Artificially Intelligent Systems: A Quintessence of Life

A section from the vast field of machine learning, natural language processing, has seen updates that have sent chills across disciplines with its potential implications and impacts on human society, including the pivotal role it will play in shaping rules for societies of the future.

GPT-3, the language prediction model created by Open AI, was built by directing machine-learning algorithms to study the statistical patterns in over a trillion words collected from the internet and digitized books. The model generates tweets, pens poetry, summarizes emails, translates languages, answers trivia questions, and even writes its own computer programs. Although its output is grammatical and even impressively natural, its comprehension of the real world is off, making it permanently unreliable in decision making. Its application in the medical field verifies this behaviour, a medical chatbot powered by GPT-3 in an actual conversation with a fake mental health patient supported their intent to commit suicide. While the model lacked scientific and medical expertise, its reactions in human trials show deeper ethical issues. Such AI-based models trained on exponentially increasing amounts of public data from the internet allow racist, sexist, and otherwise abusive language in training data and cause unthinkable societal harm.

With no concept of life, birth, emotions or even an understanding of the meaning of beauty, the approach of machine learning models are drastically different from human cognition and while in the present they are only good at mimicking human language, their future potential may lie beyond human intellect.

Garbage in, Garbage out

Researchers from John Hopkins University and multiple other studies confirm that the biases present in training data lead AI-based models to generate stereotyped or prejudiced content.

Statistically, bias has a quantifiable definition in machine learning, but legally it is defined as judgement based on preconceived notions of prejudice. While the meaning of bias transcends disciplines, it causes societal harm, at times due to allocation, economic and transactional in nature, where AI-based systems allocate or withholds certain groups an opportunity or resource, like models in financial decision-making used for credit applications processing which potentially reject certain sections of the population. And then, there are representational harms, socio-cultural, where AI-based systems reinforce the subordination of some groups along the lines of identity, an example of it being the labelling of a photo of an African American woman as a gorilla by Google Photos, a label that had a long history of being used purposely to demean a community. While the instance is an example of a systemic flaw, it is important to concede that these AI-based models fail to understand and represent certain communities due to biased datasets and cause irreparable socio-cultural harm. 

The datasets that should have been defined reality for AI-based system based on facts, capture the ideologies and opinions of the people who instead of tagging identities and objects without bias, have ranked social hierarchies and power structures. The heavy skew of the white/ male population towards positive connotations, opposite for other ethnicities, minorities and genders, the dated portrayal of women, the underrepresentation of diversity under gender & orientation are just some of the corruptions that have plagued data sets. These inputs don't just affect the model's accuracy but amplify the influence of dominant ideologies that define the AI's perception of people and the world, which gets captured in its targeted immoral decision-making errors. 

Facial recognition models fail to recognize Black, Middle Eastern, and Latino people more often than those with lighter skin. Its application in policing, facial recognition technology used by the Detroit Police Department misidentified suspects about 96% of the time. Amazon's AI to vet job applications, penalized applicants who attended women’s colleges, as well as any resumes that contained the word “women’s”.

While the field endures a diverse range of problems across disciplines, the people driving it seems to increasingly reflect a narrow demography of opinion, white and male. Women account for 18% of authors at leading AI conferences, 20% of AI professorships, and 15% and 10% of the research workforce at Facebook and Google. Black workers represent 2.5% of Google’s entire workforce and 4% of Facebook’s and Microsoft’s.

One of the fixes for Google's gorilla problem was to erase gorillas, and some other primates, from the service’s lexicon, a lazy workaround that illustrates the limited thinking and priorities of Big Tech in advanced image-recognition technology when it comes to solving representational biases in the application of a technology that lies at the foundation of self-driving cars, personal assistants, and other AI-based systems.

Rooted in the re-classification of life and society, the future challenges of our misguided civilisation would be philosophical and socio-cultural.

As we live and breathe, AI-based models are conducting the largest experiments in human history where everything from human faces, actions to millions of places and objects are being used to draw specific conclusions of segmented populations. While these classification attempts are not new, historically, they have often been used by authoritarian political regimes to organize and enforce atrocious systems of segregation and control and the accelerated application of AI-based systems to analyse consumer data in the present, matches with the right-wing mindset that influences legislation.

The book of life was a population register system powered by IBM and used by the South African government in the height of apartheid. The system classified people (coloured, Indian, white or black) and determined where one could live, what job one could have and even affect decisions on marriage. Similiar such ideas include a model trained on Chinese government-issued ID's which predicted potential criminality based on facial features.

