Is It Possible for an Autonomous Automated System to Make Ethical Decisions?

Is it possible for an autonomous automated system to make ethical decisions?

Introduction:

*Analyze the question in Topic*

Is it possible for AI to make ethical decisions? It may be possible at some stage in the future but a lot of work is to Pre-empt AI making ethical decisions.

Ethics: Definition

Putting Ethics into a machine: Humans programming set of variables, questions, limits etc. Vs AI having a set of principles and coming up with the conclusion of whether an action is ethical or not.

Are there any examples of ethics in AI out there at the minute? – Testing Driverless cars how were they tested?

(Any AI examples in construction)

Is it possible for autonomous automated systems to make ethical decisions?

This question requires a lot of introverted analysis as well as review and research into the advances in technology happening around the world today. There are so many advances taking place around the world today, it is hard to know where to stop the research and reviews, from books to scholarly articles, to newspaper articles, to TED talks, and podcasts etc.

I feel this question could or should be split in two, is it possible right now? No, I don’t believe an AI is truly ethically motivated at this moment in time. And the second question; will it ever be possible? Yes, I believe an ethically motivated AAS/AI system will evolve into our world at some stage, when or where I could not pinpoint but I do believe it will. A large amount of research and analysis will have to go into practical ethics. We could further break this down into how the autonomous automated system would become ethical? By computer scientists inputting directly our own human ethical standards which are Implicit ethical agent(Anderson and Anderson, 2011) vs inputting a set of ethical principles and having the AA system come up with its own choice Explicit ethical agents(Anderson and Anderson, 2011) through what it does best, machine learning.

We have numerous ethical dilemmas in which we try to figure out the answer to based on specific AA systems, such as the driverless car. If said car cannot stop and has to either go straight on into a 4-year-old child or if it should swerve and hit three 75-year-old persons? Who wins out? Ultimately the choice depends on which lives are worth more to society? Which is to each person a different answer as it depends on their values. If human beings cannot answer this question with a relative agreement amongst a large group of people, who we deem to have “good” ethics, how could we expect a machine too? As AI/AA systems get their learnings from humans, if we do not have the data to input, what happens with the output? A lot of areas in ethics are to be determined prior to the inputting of ethical principles to autonomous automated systems.

For an AA system to become ethical even if as stated above they are acting as implicit ethical agents, ethical boundaries would have to be in place. The universal ethical principles would generally be thought to be the first codes to be inputted such as human lives are to be protected above any damage occurring to the AA Systems. These principles are seen as a given, but what about when we dive further into ethical principles around the world? As we have seen in the USA, some bias may be introduced to these systems, which may have occurred accidentally but nonetheless these biases have occurred and an understanding of the potential outcomes of a bias based system would have to be thought out and potentially overcome, maybe a form of decoding or further coding to remove the bias, Additional ethical principles perhaps to counteract this bias.

If we further delve into the potential possibility of ethical principles becoming encoded to AA Systems, a look at what these principles should be would be of detriment. How would the morality of a machine come about? As AA Systems learn from humans, would it be an idea for a moral database cloud to come about, if so who should input to it and how would this be tracked. As we know M.I.T. has generated an online platform denominated “Moral Machine” whereby you play a “game”. Ultimately you are playing “God” by making the choice between who should live and who should not if a car accident was to occur. I don’t necessarily believe AA Systems should be based on such a whimsical website where people are disconnected to the lives in which they are choosing which gets to live or die. As humans are a self-interested group, if it was their mother, or daughter etc. in the car or on the pedestrian crossing the decision would generate a different response than one in which you are clicking through not applying real cognitive thought of the real consequences of how a driverless car should respond.

Section A: Is ethics advanced enough to put into practice? We see an example of practical ethics in the biomedical field but this is when research is determining the best outcome in the future for humans, but what about when it is based around businesses where making money is the goal? *Give an example of questionable ethics surrounding biomedical field*

Compas system in the USA where inmates were analyzed on who would be more likely to re-offend.

Contrast** Is AI advanced enough to use ethics as programming can sometimes be seen to have a bias based on the software engineers inputting the coding.

Is ethics developed enough to be deployed in the commercial world?

