Friday, 17 June 2016

Pre-empting Hindsight Bias in Automated Warfare

"His Majesty made you a major because he believed you would know when not to obey his orders." Prince Frederick Karl (cited by Von Moltke)

The killer robot community is debating concepts such as Meaningful Human Control (MHC) and 'appropriate human judgment' with a view to their operationalisation in practical use. For the purpose of this post, the various terms are bundled into the abbreviation MHC.

After things have gone wrong, the challenge for incident analysis is to avoid 'hindsight bias'. To learn from an incident, it is necessary to find out why it made sense at the time. "to reconstruct the evolving mindset", to quote Sidney Dekker. There is a long history of the wrong people getting the blame for an incident - usually some poor soul at the 'sharp end' (Woods).

In a world of highly automated systems, the distinction between 'human' and 'machine' becomes blurred. In most systems, there are a number of human stakeholders to consider, and a through-life perspective is frequently useful.

In a combat situation, 'control' is an aspiration rather than a continuing reality, and losers will have lost 'control' before the battle - e.g. after the opponent has got inside their OODA loop. What is a realistic baseline for MHC in combat? We have to be able to determine this without hindsight bias.
How would an investigator determine the presence or absence of MHC in the reconstruction of an incident? It would be virtue signalling of the lowest order to wait until after an incident and then decide how to determine the presence or absence of MHC.

One aspect of such determination is to de-couple the decision making from outcomes. The classic paper on this topic is '“Either a medal or a corporal”: The effects of success and failure on the evaluation of decision making and decision makers' by Raanan Lipshitz
There is, of course, a sizeable literature on decision quality e.g. Keren and de Bruin.

The game of 'consequences' developed here has been to provide food for thought, and an aid to discussion on what an investigator would need to know to make a determination of MHC.  It comprises short sections of dialogue. The allocation of function to human or machine, and the outcomes, are open to chance variation.
The information required to determine MHC might help in system specification, including the specifics of a 'human window'. It is not always the case that automation provides such a window - especially in the case of Machine Learning. So, how do we determine MHC in a combat situation? Try some of the exercises and see how much you would need to know. If the exercises here don't help make a determination - what would?

Please let me know in comments below, or on Twitter @BrianSJ3

As an aside, there are proven approaches to take in system development that can provide assurance of decision quality. This is not entirely a new challenge to the world of Human-System Integration. "What assurances are there that weapon systems developed can be operated and maintained by the people who must use them?"
[Guidelines for Assessing Whether Human Factors Were Considered in the Weapon Systems Acquisition Process FPCD-82-5, US GAO, 1981]