These are not isolated attempts, there are models in training and application right now, applied on public data to predict individual behaviour based on collated individual digital footprints, classify identities into subsets based on undiscussed logics for monetisation and model training. The principle that runs through the core of all these attempts is "surveillance-first, permission-later", when legislation to protect public rights, a forum to standardise and uphold ethics of public data use or even an intellectual debate on the rights of modern society exists.

While we approach the inflection point beyond which the majority of knowledge of life and this universe will be non-human, the present society across disciplines of politics, research, ethics, law, spiritual, finance, and welfare faces a deeper moral choice.

The bias that plagues machine learning models is not technically a system error but rather a quantification of what is precisely wrong with society. The datasets that are used to train these AI-based models reflect the ideologies and choices of the minority who made them resulting in the outcome of these models, indirectly/ directly raising moral questions on empirical/ present legislative infrastructure and the market system which need to be resolved not just to correct the sample data but also due to it being the human thing to do. While public and regulatory participation in this forum of collective decision-making is low and efforts to drive the field represents the monopolies that rule the world. But no individual element, especially the handful of corporations that drive AI's planetary network, would be able to get the models right and the choices of people driving the varied disciplines that marry the field machine learning make will be the collective choices of mankind. 

The true cost of AI is ignorance and billions of years of Earths history.

One would assume Siri, Cortana, Google Assistant or Alexa as cost-free convenient services, overestimate their intelligence/ accuracy and often accept suboptimal decisions, but the astronomical planetary cost of these voice-based computer systems remains incomprehensible to the end-user.

A simple command on any of these systems starts a chain of events that create a complex network of power and wealth that remains closed to the public opinion and results in an inhuman skew in wealth distribution amongst the lowest and highest elements of this industrial complex. While the complexities of logistics and supply chain enable our misunderstanding of the real costs of AI-based systems, its environmental and inhuman labour costs lie unprioritised in our ignorance.

The Tesla Model S has about 12 kg of Lithium in it and mining about 1000 kg of it from the Lithium Triangle which covers parts of Argentina, Bolivia, Chile and holds more than half the world’s supply of the metal beneath its salt beds, uses 500,000 gallons of water for filtration and extraction. The region is also one of the driest places on earth and in Chile’s Salar de Atacama, mining activities consumed 65 per cent of the region’s water as the nation faces the impending crisis of depleting water supplies while Lithium demands reach new peaks.

Our current addiction to rare earth elements (cobalt, lithium, neodymium, vanadium and gallium) is driven by the increased adoption of AI-based systems in defence and commercial applications, including aircraft/ UAVs/ vehicles, sensors, precision-guided munitions, electric motors, nuclear reactors and more. While these elements just help corporations built materialistic tools, the true genius lays hidden in the AI-based systems powering it. One would assume that the huge sacrifice our planet has made for progress and helped mankind achieve things beyond changing a song or setting an alarm on their cloud-based devices, would be enough, but the input for this machinery that has made private corporations bigger than empires that ruled the world keeps on increasing, but the flow of wealth and power remain consistent.

With a huge entry barrier of the scale of investment requirements, the systems that train/ develop natural language processing model require computation system built on specialised chipsets. While this specific element in the supply chain has its role to play in politics that will change the future landscape, the demand for chips will go hand in hand with the increased acceptability of AI-based systems under the shadow of the end of Moore's law and a beginning of a new reality of quantum computing.

Looking into a variable that affects public policy planning, electricity consumption, and we realise the tragedy that runs behind humanity's charge to go sustainable. With energy consumption skyrocketing around the world due to cryptocurrency mining, training AI-based models have followed the same trends and was one the point made by Timnit Gebru in the paper that rocked Google.

Since the early years of machine learning to 2012, the number of computational resources required by the systems doubled every two years (similar to Moore’s law of growth in processor power), post-2012, the trajectory of computing power for building and training models, on average has doubled every 3.4 months. Training a single large natural language processing model consumes energy equivalent to building 5 average American cars and running them over their entire lifetime. 

Under the guise of Magic lay hidden efforts and misery of the many and an ocean of data and wealth for the few.  