As ethics is of a theoretical basis, how can we then make ethics practical? The main example of practical ethics is based on the health system of biomedical advances. As a community, we have the utilitarian view when it comes to the biomedical field that its effects/ outcomes are of value to human beings and therefore we deem these ethically viable. However, how can we input ethical rules into an Autonomous automated system if we cannot say there are for a fact working to a globally acceptable ethical rule. Culturally there will be ethical differences from one culture to the next.

For example, which member of a society is decided upon to be more deserving of protection? If we go back to the driverless car example and the decision which was made between the child or the elderly that gets to survive based on which direction the car would inevitably go, this would differ going from one culture to the next, and in turn the input to the AA System would have to be suited to each culture in order to work in the country it is disposed to.

Ethics may not be developed enough to be deployed into the commercial world as it cannot be standardized into black and white, right and wrong, it is naturally ingrained in humans as we grow up having learned intuitively what decisions are permissible and what simply is immoral. A large number of people may have utilitarianism theoretical values, as they will choose the outcome which has the general overall positive value for the greater amount of people. However if we have a theoretical utilitarian view when It comes to an AA system making a decision based on these values, say a healthy person who did not appear to have the happiest of lives was murdered by an AA System to harvest their organs to save 5 people by giving them the healthy organs so they could go live a happy life, A general agreement would be that this is an immoral act but would have been justified by the utilitarian theory. How would the AA System know any different if solely based on utilitarian ethics?

Then we have the deontological theory in which the rightness and wrongness depend on the act itself but seems to ignore the consequences. Then we have Prima Facia duty which isn’t a single absolute duty to which we must adhere to rather a number of duties some deontological and some utilitarian. None of which are absolute, this is when ethics becomes quite messy and not a standardized set of rules to follow and in turn makes it very difficult to expect a non-sentient, non-intuitive body to come up with a reasonable ethical view or outcome as it may.

Ethical Consistency

Employees of an organisation are constrained by a set of ethical principles set out by the company along with their own personal ethical views, this is further embedded by the law in countries where industries are expected to conduct their businesses in an ethical manner which ineptly helps a business have ethical consistency. Ethical principles are not hard-wired into individuals and due to this people can act out against these principles but they face the consequences if acting in a wrong manner. (Machine Ethics Book)

Hardwiring of ethical laws (Implicitly) into an AI may make us more confident to the fact these machines will act in an ethical way, but as we know AI are learning machines, can we ever be fully confident they won’t change the laws over time.

Furthermore, if a system is explicitly expected to decide on ethical outcomes how do we predetermine that these outcomes are ethically consistent with our own morals. If an AI machine cannot be held directly accountable for unethical acts either by the legal system or by other means of facing consequences, what is to say it will abide by the ethical principles to which they were expected to base their explicit decisions on? How could we be certain that AI will always work with us for the benefit of humans and not eventually for their own benefit, although human beings are noted to be self-interested and this would be more likely an outcome if AI had all of the sentient abilities of a human, therefore this would not be the expected outcome. It could occur as a “black swan” event. (Michaels EBook)

If an AA system was in use in hospitals or caring for the elderly and their main goal was to keep the person under their care safe from harm, what happens if a patient refuses treatment that would save their life, for example, a tracheotomy procedure had to be done and the patient refuses the treatment or say a do not resuscitate request form was filled out, will the AA system respect the patients autonomy? The AA System will require to have consent ethical rules inputted to ensure it is morally withstanding. We also have to look at the principle of consent being stuck with and no intervening outside of the patients control changing the decision for them unless conducted through the legal proceedings.

The European Commission has mandated a draft of guidelines for ethical guidelines for trustworthy AI. This seems to be a good starting point as to what principles AI should follow, it lists beneficence, non-maleficence, respecting the autonomy of humans, justice and explicability. The AA System also has to be human-centric, have governance of the AI autonomy, Auditability – through logs of decisions made, traceability, has to go through an assessment list which is non-exhaustive, has to fall in line with the businesses deontology of charters and code of conduct. The ethical purpose has to respect the fundamental rights, principles, and values.

