#inextremis
A leadership perspective for European associations
When AI speaks for your association
Protecting your mandate, proving member value and leading communications in the AI era
The strategic question is not how much more you can publish. It is what your communication enables people to understand, trust and do together.
October 2026
A few words. A different mandate.
Imagine a European federation preparing a briefing after three weeks of discussion. Its approved account includes a carefully qualified sentence:
"Some members are considering supporting the proposal."
An AI tool makes the wording more direct:
"The sector supports the proposal."
The sentence is shorter. The meaning has changed.
A possibility has become a commitment. Some members have become an entire sector. A discussion has become a position.
Nothing in the new sentence sounds obviously wrong. That is what makes this fictional example important. The failure is not awkward writing. It is a claim to collective authority that nobody authorised.
For European associations, this is where the AI conversation needs to begin. Not with a list of tools, and not with an abstract choice between enthusiasm and caution. With the relationship between what an organisation knows, what its members have agreed and what it is entitled to say.
The opportunity is just as substantial. A small secretariat can make specialist knowledge easier to use, help national members work across languages and give people clearer routes into a discussion. AI can support preparation, comparison and adaptation. Used well, it can create room for the work that makes an association valuable: interpretation, relationships and collective action.
But those benefits do not follow automatically from faster production. They depend on decisions about whose contributions remain visible, which claims travel under the association's name and where any additional capacity goes.
At #inextremis, our position is that AI should strengthen an association's ability to represent, create value and earn trust. Production efficiency matters because it can serve those purposes, not because more content is an achievement in itself.
The chapters connect these choices to a self-assessment of current practice and tools for acting on the findings. Legal and institutional detail sits in a separate reference section.
The ambition is not to make communication less human. It is to make human judgement count where it matters most.
The leadership brief
Three shifts deserve a place on your leadership agenda before the next discussion about AI tools.
From fluent content to legitimate representation
An AI-assisted statement can be factually correct yet attributed to the wrong body, stripped of a negotiated qualification or presented as a consensus that does not exist. Review must cover evidence, authority and the people represented. The same discipline applies to the summaries that inform decisions and the external AI answers that describe your organisation.
From information supply to member value
As generic explanations become easier to obtain, associations have a stronger reason to demonstrate what membership enables: trusted interpretation, meaningful participation, access to expertise and the capacity to act together. Use AI to improve those experiences. Do not confuse a larger content inventory with a stronger membership proposition.
From content supervision to accountable communication
A useful test for any proposed use: will it help us serve members better while preserving their ability to understand, challenge and authorise what is said in their name?
Your reading route
The strategic argument and self-assessment
1. When better writing becomes a different mandate
2. What will members still need you for?
3. A faster consultation is not necessarily a fairer one
4. When credibility can be manufactured
5. Be represented, not just retrieved
6. The head of communication's new mandate
7. Capacity is a leadership choice
8. Responsibility travels with the message
9. Is AI strengthening your association?
Practical resources and reference
10. Four tools for your next leadership meeting
11. An association landscape, not an adoption race
12. Legal decisions, use by use
13. What institutional practice can teach you
Evidence, method and further reading
When better writing becomes a different mandate
The most consequential editing error may be a sentence that sounds better than the truth it replaces.
An association's wording can contain the outcome of negotiation. A reservation may have secured a member's agreement. A verb may separate a recommendation from a requirement. A reference to a working group may deliberately avoid committing the board.
These are not imperfections awaiting correction. They are part of the meaning.
That makes the familiar instruction to "make this clearer and more confident" incomplete. Clarity should help readers understand the actual position. It should not resolve uncertainties that the organisation has not resolved itself.
Three checks, not one
We propose a mandate check for substantive communications. It asks three distinct questions.
Evidence: is the claim supported? Examine the source, date, context and uncertainty. An approved message still needs a sound factual basis.
Representation: whose view does it express? Distinguish an individual member, a committee, the association and a wider sector or community. Approval within the organisation does not expand its mandate beyond the people it represents.
These checks cannot substitute for one another. Unanimity does not establish scientific validity. Strong evidence does not confer authority to speak. A member's experience does not establish a sector-wide condition.
This is an editorial framework, not a legal classification or a replacement for your statutes. Its value is in making a hidden change of meaning visible before publication.
Make status part of the source
A collection of accurate documents is not enough if the system cannot distinguish an adopted position from a consultation draft. Give source material a clear status: what it is, who owns it, when it was approved and whether it remains current.
Keep material still under discussion separate from the approved collection used for routine drafting. Where an analysis needs both, label them explicitly and require the output to preserve that distinction. Link substantive claims to the passages supporting them; do not accept a plausible bibliography as evidence that the wording is supported.
Across languages, protect meaning rather than word-for-word similarity. A translation may adapt syntax while preserving who is speaking, how firmly they are committing and what remains conditional. A shared glossary should record sensitive qualifications as well as technical vocabulary.
Decide what cannot be delegated
An assistant proposing a paragraph and an agent publishing one have different consequences. Access to a website, mailing list or member database must follow a separate decision about permitted actions. Drafting permission is not publication permission.
The same principle applies under deadline pressure. Reuse an approved answer when appropriate. Hold a new substantive interpretation until an authorised reviewer is available. Urgency does not transfer the board's authority to a tool.
