As AI plays a larger role in society, ethical concerns have started to come into the forefront. A great example is the ICML 2018 Best Paper being awarded to Delayed Impact of Fair Machine Learning by Liu et al. ([2018](https://arxiv.org/abs/1803.04383)). Check out the blog post [here](http://bair.berkeley.edu/blog/2018/05/17/delayed-impact/)!

Today, artificial intelligence is such a buzzword that I feel somewhat guilty for being an AI researcher. Yet, with all the money pouring into AI, there seems to be a lack of interdisciplinary practitioners who can bridge the gap between humanities and the latest technology developments.

Problems such as racial bias in judicial algorithms and face recognition systems, realistic image and speech synthesis leading to pornography of dubious legality, the muddled ethics of self-driving cars and most recently, the uncanny valley of Google Duplex, all point towards the hardcore technologist’s unwritten creed of “Can it be done?” instead of “Should it be done?”. As we stand like children today, in anticipation of tomorrow’s robots and AIs and self-driving cars and drones, we forget to wonder if there will be a day after tomorrow.

That is not to say that the Luddites have been right all along. Rather, as the oft-quoted uncle said, “With great power comes great responsibility.” Genius has to be tempered with maturity and an understanding of human nature, or risk being prefixed with ‘mad’ or ‘evil’ and one day waking up to a pitchfork-wielding mob.

Instead of the apocalyptic fears of Skynet and Roko’s Basilisk, I am primarily concerned about the very real short-term, even immediate, problems with AI that has rapidly permeated many aspects of society - bias propagation, blackbox decisions, privacy concerns, amongst other issues starving for solutions.

This does not mean I am not at all concerned with the existential threats due to AI. I do agree that the existential threats have to be prevented rather than treated, since any notion of treatment is likely to be too late. Furthermore, I think we will find that both classes of problems have related origins, such as issues of interpretability and goal-alignment.

Below I list several issues with AI in the social context. These range from AI-specific challenges to problems that tend to accompany any major technological shift. In the latter, I find that even amongst the class of major technological shifts (steam, electricity, internet), AI is relatively unique in its cultural baggage and the extent and speed of its pervasiveness.

1. Bias Propagation

It is important for people to recognize that algorithms, despite all the association with cold hard calculative logic, are just as prone to bias as humans.

Humans tend to accumulate bias in two ways - inculcation or first-hand experience. (Biological factors may also account for some form of bias, such as a predisposition to prefer certain physical features.)

First-hand experience is relatively straightforward - particularly memorable events that cause the formation of personal stereotypes. A trivial example might include how certain names might sound particularly obnoxious or humble or fat or thin or tall or short, despite common sense telling us that names should have no effect on a person’s characteristics.

Actually, name-signalling is actually a well-documented phenomenon, which means that, surprisingly, names do have some effect on personal characteristics. Ironically, these effects usually have their origins in human bias, kind of like a self-fulfilling prophecy.

When considering inculcation and upbringing, bias and discrimination can be propagated in families, tribes and cultures, through every aspect of everyday behavior - verbal teachings, gossip, nonverbal cues, songs, stories.

An interesting article on how culture might influence perceptions of AI - Why Westerners Fear Robots and the Japanese Do Not

From the perspective of an AI researcher, both first-hand experience and upbringing form the training data of a human. Certain trends in the data surface themselves as biases in the human.

In a similar manner, training data provided to algorithms can result in biases being inherited by the algorithms. These biases will stem primarily from the humans that generated the data.

Just because it is not human does not mean it is fair and without bias.

A great example are word-embedding algorithms (eg. Word2Vec and GloVe). These algorithms generate vector representations of words ie. each word is mapped to a vector of n numbers, where n typically ranges from 50 to 300. The cool thing about the vector representation is that we can do ‘word arithmetic’. A classic example is:

King - Man + Woman = Queen

Another way of interpreting the equation is to see that

King - Man = Queen - Woman

which in turn demonstrates analogy relationships (King is to Man as Queen is to Woman).

Researchers were fascinated at the prospect of such ‘word arithmetic’ until we realized that we could also calculate all sorts of other more provocative equations, for instance examining if rich/poor is analogous to white/black.

Word2Vec demo available here and the King - Man + Woman equation can be seen by entering 'man', 'king' and 'woman' in that order, in the Word Analogy section.

A more visual demonstration can be found here (can take several minutes to load).

Another related example is language models, which try to predict a next word or token, given a sequence of previous words. For example, given the words ‘The sun is’, the language model might predict high probabilities for words such as ‘hot’ and ‘bright’. Such models can also exhibit undesirable bias such as predicting certain gender pronouns as more likely given certain contexts or predicting certain attributes such as race, religion, nationality.

How do we detect bias and how can we differentiate bias from real patterns?

In particular, bias seems analagous to wrong predictions - we often consider biases as false stereotypes that lead to incorrect judgement. Predicting that a thrown rock falls to the ground is not bias. Deciding that someone, because of their race, should not be allowed to take out a loan is bias.

In the language model example, suppose a model predicts the male pronoun as more likely given that the preceding words mention ‘lawyer’. Such a model will have made a wrong prediction if the ‘lawyer’ in question was referenced using the female pronoun in the true text. But here’s the caveat, suppose it is true that lawyers are more often male than female. Then is the algorithm wrong in predicting a male pronoun with a higher probability?

Interestingly, the fact that algorithms can exhibit bias might indicate a novel way of detecting bias in training data. Specifically, if the bias did not exist in the training data, the tendency would not have surfaced in the trained model. The model can be seen as a distillation of the prejudice recorded in the training data.

How should we deal with identified algorithmic bias?

Finally, another contentious point is how to treat algorithmic bias. Algorithmic bias, once detected, might seem simple to systematically remove, as compared to human bias that is arguably resistant to change. However, the use of blackbox models might impede the diagnosis and cure. Furthermore, the solution is not as simple as simply removing the bias. In Delayed Impact of Fair Machine Learning, Liu et al. suggest that any corrective measures could have long-term repercussions that reverse short-term gains.

Links to articles documenting bias in AI:

2. Blackbox Decisions

3. Privacy

4. Weaponization

5. Media Hyperbole and Misinformation

Check out Approximately Correct! It's a blog started by Zachary C. Lipton, an assistant professor at Carnegie Mellon University, as a counter against overwhelming AI hyperbole and false hype in mainstream media.

6. Inequality

7. Playing Catch-up

My second motivation relates to my dream of using AI for social good. For instance, a team from MIT previously developed software for detecting and counting roofs from satellite images, for the purpose of estimating appropriate sites for microgrids and other infrastructure initiatives. On the same note, earlier this year, as part of a submission to a humanitarian aid and disaster relief app competition, I adapted an algorithm to detect flooded regions and structural damage from satellite images.

Recent applied-AI innovations have catered primarily to consumer delight. After all, that’s where the money is. But I see no reason why AI should not improve and amplify the efforts of aid workers and organizations. Rather than using AI as a crude tool, I seek to wield it with precision, complemented with an understanding of humanitarian problems acquired from this programme, with the knowledge that there are people, not numbers, on the other end of the equation.

AI looks to be the driving force for at least the next decade. AI engineers and experts are in high demand and low supply. But perhaps, more importantly, we need AI engineers who can also play the role of urban planner, AI experts who are also human experts and society experts.