AI adoption: what gets lost in a team summary?

Asking AI to summarise a team’s contributions seems a reasonable way to prepare for a decision. The management question is more specific: how can we check that information held by just one person remains visible when it challenges the prevailing view?

What the literature establishes

Garold Stasser and William Titus (1985, Journal of Personality and Social Psychology) showed, in a group decision-making experiment, that discussion favoured information initially shared by members over information held by only one member. This imbalance can prevent a group from making use of the evidence needed for a better choice. A summary that faithfully reflects the discussion therefore does not guarantee adequate coverage of the knowledge available.

The mechanism that can reinforce selection

Raymond S. Nickerson (1998, Review of General Psychology) describes, in a literature review, the tendency to seek or interpret evidence in ways that favour existing beliefs and hypotheses. Applied to an AI-generated summary, this mechanism suggests a risk on the reader’s side: accepting a text that fits their assessment without looking for missing objections. This is a management extrapolation, not a finding about AI from that publication.

The premature conclusion

One might conclude that automated summaries should be abandoned or every contribution retained in full. These studies justify neither prescription: condensing necessarily involves selection. The issue is to distinguish expendable repetition from an isolated piece of information that could change the decision. (our executive and employee training programmes)

What the evidence does not establish

The information-sharing experiment concerns a controlled task, and the review of confirmation bias does not assess contemporary uses of generative AI. They do not establish that an automated summary omits more objections than a human-written one. Checking only the accuracy of the sentences included nevertheless leaves another question unanswered: which relevant information is no longer in the document?

A practical check in Berne

As part of SHR’s « Conduire l’adoption de l’IA dans son équipe » programme, a team in Berne can compare source contributions with the summary prepared for a decision, using an authorised case and an approved tool. In federal administration, public health, telecoms or precision manufacturing, the information might respectively be a legal reservation, a health warning, a network dependency or a tolerance constraint raised by just one person. Before reading the summary, two reviewers identify isolated pieces of information in the sources that could alter the choice, then measure the proportion whose meaning and implications remain explicitly present in the summary. Retaining the source passages, corresponding summary passages and omissions makes the check auditable; this rate measures preservation of that information, not overall decision quality. To go further: explore the Leading AI Adoption in Your Team training in Bern in the canton of Bern, or browse our executive and employee training programmes in Switzerland.

In pictures: Leading AI Adoption in Your Team in Bern

• leading ai adoption in your team
• Bern
• research
• training bern