Under the fundamental rights enlisted in this draft it states “Equality, non-discrimination and solidarity including the rights of persons belonging to minorities” We see an example of AI discrimination in the USA judicial system where the use of a risk assessment algorithm “Compas” was used to assess which offenders would be most likely to re-offend. This algorithm was used in sentencing and parole hearings. It was shown to be twice as likely to state an African American was the most likely to re-offend than a Caucasian. A bias was obvious in this AI’s determinations. Bias may also be an assessment to pass prior to being rolled out in Europe. Although how this may be predetermined may be an issue.

In reality with deep learning Artificial intelligence, due to the fact of it always learning from intake of information, processing and concluding, and producing an output from what it has processed, we can not be sure of what the output will be and therefore cannot ensure that each output will be ethical due to the tendency of a learning machine straying from a previously programmed instruction. An example of this may be a driverless car that is always learning from its environment, which may be beneficial in times of bad weather such as snow or frost, or a change in the weather from a heat wave to a storm causing the road to have treacherous road conditions but where it is taking information in from other road users, with the likes of speeding, or tailgating etc. it will make the environment more precarious.

What about where an AA System is utilized by a multinational business, how can we control the ethical dimensions when the main goal of every business is to turn out a profit margin?

Another difficulty with ethics is it is a feature of human beings who are self-interested bodies. When challenged with the moral dilemmas of the car having to swerve save either the passenger or the pedestrian, It would depend on where that individual was, if they were not in the area it may be a dilemma for them as to who to pick, but If the individual was in the car or crossing the street, the individual would ultimately automatically choose to save themselves. It is a primitive reflex. This leads us to the conclusion that when faced directly in the midst of an accident one does not act with morals, simply a reflex to live. If moral agents cannot make a moral judgement during these situations, how can we expect a non-intuitive agent to make a moral decision?

Section B: If we prevent ethical testing within AI/AAS will this be an impediment to innovation?

We should look towards the Nazi concentration camp “Medical” experiments carried out by Nazi doctors. There were many atrocities carried out in these camps including the horrific and unethical experiments carried out there which were based on 3 types of experiments “Military 

Preventing the ethical testing of AI would most likely be an impediment to innovation. Most AA Systems in use today are of benefit to humans, such as the Roomba hoover, Google Search, The Alexa etc. these all help us save time and energy. If AI was so controlled that it was not granted the use of its own “machine learning” how advanced is AI likely to get?

Testing is required to be carried out but this would need to be in a controlled environment. AI robots which are created to care for the elderly are already in production and use. However, AI robots should not be granted full autonomy due to the fact that humans need to be able to overrule their decisions in case this is impeding on a humans autonomy. It is hard to fathom if it will ever be possible to have a fully autonomous automated system in use in any kind of situation in which will have these systems interacting with humans, animals or the environment unless its features are fully scaled back to minimal outcomes when working on any one feature.

When we look to Hanson Robotics they unveiled Sophia the Human-like Robot that can read facial expressions, decipher between tones of voice etc. they have the ultimate goal of having a robot that is sentient. However, if we have robotics which are sentient, Will this not determine that the robots themselves will work under human rights laws? AA Systems throughout the world are generally in play for the benefit to humans, but ultimately would it be ethical to in all essence rule a robot if it is sentient?

It would generally be seen to be an advantage if a sentient robot was caring for an elderly person as it would be able to converse and ultimately care for the person better than if it was non-sentient. Here we skip back to the matter of consent while working with AI. If AI ultimately becomes sentient do we then have to worry about consent from the AI, how would rights come into play with sentient beings, or would a sense of self be what is required to give a robot rights? A lot of research is in process but we have a long way to go before we can ultimately roll out fully Autonomous Automated Systems for use with people let alone who we consider being our most vulnerable such as the elderly.

AI research is clearly seen to be a great benefit to mankind due to the leaps and bounds in detecting different diseases, which in turn will increase early prevention of illnesses becoming terminal. An accident occurred in Arizona, USA where a driverless car fatally injured a pedestrian, although there was a driver in the Uber car at the time and it is noted the driver is supposed to be paying attention and take over control of the car when required, is this an example of putting less onus on the driver and put them in a situation where they think it is “safe” to not be fully concentrated on the task in hand. How do we determine the testing has gone through a rigorous and robust evaluation prior to rolling out, and when would it be an impediment not to roll out the AI worldwide.

