Artificial intelligence may never truly do anything more than what we want, but in a world where we more often than not lack comprehensiveness in spelling out what we wish for, these algorithmic minds may be led astray and deliver us product and services misaligned with what we truly desire (Quanta, 2020)
Now common across many media platforms, YouTube’s content recommendation system, one aimed at maximizing time spent on the platform, uses AI programs to deliver us content fitting our viewing history and profile. One issue that was not foreseen though, is that said recommendation become increasingly “extreme” with respect to the first video watched. Within a few clicks, one would end up watching videos promoting one-sided views on any given subjects. Watching one videos on climat change may thus lead you to climat change denials. One video on adopting a vegetarian diet would lead to veganism promoting content. Such approach, which consists of steadily using the “intensity” of videos may thus lead to increasing polarisation on all matters of subjects and undermine the efforts for healthy and constructive dialogs.
This issue, ever so conspicuous with the novel corona-virus driven pandemic and inherent heated debates, stands as only but an instance of the many externalities that arise from the broader, technical problem of alignment. While it may appear easy to point the flaws of an AI programmation, developers and engineers do not possess the luxury of after-the-facts evaluation and must therefore account for each an every ramification of the targeted wants. The stress put on designing AI with absolute certainty with little to no certain nor definitive information in conjuction with no comprehensive definition of the targeted objectives is pivotal in the deployment of large scale, complex Ai solutions such as self-driving cars.
One critical challenge of developing AIs is the implementation of nuances in achieving the often already oversimplified goals we assigned them to. Building up on the idea of self-driving cars, one must first consider what we want a self-driving car to do. While trivial at first, this question rapidly evolves into an extremely complex set of requirements. First, we will say we want to be driven from point A to to B. Then we add that want so to be done safely, after which we need to define safely. Then we realise that being driven safely may lead to the car stopping for even the slightest of reasons such as a plastic bag floating in the wind, creating congestion and uncomfortable journeys. We now realise we want to be driven rapidly, which too needs definition. So on and so forth, designers and engineers are soon to face a maze-like set of requirements, rules, and constraints that, if overlooked or oversimplified, lead to misaligned AI-driven services.
While AIs shine in performing narrow, repetitive tasks at record-breaking speeds and volumes, they are still severely inconsistent and limited in performing even mundane task with no inherent predictability. That is because, since AI are computational aggregate constructs, the reward function which they seek to solve is one that deals in absolutes, thus making the perfect weighting of ever so complex set of goals impossible, especially as these goals progressively entail less exactitude and more subjectiveness.
Quite an oversimplification but compelling an argument nonetheless, these limitations find resonance in the word “creativity” one for which we still lack proper neurological explanation, even less so an algorithmic one.
This post is only but the mere introduction to the complete spelling out of the alignment problem, one which lies at the core of modern AI developments.If you wish to know more about the technical complexity and go beyond the buzz-wordy nature of new technology, I can only recommend you give the website “quanta magazine” a read as they delve deeper into these questions.
Reference:
https://www.quantamagazine.org/artificial-intelligence-will-do-what-we-ask-thats-a-problem-20200130/
Hi Hadrien and nice post!
The self driving car example is indeed a very good one to understand how differently and specifically humans have to design AI and its applications in order to mimic not only human behavior but provide us with a service we actually want to be done a certain way. For us it seems obvious that a bag on the road possesses no real danger to us or the car, but this is very difficult to implement and design in a system that largely depends on censors to “make decisions”.
One thing I wanted to comment on was the Youtube algorithm system that handles recommendations. I think that as long as a system is developed by a organization or enterprise that has the maximization of profits as its main incentive, an AI program can never be ethically sound. And that is because it will not be designed to be so in the first place, but rather adhere to ulterior motives that benefit the respective company or organization. For this example, the goal of the algorithm is not to find out the “best” and most accurate recommendations for the viewers, but rather as you said more “extreme” ones that will make the user stay on the platform the longest, while continuing to watch similar content on their next sessions.
Hello Hadrien and thank you for such a good post!
