Researchers say some elected representatives use AI tools to help produce speeches and legislative text without telling the public. The tools themselves are ordinary; what is missing is any settled rule on when such assistance must be disclosed.
Key takeaways
- Reporting on recent academic work suggests that generative AI is being used inside legislatures to help draft political speech and legislative material, in at least some European parliaments.
- The central concern raised is not that AI wrote something, but that voters have no reliable way of knowing when it did.
- Most parliaments have no binding rule requiring a member to declare that a speech, motion or amendment was drafted with machine assistance.
- Legislative work has always involved ghostwriting by staff, party researchers and civil servants, which makes the disclosure question harder than it first appears.
- The precise scale of AI use inside any particular parliament is not publicly established, and figures circulating in social media summaries of such studies should be treated with caution.
What is actually being claimed?
The claim circulating is that a study examined how members of parliament and their offices use generative AI systems, and found that such tools are being used to help write speeches and draft legislative texts, without any corresponding declaration to the electorate. Sweden and the United Kingdom have been named in coverage of the work.
It is worth being precise about what a claim of this shape can and cannot establish. Research into professional AI use normally relies on surveys, interviews, or textual analysis of published documents. Each method has limits. Surveys depend on self-reporting by people with an incentive to understate. Textual detection of machine-generated writing is unreliable, particularly for the formulaic register in which political and legal documents are written. So the honest summary is that AI assistance appears to be present in parliamentary work, that it is largely undeclared, and that the exact extent is not publicly known.
Why is this surfacing now?
Two things changed at roughly the same time. General-purpose language models became competent at exactly the sort of text that parliamentary work generates — short speeches, constituency correspondence, committee questions, explanatory notes, amendment wording. And those models became available through ordinary consumer subscriptions, meaning an individual staffer can adopt one without any institutional procurement decision, security review or policy sign-off.
That combination produces a gap. Institutions typically regulate technology they have bought. They regulate far less effectively technology that arrives through a browser tab on a personal account. Attention to the issue has grown as parliaments across Europe have begun drafting their own internal AI guidance, and as researchers have started asking what is already happening rather than what should happen.
How did AI reach the legislative desk?
Legislatures were not caught unaware of automation. They have used digital drafting systems, standardised clause libraries and template-driven document management for decades. Legislative drafting in particular is a highly conventionalised craft, with fixed formulae for commencement, definitions, amendments and repeals. That conventionality is precisely what makes it tractable for a language model: much of the surface form is predictable.
Political speech is similar in a different way. A backbench speech has a recognisable structure — a local example, a national statistic, a request to a minister, a closing line. Producing plausible drafts of that structure is close to the core competence of current models.
So the technology did not have to be pushed. It fits the existing shape of the work, and it entered through the same door as any other productivity tool: individuals trying to keep up with volume.
Is drafting a speech the same as drafting a law?
No, and conflating the two makes the debate murkier than it needs to be. A speech is an expression of a representative’s position. It has long been drafted by other people — speechwriters, advisers, party communications staff — and no democracy has treated that as fraud. The convention is that the member owns the words when they say them.
Legislative text is different. Its wording has legal effect, is interpreted by courts, and interacts with a body of existing statute in ways that are not obvious from the text alone. In most parliamentary systems, government bills are written by specialist drafters with formal legal training, and the drafting office functions as a quality control layer. Material originating from individual members — private members’ bills, amendments, motions — often passes through fewer checks.
The risk profile therefore differs. For speech, the question is authenticity. For statute, the question is correctness: whether a fluent, confidently worded clause actually does what its sponsor believes it does.
Who is affected, and how?
Voters are affected in the sense that a representative’s own reasoning is one of the things being elected. If the reasoning is outsourced and undeclared, the electoral signal degrades, though it is difficult to measure by how much.
Parliamentary staff are affected most directly. They are the people using the tools, usually without formal guidance, and they carry the professional risk if something goes wrong. Clerks, drafting offices and legal advisers absorb the downstream consequences of poorly constructed text.
Beyond the legislature, courts and regulators inherit any ambiguity that survives into law. Confidentiality is a further exposure: material pasted into a consumer AI service leaves the institution’s control, which matters for unpublished amendments, committee material or constituent casework containing personal data.
Where do informed people disagree?
There is a genuine, unresolved argument here rather than a settled consensus with holdouts.