Onora O'Neill's work on trust offers a useful foundation: more information does not, by itself, make trust well placed. People need to examine claims and those making them. Our application is straightforward: a reader should be able to identify the speaker, understand the evidence and find a route to question the account. [1]
The leadership decision is to name the statements that always require explicit approval, then give reviewers the time and authority to protect them.
What will members still need you for?
A faster briefing is useful. A clearer reason to belong is strategic.
An external AI assistant may help a member explain a policy dossier. Its answer is not a position adopted by that member's federation. It does not provide the relationship through which a concern reaches a committee, or the agreement that enables several organisations to act together.
That distinction is an opportunity. It also asks something of associations whose communication proposition rests heavily on collecting and redistributing information.
Our argument is not that AI makes membership redundant. It is that easier access to generic explanations raises the value of being explicit about what membership adds. The association should not ask its newsletter to carry a value proposition that belongs to its whole organisation.
Move from what you send to what members can do
Consider a fictional national member that receives every briefing but rarely joins technical consultations. It lacks the specialist time to work through a long dossier before the deadline. Its practical experience never reaches the debate.
The federation redirects effort from a duplicate summary to a guided route into the consultation. Using approved sources, AI helps staff prepare a short explanation around the member's question and identify points requiring expert input. A specialist checks the interpretation. The member can then submit a practical example, and the secretariat shows how it informs the committee's discussion.
The change is a member able to contribute, not another document sent. AI supports preparation; the service also depends on expert access, clearer questions and follow-through. The investment choice is to support participation rather than add another summary to the inbox.
Ethan Mollick's experimental approach to working with AI is useful here: explore where assistance helps rather than limiting the discussion to one task. For associations, the test should extend beyond drafting to the service members actually experience. [2]
Make the distinctive contribution visible
Trusted interpretation means explaining what an issue changes, what remains uncertain and which decision matters next. Access to expertise means a member can reach someone qualified to address the difficult question, not simply receive a confident answer. Collective deliberation means different experiences can influence an outcome. Relationships make it possible to work across organisations and borders.
Show these benefits in use. A briefing can identify questions members raised; a consultation can explain how their contributions reach decision-makers; an annual report can connect activity to a change they experienced. Make the association's contribution precise enough for members to judge it, without claiming every success.
Decide what to open, what to offer and what to protect
Separate three choices. Public knowledge can help others understand the issue and recognise the association's expertise. Member services should solve a need that merits membership. Confidential deliberation needs space for people to explore differences before a position is settled.
Publishing everything is not a strategy. Neither is putting undifferentiated information behind a login. The boundary should follow purpose, rights and the expectations of those contributing.
Before commissioning an AI service, compare it with simpler changes. A current, searchable document library may outperform a chatbot built over a neglected archive. A glossary may resolve recurring translation problems. More specialist time may be the right answer to a complex dossier.
Ask a different investment question: what will a member be able to understand, contribute or accomplish that they cannot do as well today?
A faster consultation is not necessarily a fairer one
When a summary determines what reaches the board, summarisation becomes part of governance.
Imagine a federation collecting views from 50 members. Most submissions are detailed and written in polished English. Three smaller national organisations submit short responses describing an implementation problem. An AI-generated synthesis reports broad support and groups the remaining comments under "minor practical concerns".
The board receives a clean account. The three members see something else: the issue they raised has lost its significance before anyone has examined it.
This is a fictional scenario, not a finding about how often tools behave this way. It illustrates a failure worth testing. A system can preserve many individual facts while changing the balance of the discussion.
Frequency is not the same as significance
The most repeated theme may be important, but repetition alone does not establish its weight. A single contribution may identify an obstacle affecting an under-represented group or a condition on which implementation depends.
Length, language and confidence can also influence what is easy to summarise. Do not allow those features to become accidental voting rules. Equally, do not give every comment identical decision-making weight where the association's governance provides otherwise. The point is to make the basis of judgement explicit, not to replace one hidden weighting with another.
Dominique Cardon's examination of algorithmic classification encourages attention to the choices behind what becomes visible. Applied to a consultation, the question is practical: what does the method treat as important, and who approved that choice? [3]
Design for disagreement, then close the loop
Ask the analytical process to distinguish points of agreement, conditions attached to support, unresolved objections and contributions requiring clarification. Preserve traceability to the originals within appropriate access rules. Have a reviewer examine unusual or low-frequency concerns rather than validating only the dominant themes.
A summary should also identify missing evidence. No response from a group is not proof of agreement. A simulated member persona may suggest questions to ask; it is not a substitute for consulting the people concerned.
Give participants a way to correct how their contribution has been represented. Then show what happened: which questions reached the committee, what changed and why some suggestions were not taken forward. Confidentiality may limit what can be disclosed, but it need not make the process invisible.
Jim Macnamara's research treats organisational listening as work requiring culture, resources, skills and links to decision-making. An analysis tool supports that work. It does not complete it. [4]
Use AI to widen the route in
The opportunity is to lower barriers to participation: a clearer invitation, an initial translation, a plain-language explanation or an alternative format. Test those materials with the people expected to use them. Accuracy is part of accessibility; W3C guidance warns that automatic captions need correction when they fail to convey speech accurately. [5]
In our opening federation, a smaller member points out that a consultation question assumes access to specialist legal support. Rewriting it to invite practical experience allows a different kind of knowledge into the discussion. AI can assist the drafting. Listening reveals what needs to change.
