AI assistants: what gets lost when summarising a team’s contributions?
A no-code assistant can bring together a team’s contributions and prepare a decision brief. The precise management question is this: how can we prevent that summary from dropping a decisive piece of information simply because only one person supplied it?
What the literature establishes
Garold Stasser and William Titus (1985, Journal of Personality and Social Psychology) studied group decisions in which some information was known to everyone and other information was held by just one person. Their experiments show that discussion favours shared information at the expense of unshared details that could support a better-informed choice. Bringing contributions together therefore does not guarantee that decisive facts actually become available for decision-making.
Why familiar information gets more attention
Gwen M. Wittenbaum, Andrea P. Hubbell and Cynthia Zuckerman (1999, Journal of Personality and Social Psychology) show that exchanging shared information can reinforce participants’ assessments of one another’s competence. A collective preference for what is already known is therefore not simply a matter of information availability: it also involves social validation. When designing a summarising assistant, this suggests that repetition should not be treated as evidence of a fact’s importance.
The premature conclusion
One might conclude that an assistant should systematically highlight every isolated piece of information. That would confuse rarity, reliability and relevance: something mentioned only once may be decisive, but it may also be wrong or unrelated to the decision. A more defensible design rule is to make uncorroborated information that could change the choice visible, while retaining its source and verification status. (our executive and employee training programmes)
What these findings do not demonstrate
These studies concern human groups, not summaries produced by generative models; they do not demonstrate that an assistant reproduces the same bias. They identify a risk worth testing, not an established effect of AI. Assessing a summary solely for readability or coverage of dominant themes cannot establish whether it preserves decisive facts mentioned only once.
A practical check in Sion
Within SHR’s programme « Créer ses assistants et agents IA sans coder », an exercise in Sion could compare an unrestricted summary with one containing a section headed “Isolated information that could change the decision”. Fictional case files rooted in hydropower, healthcare, viticulture and Alpine tourism would each contain a decisive fact appearing in only one contribution, identified beforehand by a professional in the relevant field. The measure would be the proportion of these facts retained with their source and verification status, without being incorrectly presented as corroborated. Both versions would use the same files and model, with repeated runs to check whether any observed difference persists. To go further: explore the Building AI Assistants and Agents Without Coding training in Sion in the canton of Valais, or browse our executive and employee training programmes in Switzerland.
In pictures: Building AI Assistants and Agents Without Coding in Sion
- building ai assistants and agents without coding
- Sion
- research
- training sion