The guidance note, Client data, AI and the boundaries of consent, states that while the market research sector has historically operated on a “clear assumption” that client data belongs exclusively to the client, that assumption is being tested in the context of data being used to train AI models and processes.

The guidance says: “The historical default that client data belongs to the client remains a sound starting point. What has changed is that it can no longer be left unsaid. This is not necessarily good or bad, but it needs to be acknowledged and dealt with.”

To address concerns about client data being used in the training of AI, the guidance suggests that contracts should plainly state that the data produced during a research project is owned by the client to avoid any confusion. 

Agencies asked by intermediary platforms to agree to a particular use of project information, such as for AI training, should clarify whether participants, clients and any upstream partners, panel providers, recruiters or sub-processors have been asked to provide consent for their data to be used in this way.

Anonymisation, while useful for addressing issues with personal data, does not address issues with a client’s intellectual property rights, according to the guidance.

“A de-identified dataset, an aggregated summary, or a synthetic output can still carry the client’s strategic thinking and the answers they paid to discover,” the guidance explains.

“If those outputs feed a model that is then available to unrelated third parties, the client may reasonably feel they have funded a competitor’s advantage.”

The guidance adds: “Demonstrable and robust anonymisation is that which reduces the likelihood of identifiability to a remote level, and this needs to be applied at every stage of the data lifecycle.”

Other areas covered include scrutinising what tools, platforms and AI vendors are contractually entitled to do with the data they receive, and making disclosure of the platforms used by an agency the norm when agreeing contracts with clients.

Internal AI frameworks should cover what data may be entered into which tools, who approves harder cases, when clients should be told and how decisions are recorded, the guidance notes.

Additionally, training should use real examples from research practice rather than “abstract principles”, and minimum controls should be put in place to increase efficiency when using such frameworks, such as approved tools lists, data classification, approval/escalation routes, data retention and deletion requirements.

Qualitative data, such as focus group recordings, depth-interview transcripts, ethnographic notes and community board content, could be just as exposed to the issue of AI training as survey data, according to the guidance.

Debrah Harding, managing director at MRS, said that the research sector was an early adopter of large language models, and that knowledge and practices have to evolve alongside changes in AI technology. 

“This guidance is about ensuring expectations are aligned across clients and suppliers, and building and maintaining trust and transparency. That relies on good communication,” Harding said.

“Practitioners should approach this as part of a broader conversation with clients about how work is designed and delivered, including being clear about methodologies, technologies used, and steps that have been taken to safeguard data quality and integrity. The sooner practitioners have these discussions with clients, the better. Stronger partnerships are built on mutual expectations, and rigorous reporting which proves these expectations have been met and therefore builds client confidence.”

What ramifications will this guidance have for AI adoption?

Harding said: “The aim of this guidance isn’t to stifle innovation, but to support responsible, confident adoption of AI. That means being clear with clients about what AI tools have been used in their project, at what stage and for what purpose. Where they have been used, human oversight and robust quality controls are, of course, non-negotiable.

“Greater transparency and consistency can only be a positive influence on our ability to use AI to its full potential, building knowledge and confidence across the sector, and maintaining the high standards and innovative mindsets on which our profession prides itself.”

Ray Poynter, chair of the Esomar Professional Standards Committee, said that the question of data ownership was believed to have been a settled issue, but that using it to train AI is a new use and raises additional questions.

Poynter said: “The old saying that good fences make good neighbours comes into mind. Contracts need to be clear about who can do what with the data, and rights need to be dealt with from the initial data collection right through to the final uses.”

When asked about the impact the guidance could have on AI adoption, Poynter added: “Companies will have a greater range of models to offer and clients will be free to agree to or to reject the terms offered to them. Contracts will be a little longer, and a little more detailed, but none of this should raise any problems for organisations and will probably result in increased usefulness for AI models.”

Anita Watkins, chief executive at the Insights Association, said: “The industry is rightly focused on this issue.  Being open about AI use and obtaining client consent shows respect for the people behind the data – their trust, expectations, and responsibilities. This transparency builds stronger relationships, fosters collaboration, and helps clients feel confident their information is handled with care and integrity, leading to more strategic, long-term partnerships and better work overall.”

• The guidance is published on the MRS website.