A consultation succeeds when the association can explain whose views informed the decision, not merely how quickly it processed their words.
When credibility can be manufactured
The challenge is not only a false claim. It is a false claim surrounded by apparent reasons to believe it.
Picture an association preparing a public event. A prospective speaker has a polished biography, several articles mentioning their expertise and a convincing account in an AI-generated search answer. The sources appear to confirm one another.
A closer check reveals that the articles repeat the same unsupported biography. The generated answer has recycled those references. What looked like independent corroboration is a circle.
This fictional example illustrates manufactured authority. The same pattern can invent a relationship, imply an endorsement or build a controversy around something an organisation never said. AI can assist production and translation; it does not make every deception successful. Distribution, independent scrutiny and the audience's response still matter.
Check the evidence behind the reputation
For consequential claims about expertise, affiliations or support, look for an origin that can be independently checked. A cited institution should be able to confirm a role. A claimed publication should exist in the relevant publisher's records. Several websites repeating one account should not be counted as several witnesses.
Apply the same discipline to your own campaigning. A generated testimonial is not evidence of a member's experience. Simulated supporters do not establish public support. The fact that a campaign's objective is legitimate does not legitimise fabricated participation.
Audiences should also be able to distinguish illustration from documentation. A generated image may explain an abstract concept. It should not appear to record a real conference, community or endorsement. A small "AI-generated" label cannot repair a false claim about what happened.
For synthetic voices and dubbing, agree the speaker's permitted uses, languages and distribution. Approval of one script should not become indefinite permission for future performances. These are our recommended editorial standards; the separate legal transparency rules are addressed in Section 12.
Respond without turning criticism into an enemy
Keep honest error, deliberate deception and legitimate disagreement distinct. Propaganda may combine true statements with selective framing; criticism may be uncomfortable without being false. Wardle and Derakhshan's work on information disorder supports examining origin, intention and circulation rather than treating every unwelcome claim as "fake news". [6]
Prepare the practical response in advance. Preserve the material and its timing. Check authenticity and source independence. Establish who has encountered it and what confusion or harm it is causing. Then choose proportionately between private clarification, platform reporting, public correction and specialist support. Protect people being impersonated or targeted.
Sometimes a public rebuttal gives an obscure claim its first substantial audience. Silence is not always right, but visibility is not always the remedy either.
Build a public record that can withstand questions: current positions, identifiable spokespersons, authentic evidence and a usable route to correction.
Be represented, not just retrieved
In an AI-generated answer, appearing under the wrong description is not a communications success.
A journalist asks an AI assistant what your association stands for. The answer combines an old position paper, a national member's statement and a third-party description. Each source may be real. The resulting account may still attribute a position your association has never adopted.
This illustrative scenario brings Generative Engine Optimisation, or GEO, into the same discussion as the mandate check. The issue is not only whether the organisation appears. It is whether an audience can recognise its identity, understand its current position and verify what has been attributed to it.
An association therefore has two AI agendas: the tools it chooses to use and the external systems already capable of interpreting its public information. Choosing not to use AI internally does not remove the second question.
Make the official record unambiguous
Start with information you control. Explain whom the association represents, how it reaches positions and where its remit stops. Distinguish adopted policy from consultation material, research findings and individual contributions.
Give important documents a date, an owner and a visible status. When a position changes, mark the older version as superseded and direct readers to the current account. Keep a traceable archive where needed; removing history is not the same as clarifying the present.
Make claims easy to check through relevant original evidence. Align language versions and explain which takes precedence where necessary. Encourage accurate descriptions through genuine partnerships and contributions, not manufactured mentions.
These actions serve people first. They also give retrieval systems a clearer record to interpret. They cannot force an external service to use it correctly.
Measure the quality of the answer
Use a repeatable set of questions about your organisation, its remit and its principal subjects. Record the service, date, language, wording of the question, answer and cited sources. Repeat observations rather than treating one answer as a stable verdict.
Assess three dimensions separately. Visibility: does the association appear when relevant? Accuracy: are its identity, mandate and current position right? Usefulness: can the reader reach suitable evidence or a meaningful next step?
Do not let a high visibility score compensate for an invented commitment. A serious attribution error deserves attention even when the organisation is cited frequently. The scorecard in Section 10 keeps that distinction explicit.
Correct the underlying public record where it is unclear. Where the record is already clear, use an available feedback or correction route and monitor subsequent answers. Record what you cannot control rather than promising that one intervention will fix every service.
Keep GEO in its place
The original academic GEO research investigates visibility in a defined experimental setting, not guaranteed performance for every organisation or service. Google's guidance says its own generative search features do not require special AI files or content rewritten solely for AI, and warns against inauthentic mentions. Other services may work differently. [8, 9]
For #inextremis, this makes GEO an extension of editorial quality and institutional representation, not a reason to replace a communication strategy with optimisation tactics. Keep direct member relationships, your website, email, media relations and events central.
The objective is an accurate public account of who you are, not a larger number of appearances you cannot stand behind.
The head of communication's new mandate
AI makes the conditions behind a message as important to manage as the message itself.
When more people can generate publishable-looking material, the Head of Communication faces a choice. Become the final checker of an expanding queue, or shape how reliable communication is produced in the first place.
We argue for the second. It requires leadership support, not a heroic individual expected to inspect everything. The role is to establish an editorial system in which sources, responsibilities and escalation are clear enough for colleagues and partners to work well.
