How Do You Govern Machines That Can Learn? Policymakers Are Trying to Figure That Out

Good to see Hal Abelson featured so prominently in this article. I had the pleasure of meeting and talking with him back in the ‘80s. His class is in good hands.

You can’t govern machines that can teach themselves. And they will learn at an exponential rate. Without morals or ethics. Logically, the machines can find many reasons to terminate the human race. We’re destroying earth’s resources and climate for example. It’s a very dangerous and scary world to live in already. I can’t imagine the threat of AI on the human race.

With all respect, there is no evidence to support the claim that AI is only inaccurate in face recognition and crime prediction because "..data is the problem. The results were biased because the data that went into them was biased". These bias problems have been known for years and there is still no fix. More data will not fix the problem. This is because the problem is in the widely used deep learning and machine learning technologies - these widely used AI technologies are not that intelligent. If you do not believe me, try to get your Google Android-based phone to answer a "how to" question the next time that you really need help.

Whatever rules or limitations Western countries might try to impose on AI research or implementation will not be respected by hostile powers like China and Russia. Their AI research will not be bounded by any rules. So we must consider that if we limit ourselves too much, we would likely be committing unilateral AI disarmament. We need to allow research to the fullest depth of AI capability in order not to put ourselves at a potentially serious military disadvantage to those countries.

Whoever owns/is responsible for the Intellectual Property must be held responsible. no different than a medical device or normal (non-intelligent) software

There are a large number of current University research grants for "the hardening of machine learning". Making it more difficult to "hack" ML. One term of art used is, adversarial machine learning. Its purpose is to allow rapid crafting and analysis of attacks and defense methods for machine learning models. IBM has contributed some freeware tools here: https://github.com/IBM/adversarial-robustness-toolbox. A more theoretical paper on the general problem here: https://arxiv.org/abs/1712.04248 In all my systems and network projects since 1986, cyber security has always been the #1 or #2 design goal, but it was always gone to the bottom of my clients priorities because of high cost and the lack of perceived benefit. We are much further ahead on securing AI and ML technology than any US business is willing to pay for.

Need to read Nick Bostrom's SUPER INTELLIGENCE. Just the thought process of programming 'morality', or even 'preserving humanity' in an AI arena is quite daunting. Fortunately we are nowhere near the 'singularity' (Kurzweil predicted 2025 -- not going to happen). If the General AI singularity ever does occur (replication of singular human consciousness in a machine), humanity will be eaten for lunch...

I make no claims as to profound understanding of AI potential but my information from various sources as a layman seems to indicate the processes produce results in a somewhat mysterious way and those results are frequently better than humans can produce. Since they are mysterious why is there an assumption they can be safe and controllable by humans who have a consistent history of creating terrible disasters like nuclear weapons and global warming and industrial destruction of the planet and an internet that is rife with a large variety of innovative criminality? Why is it assumed that human control is effective in making it useful rather than disastrous?

The impact of AI on our socioeconomic system is one of the immediate tsunamis that face our species - climate change being the other. AI itself is in a state of transition whose outcomes may ultimately prove to be threatening. The futurist Ray Kurzweil provides a scenario wherein “evolution” leads to a “Singularity” in which “technology” is indistinguishable from human intelligence and in fact exceeds that of its organic masters. It is defined as a “Superintelligence”. We have just begun to address the implications of such developments. While a number of experts have addressed these issues, Nick Bostrom has provided a framework in which the discussion should proceed (“Superintelligence: Paths, Dangers, Strategies”). These depend broadly on three mechanisms: limit data acquisition; imbue the Superintelligence with “worthwhile” objectives; maintain oversight. However, for each mechanism he also provides potential countermeasures by the Superintelligence with dystopian results for the species. The possible emergence of such a “machine” should also be a guide for our approaches to AI which, as it stands, is principally an “imitative” technology at the moment. Gordon Silverman, PhD Professor Emeritus, Electrical & Computer Engineering Manhattan College

How? the simplest way - pull the plug! Simplistic as that, turn the machine off and that should stop it from doing things you do not want it to. It is not a person, it has not 'rights'. So kill it and be done.

The underlying issue with AI is that it is a black box. Data in, recommendations out. How one led to the other is opaque. The learning of the machine is based upon a “reward” system, but whether the rewards are wisely weighed is uncertain. Nothing much in the way of regulation can inform how such black boxes should be constructed. However, much can be said about how carefully the recommendations are implemented. One item: humans should be in the loop.

1. To be clear, at this point, policy or regulation would not apply to A.I. per se. It would apply to the people developing it. One part of policy should focus on how to hold those people responsible for whatever happens. 2. If A.I. has as much potential as the experts say, for good or ill, we should be exposing it to "data" from the entire breadth and depth of human experience, and possibly other forms of experience. It should not be shackled to the narrow pragmatics of mere "business and innovation." Those are just the buzz words of the moment, from the perspective of some groups of people, in some countries. Instead, A.I.should be exposed to all the forces that make life worth living, to the tides of history, to the depth of compassion, to the sublime, to profound suffering, to the ethics and sentiments of death. If all you feed to the A.I. is the competitive tools of capitalism, you will predictably get an entity that can out compete anything on earth without remorse. 3. Is it too early to start talking about rights for A.I. entities? That could be a way to establish the precedent of compassion for consciousness that we might want such entities to embrace.

