What AI needs to learn about dying and why it will save it

Those programming truthfulness can learn a lot from mortality.

Andrew Steane

Andrew Steane has been Professor of Physics at the University of Oxford since 2002, He is the author of Faithful to Science: The Role of Science in Religion.

An angel of death lays a hand of a humanioid robot that has died amid a data centre

A digital momento mori.

Nick Jones/midjourney.ai

Google got itself into some unusual hot water in recently when its Gemini generative AI software started putting out images that were not just implausible but downright unethical. The CEO Sundar Pichai has taken the situation in hand and I am sure it will improve. But before this episode it was already clear that currently available chat-bots, while impressive, are capable of generating misleading or fantastical responses and in fact they do this a lot. How to manage this?

Let’s use the initials ‘AI’ for artificial intelligence, leaving it open whether or not the term is entirely appropriate for the transformer and large language model (LLM) methods currently available. The problem is that the LLM approach causes chat-bots to generate both reasonable and well-supported statements and images, and also unsupported and fantastical (delusory and factually incorrect) statements and images, and this is done without signalling to the human user any guidance in telling which is which. The LLMs, as developed to date, have not been programmed in such a way as to pay attention to this issue. They are subject to the age-old problem of computer programming: garbage in, garbage out.

If, as a society, we advocate for greater attention to truthfulness in the outputs of AI, then software companies and programmers will try to bring it about. It might involve, for example, greater investment in electronic authentication methods. An image or document will have to have, embedded in its digital code, extra information serving to authenticate it by some agreed and hard-to-forge method. In the 2002 science fiction film Minority Report an example of this was included: the name of a person accused of a ‘pre-crime’ (in the terminology of the film) is inscribed on a wooden ball, so as to use the unique cellular structure of a given piece of hardwood as a form of data substrate that is near impossible to duplicate.

The questions we face with AI thus come close to some of those we face when dealing with one another as humans.

It is clear that a major issue in the future use of AI by humans will be the issue of trust and reasonable belief. On what basis will we be able to trust what AI asserts? If we are unable to check the reasoning process in a result claimed to be rational, how will be able to tell that it was in fact well-reasoned? If we only have an AI-generated output as evidence of something having happened in the past, how will we know whether it is factually correct?

Among the strategies that suggest themselves is the use of several independent AIs. If they are indeed independent and all propose the same answer to some matter of reasoning or of fact, then there is a prima facie case for increasing our degree of trust in the output. This will give rise to the meta-question: how can we tell that a given set of AIs are in fact independent? Perhaps they all were trained on a common faulty data set. Or perhaps they were able to communicate with each other and thus influence each other.

The questions we face with AI thus come close to some of those we face when dealing with one another as humans. We know humans in general are capable of both ignorance and deliberate deception. We manage this by building up degrees of trust based on whether or not people show behaviours that suggest they are trustworthy. This also involves the ability to recognize unique individuals over time, so that a case for trustworthiness can be built up over a sequence of observations. We also need to get a sense of one another's character in more general ways, so that we can tell if someone is showing a change in behaviour that might signal a change in their degree of trustworthiness.

In order to earn our trust, an AI too will have to be able to suffer and, perhaps, to die.

Issues of trust and of reasonable belief are very much grist to the mill of theology. The existing theological literature may have much that can be drawn upon to help us in this area. An item which strikes me as particularly noteworthy is the connection between suffering and loss and earning of trust, and the relation to mortality. In brief, a person you can trust is one who has ventured something of themselves on their pronouncements, such that they have something to lose if they prove to be untrustworthy. In a similar vein, a message which is costly to the messenger may be more valuable than a message which costs the messenger nothing. They have already staked something on their message. This implies they are working all the harder to exert their influence on you, for good or ill. (You will need to know them in other ways in order to determine which of good or ill is their intention.)

Mortality brings this issue of cost to a point of considerable sharpness. A person willing to die on behalf of what they claim certainly invests a lot in their contribution. They earn attention. It is not a guarantee of rationality or factual correctness, but it is a demonstration of commitment to a message. It signals a sense of importance attached to whatever has demanded this ultimate cost. Death becomes a form of bearing witness.

A thought-provoking implication of the above is that in order to earn our trust, an AI too will have to be able to suffer and, perhaps, to die.

In the case of human life, even if making a specific claim does not itself lead directly to one's own death, the very fact that we die lends added weight to all the choices we make and all the actions we take. For, together, they are our message and our contribution to the world, and they cannot be endlessly taken back and replaced. Death will curtail our opportunity to add anything else or qualify what we said before. The things we said and did show what we cared about whether we intended them to or not. This effect of death on the weightiness of our messages to one another might be called the weight of mortality.

In order for this kind of weight to become attached to the claims an AI may make, the coming death has to be clearly seen and understood beforehand by the AI, and the timescale must not be so long that the AI’s death is merely some nebulous idea in the far future. Also, although there may be some hope of new life beyond death it must not be a sure thing, or it must be such that it would be compromised if the AI were to knowingly lie, or fail to make an effort to be truthful. Only thus can the pronouncements of an AI earn the weight of mortality.