Protect the meaning behind the brand
A consistent tone can conceal inconsistent commitments. Two texts may sound equally confident while making different claims about what the association supports.
Brand governance therefore needs to reach beyond vocabulary and visual identity. Clarify the positions that may be adapted, the qualifications that must survive and the points at which adaptation becomes a new interpretation requiring approval.
A campaign can use different examples for different audiences without promising different things. When national members genuinely disagree, attribute the difference. Do not ask a model to manufacture a common voice that the governance process has not produced.
Originality matters too. In a short-story experiment, Anil Doshi and Oliver Hauser found that AI assistance improved assessed individual creativity while making stories more similar to one another. That is not proof that association campaigns will converge. It is a reason to ask whether your editorial choices remain distinctive. [10]
The answer is not artificial eccentricity. It is a recognisable perspective: the questions you choose, the evidence you examine, the expertise you bring and the contribution you ask audiences to make. A prompt requesting a "bold tone" is not a substitute for that work.
Give editorial authority practical form
Heads of Communication should be able to challenge an unsupported attribution, pause a misleading publication and require qualified review where the substance demands it. They also need access to the people who know when a position has changed.
That does not make communications responsible for every legal, technical or policy decision. Leadership authorises the purpose. Subject experts validate interpretation. Legal, data protection and IT specialists assess matters within their competence. Communications connects those decisions to what audiences receive.
A small team can make this workable without a new committee for every draft. Define routine adaptations, sensitive interpretations and actions requiring explicit approval. Agree who decides in each case and what happens when that person is unavailable.
Report on what communication enables
A dashboard that rewards only output may hide deterioration elsewhere. Pair production measures with questions about understanding, participation, material corrections and the accuracy of external representation.
This gives the function a more useful leadership conversation. Rather than "we produced twice as many versions", it can explain whether members recognised the decision, found the evidence and knew how to contribute.
The strategic contribution of communications is to make the association's voice useful, recognisable and accountable, however the first draft was produced.
Capacity is a leadership choice
Time saved becomes value only when someone decides what it is for.
A briefing takes less time to draft. The team can now investigate a difficult question, speak to a less-engaged member, improve accessibility or learn a new skill. It can also be asked to produce more briefings.
Neither outcome is built into the technology. Treat the allocation of capacity as part of the investment decision, not a conversation to have after the pilot.
Measure the whole job
Research gives good reasons to test AI, but not to promise a ready-made return. Noy and Zhang found shorter completion times and higher assessed quality in an experiment on professional writing. Brynjolfsson, Li and Raymond found an average productivity improvement of around 15% among customer-support agents at one company, with substantial differences between workers. Those settings are not a forecast for your policy paper or crisis response. [11, 12]
Count preparation, generation, verification, correction, approval and maintenance. Add licences, training, security support and expert attention. A striking first draft can still be expensive if one specialist spends the afternoon identifying subtle errors.
Measure the route to an approved, usable result, not the route to generated text. Agree in advance what would justify continuing, narrowing or stopping the use.
Your existing strengths and weaknesses matter. Specialist knowledge, trusted relationships and an approved archive can support useful assistance. Outdated documents, weak language coverage or dependence on one reviewer can undermine it. Better retrieval is not a cure for poor source management.
Keep the conditions for expertise
Junior colleagues learn by researching, constructing arguments and receiving feedback. Preserve opportunities to work from original sources and explain why a plausible answer is inadequate. Asking people to review material competently while removing the work through which competence develops is a poor long-term bargain.
Lee and colleagues found relationships between confidence in AI and self-reported critical-thinking effort. The study does not establish inevitable cognitive decline; it reinforces the case for deliberate work design. The ILO likewise distinguishes task exposure from the disappearance of whole jobs. [14, 15]
Involve staff in deciding which tasks to change. Reviewers need knowledge, time and permission to reject an output. People should be able to report an unsuitable use without being blamed for finding it.
Decide who benefits
Acemoglu and Johnson's work challenges the assumption that technological progress automatically distributes its benefits widely. Applied here, the management question is whether efficiency improves service and working conditions or simply intensifies production. [13]
There are material costs beyond the team. Kate Crawford places AI within wider systems of labour, infrastructure and resources. Proportionate use starts with choosing worthwhile tasks and asking meaningful supplier questions, rather than treating endless generation as a goal. [16]
Put the destination of any released capacity in the work plan. Otherwise, the next deadline is likely to decide for you.
Responsibility travels with the message
Outsourcing a task should not make responsibility for its meaning harder to find.
A statement begins in the secretariat, passes through an agency, is translated by a freelancer and is adapted by a national member. AI may assist at any point. The final wording may reach its audience far from the people who approved the original.
A policy addressed only to employees leaves much of this chain untouched. The answer is not to document every spelling suggestion. It is to agree what must remain under control wherever the work happens.
Make the handover explicit
For substantive material, pass on the current sources, the status of the position and the qualifications that must survive adaptation. Specify permitted tools and information, when material AI use must be disclosed, who approves the final version and how an error is reported.
A confidential draft entering an unapproved service is materially different from an approved tool correcting punctuation. Focus reporting on consequences for data, rights, meaning and verification.
National members also have their own mandates. Make clear when they reproduce a federation's statement and when they speak for themselves. The association should not erase those differences in pursuit of a uniform campaign.
