passages retrieved from our index of EP Material and Parltrack Data, not from the model’s general knowledge alone.
• Citations. Answers include numbered citations where the model links statements to sources, so that you can check them.
• Checks on questions. We may apply automated checks to questions (for example keyword rules) before they reach the AI models, and we reserve the right to do so. These checks do not apply to AI Output.
• Named persons. For Summaries that name a person we apply two safeguards. Key points about what a speaker said are presented as AI paraphrase, not as quotation (Section 6). Where a credible report says a Summary misrepresents a named person, we hide the Summary while we review it (Section 9).
• Research on accuracy and fairness. Members of the ParliView research team have published research on how AI models summarise European Parliament debates:
  • Fair Representation in Parliamentary Summaries: Measuring and Mitigating Inclusion Bias (arXiv:2507.14221), which measures how fairly speakers are represented in AI summaries of debates; and
  • Evaluating LLM-Driven Summarisation of Parliamentary Debates with Computational Argumentation (arXiv:2604.19331), which tests whether summaries keep the arguments made in a debate.
  This research informs our work but has not yet been applied to the Platform in its current form.
• The Study itself. Comparing study conditions is one of the ways we find out which models and settings give more accurate and useful Answers. Findings from the Study feed back into how ParliView works.
• Review of reports. We review error reports and feedback, and use them to improve our instructions, data and processes.

What we do not do. No person reviews AI Output before you see it. We do not filter the content of AI Output; we only check its formatting (such as the format of citations) automatically.

---

11. Bias and Neutrality

• Our aim. We aim to present the work of the European Parliament fairly, without favouring any political group, party, MEP, country or point of view.
• The limits. We cannot promise that AI Output is neutral. AI models learn from very large amounts of text and can reflect biases in it. The choice of which sources and speakers to include can also tilt a Summary or Answer. In the research listed in Section 10, the team found that the AI models it tested represented speakers differently depending on when they spoke in a debate, the language they spoke and their political affiliation. We have not yet repeated these tests on the models ParliView uses.
• What we do. ParliView does not give voting advice and links to sources so that you can judge for yourself. Within each study condition, the same instructions apply whichever political group, party or MEP a question is about. Our project partners and funder have no editorial control over AI Output.
• Tell us. If you think AI Output is biased, email corrections@parliview.org with examples: the question, the Answer or Summary, and the date. For general comments, email support@parliview.org.

---

12. Questions and Complaints

• Questions or complaints about ParliView’s use of AI, this page or how we handled a report: support@parliview.org.
• Questions about the Study, including study conditions and withdrawal: see the Participant Terms and Information, “Taking part”, or email privacy@parliview.org.
• Data protection questions and requests: privacy@parliview.org. You can also contact UCD’s Data Protection Officer at gdpr@ucd.ie.
• Complaints to a data protection regulator: see the Privacy Notice, “Complaints”.
• AI transparency rules: You can also raise a concern with the national authority responsible for the EU AI Act where you live.

---

13. If You Are in the UK

• AI law. No UK law in force requires the notices and labels described on this page. We apply them to participants in the UK in the same way as to everyone else.
• Accuracy and correction. UK data protection law (the UK GDPR) requires personal data to be accurate and gives you the right to have inaccurate personal data about you corrected. The correction route in Section 9 is how you exercise that right in relation to AI Output.
• Complaints. You can complain first to UCD (see the Privacy Notice, “If you are in the UK”) and then to the Information Commission, the UK’s data protection regulator.

---

14. Version and Changes

• This is version 2.0 of this page, effective from 30 September 2026. It replaces the unversioned page “AI Accuracy & Guidance”.
• We update this page when the way ParliView uses AI changes: for example, when we change models, labels or marking. We also update it when the Study adds or ends a study condition that changes what this page says. The date of the last update is shown at the top.
• Earlier versions are available on request from support@parliview.org.

Main changes from the previous page:

• ParliView is described as a UCD research platform, and users as participants in the Study, including that models and settings can differ between study conditions.
• New sections on who is responsible under the EU AI Act and on what ParliView is for and what it is not, including that it does not give voting advice.
• New explanation of notices, labels and machine-readable marking.
• New explanation of how and where Summaries are produced, and that Summaries contain AI paraphrase, not quotations.
• Corrected statements about source links and about automated filtering.
• New process for reporting errors, with one set of response times, a rule for hiding a reported Summary during review and statements of correction, and a new section on bias and neutrality.
• New UK annex.