We can contrast and compare with that of the experiments carried out in Nazi concentration camps, ultimately these “Experiments” are looked upon as bad data due to the unethical ways in which the data was gathered. A large number of experiments carried out are of no benefit to the general population due to the advances in science, but where the data could be used such as in the experiments in exposing people to the extremes of cold temperatures which was completely unethical in the way it was gathered, could be of use to researchers today especially for the likes of coast guard teams who undergo training on resuscitation and attempts of re-warming victims exposed to hypothermia. The non-use of this data may really be a serious impediment to saving lives, as there was data gathered on the rapid re-warming of victims noted as the best outcome of the re-warming techniques which were tested. The dilemma here is if it is ethical to use data which has been gathered in an unethical way. Many scientists and doctors argued that is served no purpose to science to ignore data that could help people. Dr Hayward was one of the doctors that had this view as he had used the data gathered in his own research into the planning of naval rescue missions but due to the data’s unethical background, this research could not be published. Dr Hayward considered this to be criminal not to use the data if it saves lives. When the data may be critical to saving lives could it be considered ignorant not to use it? The premise of analysis may be when the medical crisis is real and the benefit to society is great, the data should be used. When the medical problem is not pressing and the benefit to society is relatively marginal, the need to preserve the integrity of the medical profession and the victim’s memory outweighs the potential benefit to society. Another argument against the use of the data was the validity of the data, as the experiments could not be repeated how does one know this data is valid.

Deep Learning of AI is the intake of large amounts of data, processing it, and creating a conclusion which leads to an output. If AI is ever to be made fully autonomous, the way in which AI comes to a conclusion needs to be understood. As of today, there is no understanding of this output.

Deterring from testing the ethics of an autonomous automated system would be an impediment to innovation that’s to say if we all were under the agreement that innovation would halt due to the averting of testing. None the less this is unlikely to be the case as artificial intelligence is advancing each day around the world at a speed that is unfathomable. Merely deterring of the testing would put more of an impediment to the ethicality of autonomous automated systems.

When we talk about the testing of AA systems, the testing would be one of the main challenges to innovation how do we test a system which is ever changing? How do we ensure it does not stray from the ethics which it may have prominently shown during the evaluation and subsequent passing of the test?

Conclusion: Summary of Main Points & Personal Conclusions

Is it inevitable that we have to make considerations for ethics to be programmed into AI? It would be irresponsible not to. Autonomous automated systems are ultimately tools of the human beings who design and manufacture them, therefore programmers producing them have a moral obligation to make these machines ethically abled.

Is ethics sufficiently developed to be deployed in the commercial world? Due to the diversities around ethics from the different ethical theories to how each individual would see what an outcome should be, the prematurity of having an autonomous automated system decipher on what ethical outcome a situation should have seems to be a high peak to overcome before considering an AI to have full autonomy.

Will it ever be fully possible to rely on AI without human input or will human input always be a consideration for AI? Ultimately would it be completely irresponsible to rely on deep learning machines which are taking in data, processing it, coming up with a conclusion and then producing an output? The main element of these machines is to keep learning, updating and evolving. How could it be possible to ensure these machines are continuing to learn while also keeping certain values in check? They are required to come up with better conclusions but ultimately keep some constants, how could this be programmed into a machine to ensure ethical consistency.

It may be a good idea to come up with a moral database where the AI could go to and learn from for how to prevent harm coming to humans, although it would also have to be looked at to where does it need permission to prevent harm coming to a human, such as in the medical field if an AI machine/robot was at work in the hospital and a patient does not want to have an intervention to save their life. Large reviews would be required as to what would be inputted to this database. We need to look at the matter of consent and where will an automated machine know to accept the consent of an adult or more so when will it know that intervention is needed where competence is not there to give consent.

Ethic programming should not be done where ethicists cannot agree on for an example of ethical dilemmas.

References:

• Anderson, M. and Anderson, S. (2011). Machine ethics. 1st ed. Cambridge: Cambridge University Press, pp.1-28.
• ROZENBERG, J. J. (2003). Bioethical and ethical issues surrounding the trials and Code of Nuremberg: Nuremberg revisited. Lewiston, Edwin Mellen Press.
• Michalczyk, JJ (ed.) 2017, Nazi Law: From Nuremberg to Nuremberg, Bloomsbury Publishing PLC, London.