Govern the tools people actually use
Inventory AI embedded in design software, translation services, meeting tools and member-management systems, not just named assistants. Ask colleagues what they use and which problem it solves. Rules that cannot accommodate real work encourage a gap between policy and practice.
Limit permissions deliberately. External documents or web pages may contain instructions intended to mislead a connected system. Analysing those sources should not automatically authorise the tool to retrieve unrelated files, send messages or publish. ANSSI's guidance treats security as a property of the wider system and its access arrangements, not just the model. [17]
Before pooling member information, examine whether it is personal, confidential, rights-protected or commercially sensitive. For a trade association, even non-personal information can raise competition questions when it concerns competing members. Section 12 sets out the issues requiring specialist assessment.
Keep an exit and a correction route
Supplier review should cover access, storage, retention, subcontractors, changes to terms, export, incident support and exit. European hosting or a promise not to train on inputs does not answer every question.
Name who can pause a use, preserve evidence, assess consequences and coordinate corrections. Agree how a substantive correction reaches translated and adapted versions. Record any legally required notifications separately from the editorial response.
Finally, retain a workable alternative when the tool is unavailable or unsuitable. A communication process that nobody can operate without one supplier is a dependency leadership should see clearly.
For every significant handover, someone should be able to answer: what may change, what must not change and who accepts responsibility for the result?
Is AI strengthening your association?
A leadership self-assessment of value, representation and responsibility.
You may already use AI every day. You may have approved a policy, delegated the choices or deliberately limited its use. The useful question is not how much AI your association uses. It is whether you can show what it improves and how responsibility works in practice.
Use the following questions to examine communication, representation and services to members. This is a discussion tool, not a ranking of associations, a technical audit or a legal compliance test. Limited use can be a sound choice. A written policy alone is not evidence that a practice works.
Start with what actually happens
The secretary-general and head of communication can answer separately, then compare. Involve the colleagues and partners doing the work where their evidence is needed. Record the significant uses and recent examples you reviewed; a successful example does not establish that every workflow is reliable.
Choose one answer for each question
Yes. We can point to a current example or record showing this works.
Partly. Some elements work, but coverage, application or evidence is incomplete.
No. The practice is absent or does not work for the use examined.
Unknown. We cannot yet establish what happens or who is responsible.
Not applicable. The question does not fit the use examined; record why.
For each answer, note the question number beside the evidence or gap. A "Yes" covers every element asked about; for "Partly", state what works and what is missing. Refer to records without copying confidential information. Explain deliberate non-use; still examine external AI representation.
Your current practice: dimensions 1 and 2
Tick one answer. For "Partly", note the question number and the element that is missing.
Dimension 1: uses and responsibility
1. Can we identify where AI influences our communications and member services, including embedded software and work undertaken by partners?
2. For each significant use, can we name the owner and the person authorised to approve or stop its actions?
3. Have we checked that the system's actual access and permissions match the actions leadership has approved?
Look for: A current use list, named owners and a permissions check.
Dimension 2: value for members
4. Can we show a concrete improvement that members experience, beyond producing more content or drafting it faster?
5. Have intended users checked whether the AI-assisted service is understandable, accessible and useful in the languages and formats they need?
6. Have we compared the use with a simpler alternative against the same member need?
Look for: Member feedback, an observed service improvement and the alternative considered.
Your current practice: dimensions 3 and 4
Tick one answer. For "Partly", note the question number and the element that is missing.
Dimension 3: mandate and participation
7. For our last substantive AI-assisted communication, can we retrieve the supporting source, the status of the position and the final approver?
8. Have we checked that adaptations and translations preserve who is speaking, the agreed position and its essential qualifications?
9. When AI helps summarise member views, can we trace important reservations and unresolved objections to the originals and correct misrepresentation?
Look for: An approved communication, a checked language version or a consultation trace.
Dimension 4: work and expertise
10. Do we measure the full route to an approved result, including preparation, review, correction, maintenance and relevant costs?
11. Can we show how any released capacity is being used, rather than merely assuming that time has been saved?
Look for: A complete workflow review, a capacity decision and an example of professional judgement.
Your current practice: dimensions 5 and 6
Tick one answer. For "Partly", note the question number and the element that is missing.
Dimension 5: information and partners
13. Can staff and partners show which services and information are permitted for the work they undertake?
14. Have we checked the relevant supplier arrangements for access, retention, reuse and exit, rather than relying only on a no-training promise?
15. Can an error be contained and corrected across versions and partners, with named responsibilities and a workable alternative to the tool?
Look for: A working brief, checked supplier terms and an incident or continuity procedure.
Dimension 6: trust and public presence
16. Have we sampled how external AI services describe our identity, mandate and current positions in relevant languages?
17. Does our public record clearly distinguish current positions from superseded material and individual members' views?
18. Where we use generated images or voices, do we check permissions and avoid misleading audiences about real people, events or support?
Look for: A recorded AI answer, a clearly labelled position and a synthetic-media check where relevant.
Read the pattern, not a total score
Review each dimension separately. Strong answers cannot cancel a serious mandate error or uncontrolled disclosure. Investigate "Unknown" answers and justify "Not applicable". Compare different answers against actual work rather than negotiating a score.
If an affected claim, data use or action cannot be justified, pause it or use an authorised alternative while the issue is checked.
