Yup, I went there. And I went there intentionally for the simple reason that this is about Predictive Modeling Security. It comes from the dangers of predictive modeling. I suddenly realised that President Trump wasn’t all that bothered with AI (apart from the class actions in play at present), the dangers are in a much more niche setting. You see, the main dangers off predictive modeling are numerous and there is every chance that all these fake AI’s are getting toppled by them.
First we get Algorithmic Bias and Discrimination: Models train on historical data. If past decisions or societal data contain prejudices (optional predictive policing), the model codifies and scales them.This is a round about setting of the errors of historical data is a dangerous slope. It come clear through the absence of clear validation and verification (a setting I told you all about in my past stories going back months, up to a year ago I saw that danger pop up. But when we get away from all the fake AI and we concentrate on Predictive Modeling (what is more and more used) validation and verification becomes a much tougher nut to crack. But where is that and the answer is a speculated everywhere. Because when we ‘kinda’ accept bias in these fake AI settings, the presented concentrated in predictive modeling becomes a lot more dangerous and some make away with jokes. Jokes like “70% of women are medicating, this tells us that 30% doesn’t realise that they need medication” a simple joke that paints over the setting of bias and the dangers of what medication we are talking about. Two generalized settings and it is ‘laughingly accepted’ with the offsetting “It’s just a joke, don’t worry about it” but that is the problem and when we take a less simple setting like credit scores and optional housing scores. The setting intensifies, because now banking references are set in that same policing setting. So when were you last set in a banking dispute and someone told you, this is an obvious error and we will adjust that. But that is the ‘paint over’ setting, but who was thrown into that mix? Australia has a huge housing problem and that is where this setting of bias hurts the most. I reckon the United States as well, but that is speculative, because I have never lived (or ever want to live) in the United States.
Then we get to Self-Fulfilling Prophecies: Accurate prediction models in healthcare or social services can alter human behavior or clinical care paths (e.g., withholding treatment based on a predicted poor outcome). This causes the prediction to become true even if the original forecast was flawed. Which is in a serious setting in any aging population, which by all accounts was most of the world. Apart from the setting that I do not call them ‘accurate prediction models’ the setting is that fake AI is riddles with these passages and it is not just withholding treatment, the setting of treatment that a person is rightfully entitled to is getting shoveled under the carpet in way too many expected cases and the danger here is seen in “This causes the prediction to become true even if the original forecast was flawed” but anyone who knew IT in the 90s knew this setting that not just flawed, it grew the setting of the GIGO law (Garbage In, Garbage Out) as such the fake AI which was flawed became unreliable at best and flawed data never ever becomes a true prediction. It is like doing the right thing for the wrong reason, the reasoning becomes corrupt and at best defeating the purpose. Make sure that you take notice of the ‘at best’ setting, because in this age where validation and verification are nowhere to be found the fake AI becomes a point of hindrance. And Personally I think that this is now becoming an issue and the is why we see people like Sam Altman, Dario Amodei, Elon Musk, Mustafa Suleyman, Sundar Pichai and Demis Hassabis becoming more and more cautious in all this, or at least that is how I see it. They all most have been aware of these settings, but I reckon that these settings are taking rather large settings and they are ‘whisking’ it away in the hallucinations setting of perhaps something like a rogue AI settings, but when you realise that this is all driven by advanced predictive analytics. The stage becomes a dangerous one. We see the joke with the setting ‘Is this mushroom edible’ and we laugh, but the underlying setting is not a joke and it is hammering on our rights in numerous directions.
Then we get a part I am not too familiar with. It is the setting of Runaway Feedback Loops: In areas like policing or resource allocation, sending more personnel to targeted neighborhoods generates more recorded incidents in those specific zones. The model misinterprets this as proof of higher crime rather than a product of increased surveillance. (Source: Gilbert and Tobin) there are two settings in play, proper administration of case identifiers and the second are the sources under the cause of the action. I still think it comes from inadequate validation and verification, as such I could be wrong and the setting could be due to predictive analytics, but I reckon that this interacts with systems like Palantir and other systems. As such the interaction of more than one fake AI. But it could be relevant to this setting.
Then lastly we get two elements and they seem to be connected, but they are not. They are the stages of confusing correlation with actionability and the actions of the premise of what correlates and what is actionable. It feels like Palantir, but it is not merely them, it is mostly seen in hospital records and the healthcare records and optional insurance records, this is what is the likely culprit in what I speculatively call the pretense of cause and effect and whilst insurance is setting the premise of ‘This is not covered’ and the other systems trying to adjust, because the doctor said it was essential. Most countries have a good handle on this, but I am not sure the United States is on that par. And it all interacts with a high risk of methodological flaws: Studies show that many machine learning prediction tools suffer from poor data handling, overfitting, or inadequate sample sizes, leading to a high overall risk of bias when deployed and that is where I am different. I always considered all AI (which I call fake AI) as a joke and I am not taking it seriously, especially after I got word of some acts by said bank. I was never able to get the released goods, but that is the setting we are in. And what I have stated for the longest time “It all comes down to verification and validation” until that happens you have a trillion dollar failure on your hands. But that might merely be me. So when did you get any good answer from any vendor on AI? As a funny reference I am going to the Changeling (1980) with George C Scott, a decent movie and I enjoyed it, but what is Eric Winter doing there? He was born in 1976 and the movie was made in 1979, he would have been 2 years old. He seemingly started acting in 1999. If Google cannot get that part right (IMDB did make the correct references, like the unnamed actor being Erick Vinther) how much faith can you have in any fake AI? A simple setting of validation and verification. Oh, and I called out that mistake close to two years ago.
So have a great day and maske sure that you verify that it is Friday, because you never know who hands you the data.
