IA on the job: does a collective summary preserve decision-critical information?
Asking AI to prepare a summary seems a reasonable way to reduce the burden of collaborative work. The managerial question is more specific: how can a team check that relevant information held by just one person reaches the document on which its decision will be based?
What the literature establishes
Garold Stasser and William Titus (1985, Journal of Personality and Social Psychology) demonstrated experimentally that group discussions can favour information already known to several members over information held by just one. In their study, this incomplete pooling could prevent the group from identifying the option supported by the full set of available information. A substantial discussion therefore does not guarantee that the information needed for the decision has been brought together.
Why shared information receives more attention
Gwen M. Wittenbaum, Anne P. Hubbell and Cynthia Zuckerman (1999, Journal of Personality and Social Psychology) examined a social evaluation mechanism: exchanging shared information can foster mutually positive assessments of participants’ competence. Information that others recognise thus enjoys an advantage that does not necessarily reflect its value for the decision. Applied to AI-assisted summarisation, this mechanism suggests a hypothesis to test: the material supplied to the tool may already be unbalanced before generation begins.
The premature conclusion
One might conclude that it is enough to ask AI to highlight uniquely held information. But an instruction cannot make accessible something that was never supplied to the tool, and information held by just one person is not necessarily accurate or important. The aim is not to give rare information systematic priority, but to subject relevant information to the same scrutiny, regardless of how many people know it. (our executive and employee training programmes)
What the evidence does not establish
These studies examine human exchanges in experimental settings, not contemporary uses of generative AI in organisations. They demonstrate neither that an AI tool amplifies this bias nor that it corrects it. In practice, reviewing only the final summary also makes it impossible to distinguish an omission by the tool from information missing in the initial contributions.
A practical check in Zurich
As part of “IA on the job”, a Zurich team in financial services, insurance, tech or the European headquarters of an international group can test the following protocol on a clearly defined decision: before the group discussion, each participant independently records the facts they consider relevant, using only an authorised environment and data permitted under internal rules. A designated reviewer identifies facts appearing in just one contribution and validates their relevance; the team then checks whether they appear in the discussion, in the material supplied to AI and in the final summary. The measure is the proportion of these relevant facts retained in the summary relative to their initial number; if none are identified, it cannot be calculated. This check locates information losses without treating retention as proof of a better decision. To go further: explore the AI on the Job training in Zurich, or browse our executive and employee training programmes in Switzerland.
In pictures: AI on the Job in Zurich
- ai on the job
- Zurich
- research
- training zurich