How can an interest organisation use AI?

Most obviously for four things: reading and structuring large documents such as bills and consultation material, drafting member communication in the association's own voice, following the political agenda, and easing the running of the secretariat with minutes and knowledge look-ups. Positions and political priorities must still come from the political system.

May the secretariat use ChatGPT for consultation responses?

As an aid to reading and structuring, yes – to summarise bills, find relevant sections and build an outline. But the position in the response must have a political mandate, and confidential or personally identifiable information must not be entered into public AI services without an agreement that protects it. Write the ground rules down, so the judgement does not rest with the individual employee.

How does the association keep its voice when AI writes along?

By making the voice explicit before the machine writes: written-down positions, word choices and words it avoids, against which prompts and drafts are held. And by insisting that AI delivers drafts – the sender is always the association, and approval lies with people who have a mandate.

What does AI mean for work with language?

Who is actually speaking when the organisation communicates? In a private company, the answer is usually simple – the company speaks, and management owns the voice. In a membership organisation, the answer is never simple. The secretariat writes, the chair is quoted, the board has adopted, and the members must be able to recognise themselves in all of it.

The question has become pressing because artificial intelligence has made text cheap. Work with language – writing, phrasing and communicating on behalf of a sender – can increasingly be done by machines, and that makes it tempting to measure the technology by what it does best: speed and volume. But work with language has never consisted only of producing text, and that is why it is worth looking more closely at what actually happens to the work when the machine writes along – and what cannot be carried over.

Three questions every text must be able to answer

All professional work with language must be able to answer three questions. Why are we doing this work – does it make sense? Why is this the right way to solve the task – is it legitimate? And why is this output the right one – does the quality hold? The three questions correspond to the why, how and what of the work, and the answers constantly affect one another: a text can be linguistically flawless and still be wrong, because the way it was made or the mandate behind it is not in order.

Artificial intelligence does not change the three questions – but it changes how easy it is to overlook two of them. The machine takes care of production, and so attention lands almost by itself on quality: is the output good enough, and does it solve the task? In much work with language, however, legitimacy weighs more heavily. The way the task is solved and the mandate behind the words are part of the product itself – the text is not right just because it is good. This becomes clearest in associations and politically led organisations, where many voices must speak as one. That is why the rest of this article draws its examples from there – but the shift from quality to legitimacy exists in all work with language where someone speaks on behalf of others.

Three conditions shift the weight towards legitimacy

The first condition is political leadership. The chair and the board are elected, and the organisation's positions must have a mandate – they are decided in a political system, not formulated at a desk. That means something very concrete for AI use: a language model can formulate a convincing position in ten seconds, but the mandate does not come with it. The text can be finished long before the decision has been made – and that sequence is dangerous in a politically led organisation.

The second condition is member legitimacy. An association exists by virtue of its members' trust and subscriptions, and communication with members is not marketing – it is part of member democracy. A text that is effective but sounds foreign therefore costs on a different account from the conversion rate: it wears down the sense that the organisation belongs to its members.

The third condition is the public. Interest organisations work in the open – consultation responses are published, statements are quoted, and opponents are reading along. A factual error from a language model is an internal blunder in a company; in a consultation response, it is a public document that can weaken the organisation's credibility in precisely the area where it needs to be strongest.

Four uses that work in practice

The conditions are not an argument against AI – they are an argument for choosing the uses with care. Four recur in the organisations I work with.

Consultation responses

The biggest gain lies not in letting the AI write the response, but in letting it read the proposal. Bills and consultation material often run to hundreds of pages, and the AI can summarise, find the sections that affect the members and build an outline. A trade association I have trained halved the time from bill to first draft this way – while the position and the priorities still came from the political system. Reading and structuring can be delegated to the machine; the position cannot.

Member communication

Newsletters, replies to members and communicating political results are recurring writing tasks where AI can deliver drafts – if the association's voice has been written down first. In a professional organisation I have advised, the most important tools turned out not to be the licences but a document: a description of how the organisation speaks, which words it uses and which it never uses. With that in place, the drafts became recognisable instead of generic.

Political monitoring

Secretariats spend a disproportionate amount of time keeping up: agendas, parliamentary questions, committee consultations, the media. AI can read the stream and prioritise what is relevant to the organisation's particular causes – not as a replacement for the policy adviser's judgement, but as a filter that ensures the judgement is spent on what matters. The typical result is that the organisation responds faster to what actually concerns it.

Running the secretariat

The least glamorous use is often the most valuable, especially in small secretariats with many tasks per employee: minutes from board and committee meetings, meeting preparation, translations and look-ups in statutes, agreements and old decisions. These are tasks where the AI saves hours without touching anything political – and therefore a good place to gain experience before approaching the sensitive texts.

The pitfalls are about the sender

The most important pitfall is not technical, and it is not about data security – it is about the sender. Who is speaking when the AI writes for an association?

A language model writes the average of everything it has read. But an association is, by definition, not the average – it exists because it holds a particular view on behalf of particular people. When the machine does the phrasing, there is a risk of what I call position drift: wording that sounds reasonable and professional, but that nobody has decided. A consultation response that is slightly more conciliatory than the board is. A newsletter that promises a little more than the congress has adopted. These are not language errors – they are small shifts in what the organisation thinks, and in a politically led organisation that is a democratic problem before it is a communication problem.

The safeguard is not to abolish the tool, but to make the ground rules explicit: what may AI be used for without further ado – reading, structuring, drafts of the non-political? What requires political approval, however finished the text looks? And who has the mandate to say no to a text that is well phrased but is not the association's? The organisations that write the answers down move the judgement from the individual employee's gut feeling to a shared standard – and that is the difference that holds when the pressure rises.

What I have learned from teaching

I teach politically led organisations about AI and communication at Altinget's masterclass, and there is one pattern I meet every time: the conversation starts with technology and ends with organisation. Participants arrive with questions about tools and leave the room with questions about mandate – who approves, who speaks, where the line goes. In my experience, that is the right place to land, because that is where the difference between good and risky AI use is decided.

I can also see that secretariats are often further ahead than their guidelines. Employees are already using the tools – the question is not whether, but on what terms. The same themes recur when I give keynotes on artificial intelligence for boards and councils of representatives, and in the AI advisory I do for secretariats: the need is rarely more technology, but clearer ground rules for the technology already in use.

If your organisation needs to start somewhere, I would suggest three questions: which tasks in the secretariat are reading and structuring rather than positions? Who has the mandate to approve a text the AI has helped write? And would your members be able to recognise you from three sentences without a logo – also in the texts the machine has had first hand in? Organisations that can answer those questions do not just get more out of the technology. They protect what is hardest to rebuild: the members' trust that it is still them speaking.