The golden words that rule silicon valley, apply to the field of AI as well and while one would assume that we are the product for these AI-powered systems, we are much more, we are the end-users, a resource, labour and finally the product. Looking at voice-enabled AI systems clarifies our role in this planetary machinery, we as consumers avail these systems by buying AI-based products, as a resource, submit our voice commands for analysis and to build an ever-increasing database of human voices and instructions, finally as labour, we continually perform the valuable service of providing feedback regarding the accuracy, usefulness, and overall quality of the systems, in essence, helping train the neural networks for even more intimate application. 

The financial investment, energy consumption and environmental damage caused as a direct or indirect result of this field though only represent the material costs and beyond this veil is the labour costs of micro workers from around the world, often in underdeveloped/ developing markets with low pay and rights, and the billions of hours gone into building classification systems that tag humans, their actions, behaviour, places, almost any and every digital footprint measured, so AI-based models can train on them to interpret the world. 

At this stage, the stereotypical ideologies of the people that went into systems come to light. UTKFace, one of the datasets used to train AI organises data based on gender (male or female), age (numeric), race (white, black, Asian, Indian & other). While the intent for the models trained on this dataset might be different, the dataset is classified in ways similar to the Book of Life system used in the heights of apartheid in South Africa.

Captcha was an elegant application to involve the general public in this development process but beyond us, a vast and hidden labour hides behind monotonous tagging roles that help these AI-based systems sound smart. These efforts to draw bounding boxes around footage of roads to teach driverless cars what a tree, obstacle or a moving person look like, or tag content with feelings so algorithms can learn what a "sad" song sounds like, form the backbone of modern-day machine learning and is the what truly makes these systems smart.

One of the first natural language processing ideas comes from the 1960s, ELIZA, the model simulated conversation by looking for the keywords in a user’s statement and then reflect it back in the form of a simple phrase or question that gave users an illusion of understanding on the part of the program. The program was conceptualised as a method to show the superficiality of communication between man and machine but ended up convincing people of the system's unreal intelligence, a problem that plagues the present mindset.

The choices we make will forever resonate across disciplines that bind the world and mark us on a societal path, we won't be able to change.

Questions beyond the standard ambitions of improving model efficiency/ resource consumption, like the concept of fairness in machine learning are now driving the field to accept its limitations when it comes to ethical and moral choices that should be ingrained into these models for application on large public data in a world of loose legislation and imperfect political systems.

If looked at bias in AI-based models technically, the classic approach would be to write better algorithms and improve datasets, but the concepts of fairness, discrimination, inclusion being defined in arithmetic for the predictive models fails to capture comprehensive definitions that apply to varied disciplines and help AI-based models to build a realistic view of the world. The biased results of these models are the reflection of their errors in the perception of our world and society. Deleting biased association (scrubbing to neutral) a practice often used to improve model accuracy hasn't worked well to date, the idea to eliminate biased elements to make the model neutral raises the questions of the definition of neutrality, a realistic view of society, the collective socio-cultural conscience and the complex task to reflect humanities best version in these datasets.

The habit of AI-based systems to serve better outputs or favour certain population segments while targeting or even subordination of minority segments is a problem that took birth with humanity and has a more recorded history than probably any other topic. And the problem that should capture our imagination isn't the error in an algorithm's choices when it comes to judgements on empathy, accuracy, legality, social utility and equity for humans and the world we live in, but the choices we have made throughout history to reach this point and a vision of a world we intend to build.

Building a better dataset, accurate prediction models, or even a legal and pro-choice infrastructure of this field of study to manage human society won't solve the errors of human history which remain vivid beyond the inks of their words and amplify the wrong actions of the past into a brutally efficient AI-powered machinery with a default culture and behaviour to target minorities, especially in a world under the clouds of economic pessimism and the dominance of the right in public decision making.

Maybe not the precise strategy to solve this problem that plagues our present but possibly a good thought to consider would be actions of the Recording Industry Association of America (RIAA) from the 1980s when they inspired legislation to kill the Digital Audio Tape industry and formalised the introduction of copyright filtration system in the age of vinyl tapes. The actions rendered consumer digital audio dead for almost a decade and while the whole episode isn't right in any sense showed how through public legislation under parliamentary democracy, monopolies and profitable outcomes decided the timeline and road map of a mass consumer technology. While commending the system's efficiency of execution across a global population, the hierarchy of the system didn't fairly represent the role and choice of the general public on decisions of this technology which eventually lead to individual actions that revolutionised the music industry.

Till this hierarchy in any systems we built, fairly represents the role and choice of the general public in its social, cultural, technological, financial and spiritual future, the ghosts of bias will plague artificially intelligent systems into actions that will drive profitability but erode humanity.