New-Tech Europe | February 2019

do is wait. This means that in some situations, humans and robots will always have to work together, with the robot taking on the heavier work and its human workmate having more time for interaction with other people and knowing how to respond to unexpected situations. AI systems need to be tested regularly As we have already said, the human always needs to understand how a robot arrives at a certain conclusion or action – and must always be able to make adjustments where necessary. Recent examples of prob-lems with artificial systems have demonstrated exactly that. For example, there is the instance of the chatbot, Tay, which began posting racist messages on Twitter after certain other Twitter users left politically incorrect posts. The chatbot had not been given any instruc-tions to recognize these types of statements as being inappropriate. ‘Norman’ also made the news in 2018. Norman is an AI system that displayed psychopathic characteristics when doing a well- known test with Rorschach inkblots. It happened because Norman had previously been shown mainly sensational and violent images from Reddit and he had built up a picture of the world based on those images. MIT re-searchers wanted to use the experiment to demonstrate the danger of ‘false data’ being used as input for AI systems. And finally, there is also the example of the COMPAS algorithm that was used by the judicial system in America to make predictions about the recidivism of convicts. What happened? Based on the historical data used as input for the algorithm, it reached the conclusion that blacks

Fig 2: The cobot developed for Audi can be operated using hand gestures and has a screen in the form of a face to communicate with the worker.

it from there. But in more complex situations, the cobot may always need a human workmate on hand to give it instructions and to instruct it how to do things. One example of this would be collecting waste in a city. It can be a complex business distinguishing what is waste and what isn’t. It’s also hard to know how to react if someone waves to the waste truck driver and then runs up behind with a bag of waste to be picked up. A robot would not know how to respond, whereas a human knows that the friendliest thing to

would like to cross. But this type of ‘unwritten rules’ in human-to- human communication is not easy to transfer to AI systems. The way humans and cobots work together on the workfloor can take the formof the humandemonstrating how something is done and the cobot learning from it so that it can then perform a particular action perfectly. In a single, repetitive process it may be that the human worker will only have to show the robot how to do something a few times and the robot will then take

Fig 3: This is Norman, the ‘psychopathic’ AI algorithm that MIT scientists trained to demonstrate the danger of AI when ‘false’ data is used as input. (copyright “Thunderbrush on Fiverr”, https://www.fiverr.com/thunderbrush). At right is one of the Rorschach inkblot tests that Nor-man was given to look at. In it, he saw a man knocked down and killed by a speeding car, whereas standard AI systems see it as a close-up of a wedding cake on a table.

22 l New-Tech Magazine Europe