One position holds that AI assistance is simply the next tool, comparable to a word processor, a search engine or a research briefing, and that mandatory disclosure would be both unenforceable and faintly absurd. Nobody declares which paragraphs a researcher wrote.
A second position holds that scale changes the character of the thing. A tool that can generate an entire argument, rather than assist with one, is not analogous to spell-check, and representative democracy has a legitimate interest in knowing whether a position was reasoned or generated.
A third position is procedural: disclosure is the wrong target, and the real fix is verification. On this view it does not matter who or what wrote a clause, provided a competent human is accountable for checking it and formal review is strengthened.
There is also honest disagreement about the evidence base itself. Some argue that studies in this area systematically undercount usage; others argue that coverage tends to overstate what limited samples can support.
What does this mean in practice?
In the near term, most institutions will not ban these tools; the demand pressure is too high and enforcement would be impractical. The likelier trajectory is internal guidance: rules on what categories of material may be processed by external services, what must remain inside approved systems, and where a named human must sign off.
Disclosure regimes, if they arrive, will probably be narrow rather than universal — applied to formal legislative instruments rather than to every piece of correspondence or every speech. Broad declaration requirements collide with the fact that political text has always been collaboratively produced.
For readers, the practical implication is modest and worth stating plainly: the presence of AI in a document is not by itself evidence of bad faith or poor quality, and its absence is not evidence of care.
What should you watch next?
Watch for parliaments publishing formal AI policies, and read what those policies actually cover — procurement and approved tools, or member conduct and declaration. The two are very different commitments.
Watch codes of conduct and registers of interest, which are the existing machinery through which any disclosure duty would most plausibly be implemented.
Watch drafting offices and legislative counsel, whose institutional response is the most informative signal about whether AI-assisted text is creating real quality problems.
And watch for replication. A single study describing a novel phenomenon is a starting point. Whether the pattern holds across more parliaments, with better methods, is what will determine if this becomes a durable governance issue or a passing observation.
Frequently asked questions
Is it illegal for a politician to use AI to write a speech?
In general, no. Parliamentary rules typically govern conduct, declarations of financial interest and the accuracy of statements, not the tools used to prepare text. Speechwriting by staff has always been accepted practice. Some institutions are now drafting internal AI guidance, but a legal prohibition on AI-assisted drafting is not the norm in European parliaments as far as is publicly documented.
Can you tell if a law was written by AI?
Not reliably. AI detection tools produce both false positives and false negatives, and they perform particularly badly on formal, conventionalised writing. Legislative text is highly templated by design, which makes it resemble machine output regardless of origin. Any confident claim that a specific document was machine-generated, based on detector output alone, should be treated sceptically.
Does AI-assisted drafting make laws less accurate?
It can, though this depends heavily on review. Language models produce fluent text that may misstate how a clause interacts with existing statute, or invent references that do not exist. Where drafting passes through professional legislative counsel, such errors are likely to be caught. Where it does not — private members’ material, amendments, motions — the risk of unreviewed error is higher.
Why does disclosure matter if staff already write speeches?
Because the arguments differ in kind, according to those who favour disclosure. A staff writer shares the member’s political context and is accountable within the institution. A model generates plausible argument without commitment or accountability. Critics of disclosure respond that the distinction is hard to operationalise and that voters judge positions, not authorship. This disagreement is unresolved.
Are Sweden and the UK unusual in this?
There is no basis for saying so. Those two countries appear in coverage of the research, which most likely reflects where the researchers looked rather than where usage is highest. Generative AI tools are widely available across Europe and beyond, and no comparative ranking of parliamentary AI use has been publicly established.
What would a sensible policy look like?
Most proposals combine three elements: restrictions on what confidential material may be sent to external AI services, a named human accountable for the correctness of any formal text, and disclosure obligations limited to legislative instruments rather than all political communication. Whether such rules are adopted, and whether they are enforceable in practice, remains open at present.
Sources and further reading
- Academic research on technology adoption in legislatures, typically published in political science and public administration journals.
- Parliamentary procedure committees and codes of conduct, which set out existing rules on declarations and member conduct.
- Offices of parliamentary counsel and legislative drafting guides, which document how statutory text is produced and reviewed.
- National data protection authorities, for guidance on processing confidential or personal data through third-party AI services.
Surfaced from the reddit:technology signal “undisclosed AI use in parliaments”. AI-assisted draft, editorially reviewed.