AMEC's Barcelona Principles connect measurement with objectives, stakeholders and outcomes. Distinguish a service improvement from additional output. [18]
What remains out of reach?
Which important member need remains poorly served, and what role, if any, could AI play in meeting it?
Consider a barrier to understanding, participation or access to expertise. Check the need with the members concerned and compare an AI-assisted option with a simpler alternative. "No useful AI role" remains a valid conclusion.
Turn the findings into three decisions
Choose one practice to preserve, one priority to resolve and one opportunity to explore, drawing on the open question above. Do not invent a finding to fill a category. Record the owner, evidence required and a suitable review date.
| Decision | What changes and how we will know | Owner and review |
|---|---|---|
| Preserve | ||
| Resolve | ||
| Explore |
Use the guide on the next page to choose a starting point. Review again when significant uses, positions, supplier arrangements or permissions change.
Match your findings to the next step
Follow the route relevant to your findings. The tools help organise action; they do not replace specialist assessment where it is needed.
| Questions | Start here | Use it to |
|---|---|---|
| 1-3 | Tool 4: responsibility handover Section 8: permissions | List significant uses, name owners and check actual access with the person responsible for the system. |
| 4-6 | Tool 1: member-value brief | Clarify the benefit, seek member feedback and compare a simpler option. Use it also for your open opportunity question. |
| 7-9 | Tool 2: mandate check Section 3: consultation | Verify claims and translations; trace reservations and give members a way to correct their representation. |
| 10-12 | Tool 1: member-value brief Section 7: work and expertise | Count the full cost, decide where capacity goes and protect the time, skills and authority needed for judgement. |
| 13-15 | Tool 4: responsibility handover Section 12: legal decisions | Clarify permitted data, supplier terms, incident handling and continuity. Involve qualified specialists where required. |
| 16-18 | Tool 3: representation scorecard Section 4: synthetic media Section 12: legal decisions | Check external answers against current positions. Examine permissions and claims made by generated images or voices. |
Confirm any improvement against a real example from the work, not simply the completion of a template.
The voice is still yours
The opening sentence involved only a few words. Those words carried the distance between a discussion and a mandate.
The same responsibility connects member value, participation and public credibility. Your assessment identifies where to strengthen it; the tools help turn that choice into practice.
The choice is not between speed and responsibility. It is whether speed serves a purpose worth pursuing, within a process capable of protecting it.
At #inextremis, we connect brand strategy, editorial direction and communication governance around that purpose. The starting point is what makes your association worth listening to, joining and working with. Technology should help more people experience that value without weakening the judgement behind it.
Practical resources and reference
Use these tools to address your findings, then consult the reference sections as needed.
Four tools for your next leadership meeting
These are proposed working tools, not validated scoring systems or legal compliance tests. Adapt them to your governance and keep the administration proportionate.
Tool 1. The member-value brief
Use this to review an existing use or plan a new one.
| Decision | Complete for this use |
|---|---|
| Member benefit | This use will help [specific audience] to [observable improvement]. |
| Baseline and alternative | Today the task takes [full process time] and causes [problem]. Compare AI with [simpler improvement]. |
| Boundaries | Approved sources: [collection and owner]. Excluded data: [categories]. Permitted actions: [limits]. |
| Accountability | Use owner: [name]. Substantive reviewer: [name/role]. Publication authority: [name/body]. |
| Decision rule | Continue if [quality, service and cost criteria]. Pause or stop if [unacceptable error or burden]. |
| Capacity and upkeep | Reinvest time in [activity]. Review on [date] and when sources, permissions or supplier terms change. |
A strong purpose names the member benefit: "Help a new national member identify the requested action in a policy briefing." "Produce more summaries" describes output, not value.
Tool 2. The mandate check
Use this for a substantive claim and repeat it after significant adaptation or translation.
| Check | Evidence to record | A reason to hold the claim |
|---|---|---|
| Evidence | The supporting passage, its date and any uncertainty. | The source does not support the wording or is no longer current. |
| Authority | The person or body entitled to approve this specific statement. | A draft, discussion or expert opinion is presented as an adopted position. |
| Representation | The member, group, association or wider constituency represented. | A narrower view is attributed to a larger group without a mandate. |
Record the final wording, its source passage, position status, approving person or body, approval date and any required review date. Add a route for questions or correction. If a required check is unresolved, hold the affected claim or use wording already authorised; do not convert uncertainty into approval.
Tool 3. The representation scorecard
Sample external AI answers about your association and its priority subjects. Keep the same core questions, record changes and review relevant languages.
For each observation, save the exact question, service and available model or mode, date, language, answer and cited sources. Record visibility, accuracy and usefulness separately.
| Dimension | Record | Follow-up |
|---|---|---|
| Visibility | Present, absent or not relevant to the question. | Check whether relevant authoritative material is publicly accessible. |
| Accuracy | Accurate; minor defect; significant mandate, identity or position error. | Compare with the approved record. Assign an owner to consequential discrepancies. |
| Usefulness | Can the reader verify the answer and reach the appropriate next step? | Check source relevance, links and the clarity of the destination. |
Do not combine these into a reassuring average. An invented institutional commitment is a significant error even when visibility and linking are strong. Log the evidence, responsible owner, correction attempted and follow-up observation. A sample identifies issues; it does not measure every answer every user receives.
