mr. hersh: two points.there is no universal definition of a bad person, and therefore, it's a social construct, and it is laden with people's moral judgments. there is no mathematical way to determine whether someone is a bad person in the abstract, in the sense that there is no universal definition of a bad person. so the second point that i would make, though, is that just because you are using data science and mathematics, you could have the most advanced neural network made yet, but if the underlying training set for that algorithm uses systemically biased information, so for example, if you are looking at criminal justice statistics and you have a model that ends up predicting that this african-american is more likely to commit a crime because more african-americans end up in prison, you create an algorithm which is simply reflecting social and moral biases of the data sets. sen. hirono: i think the predictive value of what the government is pursuing is very questionable and problematic in terms of all

mr. hersh: there is a lot we don't know. it often looks the same. responding to senator kennedy, just because the campaign spends a lot of money on a particular ad kind doesn't mean it works. for a long time campaigns were running money on robo calls. countless experiments have shown they do nothing. in an environment where there is a lot going on in the campaign, a lot of campaign ads don't work. nobody in this room changes their mind as a result of the campaign ads of the 2016 election. for someone who was a director of research for cambridge analytical, they should know the actual effect and estimate of which campaign ads do what. part of my skepticism comes because there has been a -- there has been no presentation of any evidence in environments in which someone shows, here is the effect. this ad targeting neuroticism has no effect. it is hard to sort this out as technology changes from 2004. lots of things are new. one point to come back to is that in a presidential election, when there so much going on, the effect of one kind of ad is usually

mr. wylie: yeah. i don't contest what professor herish -- hersh was saying in the fact is through persuading somebody compared to motivating them to turn out. the data sets that traditionally were used are often very sparse nd not necessarily reliable. senator coons: available through facebook. mr. wylie: yes. that's why cambridge analytica ended up pursuing that path because it found in comparison to traditional marketing data sets, the data that you could procure from social networking sites was much more dense and actually much more reliable to create a precise algorithms yism. senator coons: a billionaire megadonor in support of the trump campaign, fund add shell corporation, cambridge analytica, still run by foreign nationals,er to take advantage of this research, this understanding, and gained access to 87 million americans, facebook information, and developed some of the algorithms that came out of that. i want to ask you in the time remaining about your experience working with steve ban none, one of president trump's senior -- bannon, one of president trump's enior campaign advisors. was vote