会议领导者需主导AI伦理：Clear框架指导合规使用

Skip to content

• View Larger Image

Artificial intelligence (AI) can make business events more accessible, efficient and rewarding. But when assistance turns into monitoring, scoring, or unlimited power, meeting leaders face a question more important than what the technology can do: should they allow it?

The webinar Using AI in Meetings: How to Stay Ethical, Compliant and Human addressed this question. Pia Villför Larsson from Facilitatorhuset in Sweden led the session, focusing on facilitation and governance. Olga Sukhovskaya from Alineo in France joined her and covered measurement, consent, and automation bias. They shared a starting point: responsibility for AI in meetings cannot be delegated to suppliers or legal departments – it must remain with the meeting leader.

The promise is almost irresistible. An AI assistant joins the meeting, records conversations, generates transcripts, identifies action items, and delivers a concise summary before participants leave. For meeting leaders under pressure, this represents progress: reduced administrative tasks, fewer overlooked decisions, and more time for the actual discussion. AI can also improve accessibility with real-time captioning, and support international teams by bridging language and accent differences.

However, AI technology can also change meeting dynamics. A tool designed for note-taking may start tracking speaking time, measuring engagement, categorising participants, or inferring emotions such as positivity or reluctance. Recordings can evolve into behavioural profiles, and a supportive assistant may become an unseen evaluator.

“It is in the transition from assistance to assessment that meeting professionals must tread carefully,” says Pia Villför Larsson.

When metrics become judgements Imagine a summary report presented to a manager after a team meeting: Dave spoke for 38 per cent of the meeting and was labelled ‘very engaged,’ while Julia contributed less and was classified as ‘uninterested’ or ‘negative.’ Although these figures appear objective, their implications are not neutral.

“Speaking time is not a reliable measure of insight, preparation or engagement. Someone who speaks frequently may be helping the group – or simply monopolising the conversation. A quieter participant may be reflecting, communicating in a second language, or following a cultural norm that discourages interruption, while an experienced specialist may need only two sentences to significantly influence the discussion,” says Pia Villför Larsson.

Emotional labels can also be misleading. “I’m worried” may seem hesitant or unsupportive, but it could be the most responsible input. Constructive scepticism is essential in risk management and strategy, and assigning it negative scores risks rewarding overconfidence while penalising caution.

The risk grows when meeting analyses inform performance appraisals, promotion discussions or decisions, or project invitations. A metric designed to support collaboration may become a tool for human resources management. Participants may not be aware of this secondary use and may have little opportunity to challenge the outcome.

“Data generated during a meeting does not remain harmless simply because it is numerical. The more authoritative the dashboard looks, the easier it is to forget that those categories have been designed by somebody,” says Olga Sukhovskaya.

The law sets clear boundaries EU AI legislation now establishes a key framework. Most provisions took effect in August 2026, while prohibitions have been in force since February last year. Prohibited practices include using AI to assess emotions in the workplace or education, except for specific medical or security needs. The law also targets harmful manipulation, exploitation of vulnerabilities, social rating systems, and certain biometric categorisation methods.

However, the AI Act is only one aspect of the legal landscape. Meeting technology may also be subject to the General Data Protection Regulation (GDPR), national labour laws, employee monitoring rules, collective agreements, confidentiality obligations, trade secret protections, and internal company policies.

Recordings, transcripts, and participant profiles may contain personal data. Under the GDPR, organisations must have a lawful basis for processing this data and comply with principles such as transparency, purpose limitation, data minimisation, accuracy, security, and storage limitation. Participants must be informed of the reasons for data collection, intended use, retention period, access rights, and any transfers outside the EU.

“A tool does not meet the requirements simply because it is available on the market,” says Olga Sukhovskaya.

Organisers must know where data is stored, which suppliers process it, whether subcontractors are involved, the default storage settings, and whether customer information may be used to improve models.

Consent is not a universal solution and is only one possible legal basis under the GDPR. In employment relationships, it is often difficult to show that consent is voluntary because of the power imbalance. European data protection guidelines recommend employers rely on employee consent only in rare cases where refusal has no negative consequences.

“Everyone clicked ‘accept’ is not a comprehensive governance strategy,” says Olga Sukhovskaya.

Compliance asks “can we,” ethics asks “should we” Although related, they are distinct. Compliance addresses whether an action is permitted and under what conditions. Ethics considers whether it is fair, proportionate and aligned with the intended human interaction.

For example, a participant may not want an internal discussion recorded. While the organiser may have a legal basis for processing, it might make the participant feel uncomfortable or unable to object in front of colleagues or a senior manager. Formal notice does not eliminate social pressure.

“The result may be a technically documented meeting with a poorer-quality conversation. The organisation gets meeting minutes but loses transparency,” says Pia Villför Larsson.

Other ethical challenges exist. Simply labelling a document as “AI-generated” does not resolve these issues. Transparency is important, but responsible use also requires restraint, human judgement, and the ability to correct the minutes.