Tool 4. The responsibility handover
Attach this information to work moving between the secretariat, agencies, freelancers, translators and national members.
| Carry forward | What the next contributor needs |
|---|---|
| Tools and information | Permitted services and settings; allowed data; prohibited uploads; any rights restrictions. |
| Meaning and status | Approved sources and version; speaker and mandate; claims, qualifications and terminology to preserve. |
| Adaptation and approval | What can change; what needs renewed review; who approves the final language and publication. |
| Disclosure and incidents | When material AI use must be declared; who receives an error report; who can pause or correct the work. |
| Continuity | Source owner; review trigger; access and export arrangements; how work continues without the tool. |
A short completed handover is more useful than a long policy that does not accompany the work. Apply extra detail where the consequences justify it.
An association landscape, not an adoption race
European associations approach AI through different missions. DIGITALEUROPE's public interventions emphasise innovation, implementation and competitiveness. BEUC examines consumer protection. The European Writers' Council focuses on authors' rights and the conditions under which creative work is used. Their positions help explain priorities, not how extensively their secretariats use AI. [19-21]
A public call for safeguards can coexist with active internal experimentation. Support for innovation can coexist with strict limits on confidential data. Neither position is a reliable proxy for operational maturity.
Concrete initiatives show possible directions. On 1 October 2026, the European Disability Forum announced 18 projects in its Google.org-funded Empowered by AI programme. The European Union of the Deaf plans staff capability development and an internal AI policy. Inclusion Europe will explore AI-assisted easy-to-read information while involving people with intellectual disabilities in design and validation. These are announced activities, not measured results. [22]
The useful comparison is therefore with organisations facing similar member needs, language demands and review constraints. A scientific society, a patient organisation and a trade federation do not have identical consequences to manage when information is wrong.
Do not copy a visible AI project to demonstrate progress. Identify a service improvement your members would recognise.
Legal decisions, use by use
Describe the actual activity before trying to classify its obligations.
This section is an orientation for decisions, not legal advice or a complete compliance assessment. Requirements depend on the system, the organisation's role, the data, the audience and the contractual arrangements. Check the official text and obtain appropriate advice for the proposed use.
Role, classification and literacy
An association using a third-party assistant will often be a deployer. Developing or offering a system under its own name may involve different responsibilities. A politically sensitive briefing is not automatically a high-risk AI system under the AI Act; a recruitment application may require a different assessment under the defined categories and conditions. [23]
The amended Article 4 requires providers and deployers to take measures supporting AI literacy among relevant staff and people acting on their behalf. It does not require a guaranteed level for every individual. Training should reflect real tasks, such as checking a summary, recognising an unsupported claim and knowing which information must not enter a tool. [24]
Transparency is not one rule
Article 50 distinguishes obligations. Providers of systems intended to interact directly with people generally need to make the AI interaction apparent unless it is obvious in context. Providers generating synthetic content face technical marking requirements, subject to specified exceptions. [25]
For deployers, disclosure requirements cover deepfakes and AI-generated or manipulated text published to inform the public on matters of public interest. The text-related exception concerns human review or editorial control together with editorial responsibility held by a natural or legal person. It is not a blanket exemption from deepfake disclosure. Your editorial policy may require more transparency than the legal minimum. [25]
Personal data, rights and commercial sensitivity
For personal data, establish the purpose, lawful basis, information provided to individuals, access, retention and contractual safeguards. Examine international transfers and whether an impact assessment is required. A no-training promise does not settle storage, access or deletion. CNIL provides practical guidance on these questions. [27]
Assess intellectual property separately. Access to a text, image or recording is not permission for every subsequent use. Examine source rights, output use and agreements concerning identifiable people and voices. [28]
For associations bringing together competitors, future pricing intentions, individual strategies and detailed commercial forecasts may be sensitive even without personal data. A shared assistant can raise questions about indirect information exchange. The Commission's horizontal cooperation guidelines address exchanges through associations, platforms and algorithms. Aggregation is not a universal exemption: assess information, market context, access and purpose before combining or disclosing it. [29]
Campaign context and application dates
Certain communications intended to influence legislative or regulatory processes can fall within the EU political advertising framework, depending on the activity, services, definitions and exceptions. Assess a relevant campaign before commissioning or distributing it; do not assume that all association communication is in scope. [30]
At the reference date of 13 October 2026, Article 113 retains general application of the AI Act from 2 August 2026, with differentiated dates. Specified provisions for Annex III high-risk systems apply from 2 December 2027, and specified provisions for Annex I product-related high-risk systems from 2 August 2028. Other provisions have their own dates. This is not a general postponement of every obligation. [26]
A useful approval record describes the specific use and the assessments it required. "We use AI" is not precise enough.
What institutional practice can teach you
The EU combines regulation and oversight with support for AI development and adoption. Governance involves the Commission's AI Office, national authorities and coordinating or advisory structures. The AI Continent Action Plan and Apply AI Strategy concern capacity and adoption; participation in any programme depends on its terms and eligibility. [31-33]
Other institutions offer useful methods, not additional civilian legal obligations. NATO's revised AI strategy emphasises responsibility, traceability, reliability, governability and bias mitigation, alongside workforce capabilities and evaluation. For an association, the practical questions are whether a use can be tested realistically, whether people understand its limits and whether someone can intervene. [34]
French initiatives connect capability, evaluation and security. AMIAD focuses on defence applications; the Defence Ethics Committee addresses responsible and proportionate use. INESIA brings together capabilities involving ANSSI, Inria, the Laboratoire national de métrologie et d'essais and the Pôle d'expertise de la régulation numérique. Their contexts differ from association communication. [35-37]
The transferable discipline is modest and useful: test your own documents and languages, include difficult cases, examine permissions, record limitations and retain an alternative. Do not present institutional inspiration as certification or compliance.
