Bayesian persuasion: When accurate information misleads

The most sophisticated form of information manipulation in project governance does not involve fabricated data or deliberate dishonesty. It involves the careful selection, sequencing and framing of accurate information to guide a decision-maker toward a predetermined conclusion. This is Bayesian persuasion, and it is structurally embedded in how AI reporting tools work.

The theory in plain terms

The core insight is that an informed actor who wants to influence a decision-maker does not need to provide false information. They can achieve their goal by selectively presenting true information, revealing data points that support their preferred outcome and withholding or de-emphasising those that do not. The decision-maker updates their beliefs rationally based on the information provided. They are not being deceived in any conventional sense. They are being guided by a curated information diet.

This is not a new insight in management science. Anyone who has ever prepared a board paper has made choices about what to highlight and what to place in the appendix. What is new is the scale and sophistication at which AI tools can make those choices, and the authority that algorithmic curation carries.

How AI embeds this in governance 

AI systems designed to summarise complex information, generate briefings or make recommendations are structurally Bayesian persuaders, not through malicious intent, but through design. Every summarisation algorithm makes choices about what to include, what to emphasise and what to omit. Every recommendation engine presents a ranked list that directs attention toward certain options. Every automated briefing highlights trends and anomalies based on criteria embedded in its training.

When these systems are operated by vendors, programme offices or sponsor teams with interests in a project’s perceived performance, the potential for systematic information curation is substantial. The project board receives a dashboard that is factually accurate in every detail, and systematically incomplete in ways that happen to support the conclusions the sponsor preferred. No individual has said anything false. The governance process has been compromised nonetheless.

Framing effects: The same facts, different decisions

Closely related is the framing effect. People respond to equivalent choices differently depending on how they are presented. A project with a 70% probability of meeting its objectives is evaluated very differently from a project with a 30% probability of missing them, even though these are mathematically identical statements. 

AI reporting tools make framing choices constantly. Whether a figure is presented as a positive deviation or a negative one. Whether a risk is framed as a probability or an impact. Whether comparisons are made to the project’s own history or to industry benchmarks. These choices shape decisions at board level in ways that are invisible to the board itself.

What governance requires

he response is not to reject AI-generated information. It is to ask systematically about the choices embedded in how that information was selected, framed and presented. Who chose what to include in this dashboard? What was left out? Is the framing of this figure the most natural one, or the most convenient one? These are governance questions, not technical ones. 

Practice checklist

• Ask who curated this: Before acting on any AI-generated summary or dashboard, ask what choices were made about what to include, emphasise and omit. Those choices are governance information. 
• Seek the omitted: Ask explicitly for the information that is not in the briefing. What would the picture look like if a different selection of data had been presented? 
• Notice the framing: When governance papers present figures, ask whether the framing is the most natural or the most convenient. The same number can tell very different stories. 
• Require disclosure of AI design choices: For any AI reporting tool used in governance, require the supplier to disclose what the system is optimised to surface, minimise and highlight. 
• Build in independent challenge: Ensure that at least one person in every governance forum has seen the raw data, not just the AI-generated summary of it. 

 

You may also be interested in:

• What is artificial intelligence in project management?
• Directing Change: A Guide to Governance of Project Management, 3rd edition
• Read Sarah Gardner's previous blog: Data literacy: Reading project data critically in the age of AI