• A group discussion is transcribed in real time for the meeting leader to monitor, even though participants believe their conversation is semi-private.
• A healthcare provider or counsellor relies on an automated transcript without verifying its accuracy or ensuring key statements are captured correctly.
• Contributions are rewritten in a polished, professional tone that removes hesitation, humour, and individual voice. The final document appears impressive but no longer reflects the people who created it or the tone of voice in their statements.
• An AI summary is circulated without review and later treated as the agreed plan, even if the system omits a condition, confuses speakers, or misreads small talk as an action point.
• No retention policy exists. Recordings and transcripts are stored indefinitely because storage is easy, and no one has decided when to delete them.

A useful operational response can be summed up in one word: Clear.

A Clear Framework for AI in Meetings Pia Villför Larsson and Olga Sukhovskaya developed The Clear Framework for AI in Meetings in March to guide the ethical and compliant use of AI during meetings. Clear is an acronym and it stands for:

C: Consent and transparency

L: Limit data and storage

E: Equality and inclusion

A: Accountability and human oversight

R: Review, feedback and correction

Consent and transparency “Participants should be aware that AI is being used before the meeting begins. They should understand whether the meeting will be recorded, transcribed, summarised or analysed, and whether the output will be used solely as a memory aid, or for any other purpose,” says Olga Sukhovskaya.

When consent is the legal basis, it must be a genuine choice and as easy to withdraw as to give. In other cases, organisers should clearly explain the legal basis and participants’ rights. Meeting organisers should consult the organisation’s legal or data protection specialists, rather than creating a consent process at the last minute.

Transparency should include personal AI assistants. If participants use their own note-taking bots, this must be disclosed to the group. One disclosure does not justify others using undisclosed recording tools.

Limit data collection and storage Not every meeting requires recording, and few recordings need permanent storage. Begin by defining the purpose, then collect only what is necessary.

A meeting may begin with a routine project update and continue with a sensitive HR discussion. In such cases, it is advisable to stop recording before the HR topic. Alternatively, a verified summary may suffice, allowing you to delete raw audio files and transcripts after an agreed period.

“Data minimisation is not the same as stifling innovation. It is about disciplined meeting planning. Storing less unnecessary information reduces the risk of loss, misuse, or misunderstanding,” says Pia Villför Larsson.

Equity and inclusion AI can broaden participation through subtitles, translation, and accessible formats. However, it may misinterpret accents, overlook cultural differences, or misrepresent communication styles as indicators of engagement. Human facilitation remains essential.

“The meeting leader must be mindful of who has been interrupted, who has not had the opportunity to speak and whose contributions have been distorted in the transcript. Inclusion cannot be delegated to a participation rate,” says Pia Villför Larsson.

Accountability and human oversight AI should support, not silently make, decisions. Each transcript, summary, and action list should have a designated human reviewer before becoming part of the organisation’s records.

The reviewer should confirm that decisions were made, conditions and dissenting views were preserved, speakers were correctly identified, and actions were assigned appropriately. If outcomes may affect individuals, apply an even higher standard of review.

Accountability also requires addressing automation bias, which is the tendency to give undue credibility to machine-generated text because it appears fluent and well-structured. “A polished error remains an error,” says Olga Sukhovskaya.

Review, feedback and correction Participants need a practical way to review how AI was used and to correct its outputs. If a summary will become the official record, those represented should be able to identify omissions and inaccuracies before it is finalised.

The most effective approach is to reserve time at the end of the meeting for a joint review. Display the draft summary, confirm decisions and responsibilities, and invite corrections while the discussion is still fresh. Ten minutes spent validating the minutes can prevent weeks of confusion and foster immediate ownership of the outcome.

Organisations should regularly review their use cases Did improved transcription enhance accessibility? Did it discourage participation? Were summaries accurate enough to justify data processing? Responsible AI requires ongoing governance, not a one-time approval.

Meeting organisers have new responsibilities Pia Villför Larsson and Olga Sukhovskaya recommend that, before each session, organisers should be able to answer four key questions: 

1. What is the purpose of using AI? 
2. Is it necessary? 
3. What personal or sensitive information may be involved? 
4. What rules should govern the tool and its results?

During the meeting, provide participants with a brief, clear reminder. They should know which tools are active, what will happen to the material, and how to raise concerns. If the organisation offers a choice, it must also provide a practical alternative. Asking someone to switch off their microphone or remain silent reduces participation to observation and should not be considered a complete solution.

After the meeting, someone should review the output, correct errors, confirm decisions, and apply the agreed storage policy. Do not use raw data for employee performance reviews, model training, or unrelated analysis without proper disclosure.

A simple rule of thumb might be: Please note that this meeting is being transcribed using AI to generate a draft summary. Only designated recipients will have access to the transcript, and it will be retained for the specified period. If you are using a personal AI assistant, please inform us. AI-generated output will not be considered a decision or official record until reviewed and approved by a responsible person.

While wording may differ by organisation, jurisdiction, or purpose, the underlying principle remains the same.

The future of AI-assisted meetings will depend not only on software capabilities or legal guidance, but also on practical factors such as invitations, chairperson selection, storage settings, and time allocated to review AI-generated summaries.

The key question is no longer whether AI can record meetings, but whether participants understand what the AI is capturing, how it will use the information, and who is accountable for any misinterpretations.

Page load link

Go to Top