IA on the job: presenting risk without steering the choice?

An assistant can turn a technical note into a presentation for senior management without changing the underlying data. For SHR’s “IA on the job” programme, the management question is precise: how can we check that this rewording improves understanding of risk rather than silently changing the choice?

What the literature establishes

Gerd Gigerenzer and Ulrich Hoffrage (1995, Psychological Review) show that, in Bayesian reasoning problems, natural frequencies can make inference easier than a presentation in probabilities. This format preserves counts and subgroups drawn from the same set of cases, rather than requiring readers to combine conditional percentages. The finding concerns a specific information structure: simply adding a few numbers to a text is not enough.

Framing can shift preferences

Amos Tversky and Daniel Kahneman (1981, Science) show that different descriptions of equivalent consequences can alter preferences between risky options. Presenting outcomes as gains or losses therefore does not necessarily leave the choice unchanged. Applied to AI, this mechanism suggests that an instruction such as “make this briefing more reassuring” may affect the decision, not just the style.

The premature conclusion

One might conclude that requiring natural frequencies is enough to produce a neutral presentation. That would confuse an aid to calculation with general protection against framing effects. A briefing can make denominators visible while highlighting only the favourable consequences of an option. (our executive and employee training programmes)

What these studies do not establish

These studies concern neither current generative assistants nor decisions made in businesses in Valais. They do not allow us to quantify the effect of AI rewording in these settings. Above all, an observed frequency, an estimated probability and a hypothetical scenario are not interchangeable: asking an assistant to present them in a common format can conceal their differences.

A practical check in Sion

In an “IA on the job” workshop in Sion, use clearly labelled fictional cases or authorised data covering hydropower maintenance, health screening, vineyard crop losses and cancellations in Alpine tourism. For each case, prepare two briefings containing the same information: a free-form AI rewrite and a checked version specifying the reference population, the period, the status of the figures, and both favourable and unfavourable consequences. Distribute the versions between two groups, then measure the proportion of readers who correctly identify the population concerned and the meaning of the risk; record their chosen option separately. Comparing understanding and choice can reveal a possible presentation effect without treating a different preference as an error. To go further: explore the AI on the Job training in Sion in the canton of Valais, or browse our executive and employee training programmes in Switzerland.

In pictures: AI on the Job in Sion

• ai on the job
• Sion
• research
• training sion