Here's a simple recipe: 1. Take a ton of data – structured and unstructured 2. Clean it, making sure obvious biases are removed 3. Keep it aside as your initial training data set 4. Review the algo and modify logic that might introduce biases 5. Feed the training data from step 3 into the model and review the results 6. If the results are showing unintended bias, modify the algo and incorporate more data to make the training set more balanced 7. Repeat The most important underlying need is to have a set of standards that define various types and degrees of bias. If I don’t know why/what/how much I have to fix, I’ll use varying degrees of subjectivity which is a problem. The next most important thing is to introduce human intervention at key stages to review and correct biases.

Why is nobody discussing the near certainty that AI will either wipe out the human race, or change it beyond all recognition? Gorillas encountered a more intelligent race, how well did that go for them?

"The results were biased because the data that went into them was biased-" The most important comment in the article. Why not an algo that checks for, and identifies, bias, of any type. Once identified, then the humans take over and discuss, adjust and then run through the bias algo again until the bias has been removed. Of course, there will be times when bias is wanted, e.g., a bias towards good weather over bad. All very complicated, like life.

The first step to governing machines that can learn is simple - deny they exist. So-called "artificial intelligence" isn't dangerous, but the simpleminded conceit of programmers that biological intelligence - including the set of values, of compassion, of responsibility (yes, that's part of it), of the infinitely-variable, ever-changing chemical interactions in the human brain can be condensed into simple binary format. Non-biological learning is impossible - by definition - and as digital convenience aids (DCAs) become more sophisticated, drawing the distinction will become critical. That DCAs are already proving dangerous has never been more evident after the crash of Lion Air Flight JT610, when undertrained human pilots put their faith in an "artifically intelligent" anti-stall system that doomed both them and 187 other passengers who had entrusted their safety to them.

"Garbage in, garbage out" was a phrase I learned over 50 years ago when I first started to work with computer software. Neural networks (which is what most of AI is currently) "learn" the embedded racism and misogyny of our society when given data that is a reflection of hatred. There are some rues that should be implemented curb the worst aspects of evolving AI. 1. The input to and goals of a neural network should be reviewed by a independent organization, similar to the Institutional Review Boards in place at universities. 2. Neural networks should not be able to search the internet for data. Neural networks should be isolated in a variant of Sensitive Compartmented Information Facility (SCIF) with no outside connections. 3. The output of a neural network should not be transmitted electronically through the internet. It should only be readable in a SCIF. Will this reduce the utility of neural networks? Yes. It also prevents harm. Unfortunately, there will be some billionaire or corporation or nation who buys a ship, loads it up with computer and communication technology and will break all of those rules. If climate change does not get us, the Singularity will.

The extent to which we are able to harness the extraordinary potential of AI for the benefit of humanity will determine whether the decade of the 2020s will be the best in the history of humanity or the worst. Right now we are on a trajectory for the latter, where the massive job destruction of AI and exponential growth in inequality it will fuel will be the tipping point to chaos on a global scale that will amplify current geopolitical tensions by orders of magnitude. This scenario makes comparisons to the 1930s rather quaint. Silicon Valley has failed us, rather spectacularly. It has already been "iterating" in re-architecting civilization, capitalism, and democracy, without the knowledge or consent of the rest of humanity. All to sell advertisements. In doing so it made a mockery of "listening to the voice of the customer". Either Silicon Valley and places like MIT execute the most consequential "pivot" in history, or history will be vicious in the judgement it renders.

It seems that the expansion of Technology constantly outpaces peoples' ability to see and deal with dangerous ramifications. As the AI machines 'learn' - and I don't care if it's how humans learn or not - perhaps they need to learn a system of ethics that will prevent them from making disastrous decisions.

Artificial Intelligence is a form of capitol investment. Nothing more, nothing less. Who owns it? They're responsible. "Too big to fail" doesn't apply. But its not "too smart to fail" either. And failures will happen.

Surely, there should be a distinction here between teaching and doing. A.I. recommending a conclusion and a course of action is a lot different from A.I. taking that course of action on its own.

These machines aren't "learning" in the sense that you and I think. They're collecting vast amounts of data, running statistical analysis on it, and spitting out "answers" that are just the most statistically probable outcomes. What they learn, how they collect the data and where, and the list of possible outcomes is all controlled by human beings. In other words, AI (even in the foreseeable future) that has been programmed to analyze traffic patterns will not suddenly learn about the varying levels of heat involved in the consumption of spicy foods unless we tell it to do so. So when it's asked "How do we govern machines," you don't...nor do you have to. You use regular old laws that outline how society is to act, then you hold scientists to those laws.

In the aviation industry, we say "FAA rules are written in blood" because many people died or suffered greatly in tragic aviation events that led to changes in aviation rules. I am deeply concerned that the artificial intelligence industry is going down the same route. Like drones and social media, it is way too much of an unregulated wild west in which many people will suffer as society learns to deal with them.

The other challenge is how to collect data, either inert data like traffic data or active data that humans generate through their activities. Will I have the option to not allow others to use my data, i.e. Google, Amazon or Facebook for their own personal gain? Can I allow some of my data to be used or all of it? Can I receive compensation for my data? What if a company illegally uses my data. Can I sue them? What's the "value" of my data compared to someone in Kansas City Kansas? How do companies or organizations validate the data they receive? How data does an A.I. need in order to sustain itself and be effective? Just some thoughts.

There is a very profound legal problem to be solved here: liability. When something goes wrong, who do you sue? Who "pays"? That's the fundamental mechanism by which we try to compel ethical, moral, and inventive technical behavior in our society. The few accidents involving self-driving vehicles have resulted in complex legal cases. Tesla and others claim "the driver is ultimately responsible", but there is "shared liability" (man and machine) in most cases. Accidents happen because "systems" fail. They are "multi-factor", i.e. designers and engineers can't anticipate every mode of use. Now comes along machines that learn, adapt, and change their behavior.