Evidence, method and further reading
The argument in this publication combines institutional sources, research and documented association initiatives with #inextremis's editorial analysis. The self-assessment, mandate check, working tools and strategic conclusions are proposed applications to association communication, not independently validated research instruments.
The federation, consultation, speaker and external-answer scenarios are fictional. They illustrate decisions and consequences; they do not report client experience, measured returns or the frequency of an incident.
Association statements illustrate priorities, not a representative level of internal adoption. Announced projects are distinguished from demonstrated results. Findings from professional writing, customer support and creative tasks remain bounded by the settings studied.
EU legislation and relevant national law establish legal requirements. NATO strategies, French institutional practices and research perspectives contribute methods and questions; they do not become binding standards for civilian associations through their inclusion here.
The numbered references link to the material supporting the discussion. For a broader perspective, Bennett and Livingston's The Disinformation Age examines the institutional context of disinformation, substantially in the United States. Read it alongside O'Neill on trust, Macnamara on listening and the different accounts of technological change offered by Mollick, Acemoglu and Johnson, and Crawford. [1, 2, 4, 13, 16, 38]
Legal, institutional and product information requires review when this publication is updated. Reference date: 13 October 2026. Research, restructuring and drafting were assisted by AI. Final editorial approval remains the responsibility of #inextremis.
Sources
1. Onora O'Neill. Trust and Transparency, Reith Lectures. 2002.
2. Ethan Mollick. Co-Intelligence: Living and Working with AI. 2024.
3. Dominique Cardon / Sciences Po. La vie rêvée des algorithmes.
4. Jim Macnamara. Creating an 'Architecture of Listening' in Organizations. 2015.
5. W3C Web Accessibility Initiative. Captions/Subtitles: Making Audio and Video Media Accessible.
6. Claire Wardle and Hossein Derakhshan. Information Disorder: Toward an Interdisciplinary Framework for Research and Policy Making. 2017.
7. VIGINUM / SGDSN. Generative AI and the Information Threat. 2025.
8. Pranjal Aggarwal and colleagues. GEO: Generative Engine Optimization. KDD 2024; preprint 2023.
9. Google Search Central. Google's Guide to Optimizing for Generative AI Features on Google Search.
10. Anil R. Doshi and Oliver P. Hauser. Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content. 2024.
11. Shakked Noy and Whitney Zhang. Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. Science, 2023.
12. Erik Brynjolfsson, Danielle Li and Lindsey Raymond. Generative AI at Work. The Quarterly Journal of Economics, 2025.
13. Daron Acemoglu and Simon Johnson. Power and Progress. 2023.
14. International Labour Organization. Generative AI and Jobs: A 2025 Update. 2025.
15. Hao-Ping Lee and colleagues. The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. CHI 2025.
16. Kate Crawford. Atlas of AI. 2021.
17. ANSSI. Recommandations de sécurité pour un système d'IA générative. 2024.
18. AMEC. Barcelona Principles 4.0. 2025.
19. DIGITALEUROPE. AI Omnibus: A Necessary Pause to Enable Real Simplification. 16 February 2026.
20. BEUC. EU Adopts the AI Omnibus: What It Means for Consumers. 29 June 2026.
21. European Writers' Council. Contribution on AI and the Cultural and Creative Sectors, 2025/2180(INI). 2026.
22. European Disability Forum. Empowered by AI: 18 New AI Projects Begin Across Europe. 1 October 2026.
23. European Union. AI Act, Consolidated Text: Roles and Classification. consolidation of 27 July 2026.
24. European Commission, AI Act Service Desk. AI Act, Article 4: AI Literacy.
25. European Commission, AI Act Service Desk. AI Act, Article 50: Transparency Obligations.
26. European Commission, AI Act Service Desk. AI Act, Article 113: Entry Into Force and Application.
27. CNIL. Les questions-réponses de la CNIL sur l'utilisation d'un système d'IA générative. 2024.
28. EUIPO. The Development of Generative Artificial Intelligence From a Copyright Perspective. 2025.
29. European Commission. Guidelines on Horizontal Co-operation Agreements, Chapter 6: Information Exchange. 2023/C 259/01.
30. European Union. Regulation (EU) 2024/900 on the Transparency and Targeting of Political Advertising. 2024.
31. European Commission. AI Act Governance and Enforcement.
32. European Commission. AI Continent Action Plan. 2025.
33. European Commission. Apply AI Strategy.
34. NATO. Summary of NATO's Revised Artificial Intelligence Strategy. 10 July 2024.
35. French Ministry of the Armed Forces. IA de défense : inauguration du pôle recherche de l'Amiad. 5 March 2025.
36. French Ministry of the Armed Forces. Comité d'éthique de la défense : pour un usage de l'IA éthique et proportionné. 12 February 2025.
37. SGDSN. Évaluer et sécuriser l'IA: INESIA.
38. W. Lance Bennett and Steven Livingston. The Disinformation Age. 2020.