The topic is indeed interesting and very good points provided for the YouTube’s content algorithms. I also started to notice that the one-sided information is appeared in recommendations all the time I’ve trying to watch/read something. I was thinking that in cases of different algorithm approach, it would be better for people to consume the information via the platforms that read the data and ‘clicks’. The interest cannot be one-sided. What is your opinion about potential solution on algorithm recommendations?
Very nice example of a self-driving car since the technological disruption is everywhere now, covering the vast majority of industries. The recent articles that I’ve read was about the levels of automative vehicles, and the world is currently succeeded to have the 2nd level of automation. Indeed, AI is a critical part of the technology and now the major constraint lies in adoption of advanced AI for the 3rd level of automation. However, it is a complex procedure since it’s too early for AI to replace humans, especially the emotional and behavioural parts. Therefore, I do not think that the 3rd level can be achieved soon as well as the highest, 5th level, can be achieved at least someday.
Good morning Hadrien,
Thank you for this interesting article about AI. The blog post follows a logical and coherent structure, in which the arguments are well defined and logically built up. I concur with your notion that Youtube’s (AI) algorithm promotes increasingly extreme videos at both ends of various spectra (e.g. both videos in favour of veganism while videos despising veganism and encouraging a carnivore diet).
The problem you raise with this line of reasoning is a compelling yet difficult one. It is compelling because it appears we are all in this large-scale social experiment while nobody can exactly predict how it will unfold. It is difficult because, as you noted as well, there is not only an alignment problem with AI itself but also with the business models of most of the companies deploying AI, such as Youtube and Facebook. While the goal of such firms is to maximise time-on-sight, as you righteously observed, the goals of its users seem different. Let us see if we can take this argument to a more philosophical level.
Consider the difference between the ”experiencing self” and the ”remembering self”. The former one can be described as ‘oneself’ in a particular moment, for instance when consuming a Youtube video. The latter can be described as ‘oneself’ who remembered doing an activity, such as scrolling through Facebook. It has long been noted by philosophers and scientists that there appears a mismatch between the goals and values of the experiencing self as opposed to the remembering self. To illustrate this point, imagine you are watching a Youtube video. The video comes to an end and the algorithm suggest a new video to you, which so strongly appeals to you that you decide to click on it. In that moment, chances are that you are perfectly happy spending a few more minutes of your day on Youtube. Yet, very often, when people think back about their day and they realise that they have squandered quite some time on Youtube, they are dissatisfied with that observation. Hence the apparent mismatch between the goals and values of both ”selves”. I would argue that an awareness of such differences is the basis for changing one’s behaviour, at least in the sense that one gains a more profound understanding of how the mind works.
Finally, I would like to disagree with your notion that ”…While AIs shine in performing narrow, repetitive tasks at record-breaking speeds and volumes, they are still severely inconsistent and limited in performing even mundane task with no inherent predictability…”. AI beat the world’s number one Alpha Go player in 2017. Admittedly, there is some predictability in this game yet it also requires considering various options, choices, and moves as it is highly complex. Therefore I do not consider the claim you make well founded in the light of such events. At the same time, you might have a differing definition of ”mundane” or ”inconsistent” than I do, which might explain my disagreement.
Thank you for your blog post and making me think this morning.
Hi Jan,
Very compelling and complementary answer that you wrote here. I would definitely support the points you made. As for your disagreement, I think I do have a differing definition of mundane activities. What I meant was that AI are immensely effective at performing rapid computations, and not so much at being “creative”, a word which I think encompasses some of the most complex concepts humans have been wrestling with. Although I am not familiar with alpha go, I will make an educated guess and qualify it as a game whose characteristics resembles Chess and other games that involve both reactivity and foreseeing. While undoubtedly extremely complex to us humans, these algorithms (game rules can to some capacity be considered as algorithm) are still in essence a succession of short, rapid calculations we do with our brains. While the exponentially vast successions of computations is complex, the computations themselves are not. Think of any task that only requires you a fraction of second to process, then there’s a high chance that an AI will be good at it. In principle, the idea is the same with calculators, while calculating 1+1 will only take you a split second, 1+1+1+1+1+1+1+1+1+1+1+1+1+1+1+1+1+1+1+1+………… will become increasingly hard for us to calculate. Obviously i avant-midi no expert in this field but I hope my reply does clarify a bit.