For as long as AI is not imbued with mortality and the ability to understand the implications of its own death, it will remain a useful tool as opposed to a valued partner. The AI you can trust is the AI reconciled to its own mortality.

Explainer

Belief

Creed

6 November 2024

5 min read

Whether it's AI or us, it's OK to be ignorant

Our search for answers begins by recognising that we don’t have them.

Simon Walters

Simon Walters is Curate at Holy Trinity Huddersfield.

A street sticker displays multiple lines reading 'and then?'

Stephen Harlan on Unsplash.

When was the last time you admitted you didn’t know something? I don’t say it as much as I ought to. I’ve certainly felt the consequences of admitting ignorance – of being ridiculed for being entirely unaware of a pop culture reference, of being found out that I wasn’t paying as close attention to what my partner was saying as she expected. In a hyper-connected age when the wealth of human knowledge is at our fingertips, ignorance can hardly be viewed as a virtue.

A recent study on the development of artificial intelligence holds out more hope for the value of admitting our ignorance than we might have previously imagined. Despite wide-spread hype and fearmongering about the perils of AI, our current models are in many ways developed in similar ways to how an animal is trained. An AI system such as ChatGPT might have access to unimaginable amounts of information, but it requires training by humans on what information is valuable or not, whether it has appropriately understood the request it has received, and whether its answer is correct. The idea is that human feedback helps the AI to hone its model through positive feedback for correct answers, and negative feedback for incorrect answers, so that it keeps whatever method led to positive feedback and changes whatever method led to negative feedback. It really isn’t that far away from how animals are trained.

However, a problem has emerged. AI systems have become adept at giving coherent and convincing sounding answers that are entirely incorrect. How has this happened?

This is a tool; it is good at some tasks, and less good at others. And, like all tools, it does not have an intrinsic morality.

In digging into the training method for AI, the researchers found that the humans training the AI flagged answers of “I don’t know” as unsatisfactory. On one level this makes sense. The whole purpose of these systems is to provide answers, after all. But rather than causing the AI to return and rethink its data, it instead developed increasingly convincing answers that were not true whatsoever, to the point where the human supervisors didn’t flag sufficiently convincing answers as wrong because they themselves didn’t realise that they were wrong. The result is that “the more difficult the question and the more advanced model you use, the more likely you are to get well-packaged, plausible nonsense as your answer.”

Uncovering some of what is going on in AI systems dispels both the fervent hype that artificial intelligence might be our saviour, and the deep fear that it might be our societal downfall. This is a tool; it is good at some tasks, and less good at others. And, like all tools, it does not have an intrinsic morality. Whether it is used for good or ill depends on the approach of the humans that use it.

But this study also uncovers our strained relationship with ignorance. Problems arise in the answers given by systems like ChatGPT because a convincing answer is valued more than admitting ignorance, even if the convincing answer is not at all correct. Because the AI has been trained to avoid admitting it doesn’t know something, all of its answers are less reliable, even the ones that are actually correct.

This is not a problem limited to artificial intelligence. I had a friend who seemed incapable of admitting that he didn’t know something, and whenever he was corrected by someone else, he would make it sound like his first answer was actually the correct one, rather than whatever he had said. I don’t know how aware he was that he did this, but the result was that I didn’t particularly trust whatever he said to be correct. Paradoxically, had he admitted his ignorance more readily, I would have believed him to be less ignorant.

It is strange that admitting ignorance is so avoided. After all, it is in many ways our default state. No one faults a baby or a child for not knowing things. If anything, we expect ignorance to be a fuel for curiosity. Our search for answers begins in the recognition that we don’t have them. And in an age where approximately 500 hours of video is uploaded to YouTube every minute, the sum of what we don’t know must by necessity be vastly greater than all that we do know. What any one of us can know is only a small fraction of all there is to know.

Crucially, admitting we do not know everything is not the same as saying that we do not know anything.

One of the gifts of Christian theology is an ability to recognize what it is that makes us human. One of these things is the fact that any created thing is, by definition, limited. God alone is the only one who can be described by the ‘omnis’. He is omnipotent, omnipresent, and omniscient. There is no limit to his power, and presence, and knowledge. The distinction between creator and creation means that created things have limits to their power, presence, and knowledge. We cannot do whatever we want. We cannot be everywhere at the same time. And we cannot know everything there is to be known.

Projecting infinite knowledge is essentially claiming to be God. Admitting our ignorance is therefore merely recognizing our nature as created beings, acknowledging to one another that we are not God and therefore cannot know everything. But, crucially, admitting we do not know everything is not the same as saying that we do not know anything. Our God-given nature is one of discovery and learning. I sometimes like to imagine God’s delight in our discovery of some previously unknown facet of his creation, as he gets to share with us in all that he has made. Perhaps what really matters is what we do with our ignorance. Will we simply remain satisfied not to know, or will it turn us outwards to delight in the new things that lie behind every corner?

For the developers of ChatGPT and the like, there is also a reminder here that we ought not to expect AI to take on the attributes of God. AI used well in the hands of humans may yet do extraordinary things for us, but it will not truly be able to do anything, be everywhere, or know everything. Perhaps if it was trained to say ‘I don’t know’ a little more, we might all learn a little more about the nature of the world God has made.