The possibility of dishonest statements and dishonest behaviour is something we all have to be aware of. A particularly interesting question is how to let computers reason with this. This is less far-fetched than it may seem. After all, the Turing test (one of the basic cornerstones of Artificial Intelligence) is based on a computer successfully convincing someone he’s in fact a human. Apart from that, truly intelligent systems may want to have ways of detecting, dealing with and reasoning about the kind of dishonesty they may encounter during some of their interactions with humans.
In the recently published published paper “a formal account of dishonesty” we give a logical formalisation of different types of dishonesty, including lies, bullshit and half-truths. We study the logical properties of these, and provide a number of maxims that we feel every intelligent system that faces the possibility of dishonesty should adhere with.
On the 13th of February Nava gave a talk at HCID at City University London: Evaluating the Visualisation of Plans in the SAsSY Project (Talk slides)
Abstract: The Scrutable Autonomous Systems (SAsSy) project aims to enable the scrutiny of autonomous systems by allowing agents to generate plans through argument and dialogue, while justifying the purpose of each step within the joint plan. Humans or agents can then critique these plans by suggesting and justifying alternative courses of action as needed, thus driving the planning process. However, this requires the system to first present the plan that has been chosen for execution by the system to the user. Given that these plans are often long and complex, one key challenge of providing transparency to humans regarding the internal workings of an intelligent system is therefore (good) information presentation. One of the challenges has been to decide how to adapt the plan presentation to a user’s role and area of responsibility. This talk addresses a key question of how to address the “goodness” of an adaptation. It describes a layered-evaluation approach, measuring performance in a dual task effectiveness) to complement measures such as time to complete (efficiency), and self-reported measures. This talk describes two case studies evaluating visualisations of a plan that illustrates the value of the approach.