It might sound surprising, but AI does not have a universally accepted definition, yet. You might find some in the dictionary or Wikipedia, but in reality, we call something AI only if it feels ‘magical’. When a magic slowly becomes a ubiquitous product, we stop calling it an AI. Also, AI is an umbrella terminology, that can cover vastly different technologies.
The digital maps, for example, finding us the fastest, shortest, cheapest, least interrupted transit everyday, we do not call it AI. Because it does not feel magical anymore. Although it has all the elements that feels like an AI. Same goes for the predictive texts, our mobile digital keyboards are assisting us with writing the next word with a tap since more than a decade, it pretty much behaves like an AI, we do not call it one. Whereas, the LLMs like ChatGPT still feels magical, so it is a well accepted AI, at least for now.
I recently finished reading a book called “AI Snake Oil: What Artificial Intelligence Can Do, What It Can’t, and How to Tell the Difference.” by Arvind Narayanan and Sayash Kapoor. The book goes quite in depth in technical details, which can be fantastic for those who want to understand the real work behind the “magic” we call AI nowadays. Then they explore the social and humane nuances of AI, along with the future and the possibilities of AI in real world application.
Before going into detail, here are some disclaimers: research around contemporary AI is scars, specially the quality of research is heavily questioned and criticised in academia. Hence, it is still a topic of speculation. The field is filled with myth, fear, resistance, misunderstanding, and sometimes outright, lies. So, not just this piece of writing, or the book in discussion, anything around AI we have to be very careful, and take handful of salt.
There would also be the immediate frown by many about “who” is talking about it, in this case you might wonder, who am I to write about AI. The author of the book in discussion aptly interjects, anyone and everyone should talk about it for anyone and everyone. Since, there is a serious lack of verified facts in this field, while it is rapidly growing and occupying our day to day lives like a storm—we all should talk about it—to shape the present and the future of AI in our lives. Meanwhile, I code a little, and I built successful global tech products decades ago, and working with a technology consultancy firm for several years with experiences in building various types of applications globally. Although, I am definitely not an expert in AI. This is an anecdotal writing from the memory of the book. While the authors of the book are two Princeton University professors of computer science with vast experience of teaching and research on AI technologies.
Here are some of the key points they make in the book, written based on my own understanding,
The book talks about 3 types of AI mostly,
1. Predictive AI: where the technology tries to predict something, like the possibility of someone committing a crime in future, the possibility of someone catching a disease, how good someone might do at a workplace etc.
2. Generative AI: The technology of generating something by giving simple human language input. The way most LLMs now work, it generates text, image or multimedia responses.
3. Content Moderation AI: The technology of automating content moderation in only online social networks.
Here are some of the main points they make about these three types of AI:
1. Predictive AI
They are mostly snake oils, পুরাই সান্ডার তেল! Predictive AIs do not work now, and highly likely they never will. Some of the most commonly used predictive AIs are (currently in use, or in the near future, depending on where you are)
—Police are using AI to predict near future crimes in neighbourhoods, and about the decision of bail of a person in trial to save time and court resources.
Here is why it does not work: the system is inevitably trained by historical data, which invariably will put a specific demographic minority in the cross hair, no matter which country or culture it is applied in.
As human behaviour is not possible to predict, the system will continue to make mistakes about people who might decide to be a murid of a sufi saint the next day of their bail, or a clean slate gentleman chemistry teacher starting a meth lab in the backyard all of a sudden like Breaking Bad. No AI model ever can predict the uncertainty of human behaviour and decisions. None of the advertised and used AI tools work beyond a 50 to 60 percent accuracy, which is similar to a coin flip.
—Stock market prediction: Does not work, and likely will never work. For the similar reasons; stock market is vastly affected by completely random social, political incidents, and individual human behaviour. No amount of data, training and computational super power can predict something like a plane crash that will end up tanking Boeing’s share price. So if an AI product is claiming to be giving better decisions about stock trading than a human, you can be sure it is a snake oil.
Then of course, if AI is used to run specific algorithm with a statistical goal for data crunching, it can be pretty helpful, where you clearly know about the algorithm in use, data you are dealing with, and what are its limitations.
—Recruitment AI: It is probably one of the most commonly used predictive AI technologies now. HR departments receive thousands of applications, and they need an easy fix for sorting and selecting. The trouble is, there is no easy fix to this problem.
All AI products claiming to be able to find the best candidates from a pull of applications are most likely fake. It is a discriminatory snake oil similar to the policing one, and will continue to fuel the minority stereotypes in every society. No amount of data and computational power can predict the potential of a candidate for any job. It is simply not possible due to the same reason, human behaviour and potential are unpredictable factors.
Wait, can it at least filter out some quantifiable prerequisites and binary qualifiers of candidates, like if they have that particular degree, or the number of years of experience? Yes, they can. But that’s pretty much is the breadth of possibility AI has in this field. The rest of it, will continue to be a snake oil.
—Medical diagnosis: AI also miserably failed in most cases of diagnosis. It might be surprising to many, since diagnosis often seems like a data driven decision making, like comparing some x-rays and giving a decision, it actually involves much more complex humane and social nuances that an AI will continue to miss. And the training data statistics are often deeply misleading. We still need a radiologist, pathologist and a doctor—no, they are not going anywhere anytime soon.
In one example, the authors talked about a real life incident where a hospital was using a predictive AI tool to decide about whether certain incoming pneumonia patients should be released or admitted to hospital overnight. The AI was trained with old hospital data that indicates patients with asthma who are also suffering from pneumonia, are less likely to remain in hospital for long. The AI model started to recommend that patients with pneumonia and asthma combination should be sent home, they do not need to be admitted. But the actual reason behind it was, that the asthma and pneumonia combination is so dangerous that doctors took it very seriously, they put them to ICU immediately, and they made faster recovery—but the training data did not have this insight. Hence, when the AI was asked “who can be safely sent home,” it made the deadly mistake of marking asthma patients as low risk.
Similarly predictive AIs that relates to direct human behaviour or social nuances will continue to fail, like the ones that try to predict the outcome of an election, or how an organically grown city will look like in future.
So if anyone is trying to sell you a product that claims to effectively predict something like the examples mentioned above—run away from it. Or if you just want to test it for R&D, keep it in the R&D labs, and do not deploy them in real world application or invest a big capex.
2. Generative AI
They are great! And they have incredible potential to do more than what today’s GPTs are capable of. Go for them, use them, learn them, play with them, invest in them.
However, generative AI comes with some inherent danger that we already know about, like the rampant use of deep fake media, and exploitative use of someone’s persona, face or creative work. For those we need better local and global policies—immediately!
The book goes at length about how generative AI actually works. The text generation is mostly a massive computational work of pattern identification. It simply identifies which word has higher possibility of coming after which one given hundreds of parameters to calculate in parallel. No, it is not sentient, and technically it can not be.
And they need massive manual human laboured training to be better at predicting the next words in specific context. All big AI still go through these manual intervention trainings. These are mostly done by workers who are very poorly paid in gruelling working environment. That’s where the PT of GPT comes from, Pre-Trained, by humans, manually, one by one.
Same goes for image generation, it simply calculates the possibility of which coloured pixel is more likely to be next to which pixel in a given instruction. It mimics the understanding of an image. A big sophisticated permutation combination calculation that feels magical.
3. Content Moderation AI
It works in the background all the time as we use social media everyday. This social god decides if your content should qualify for going viral, or if it should disappear in the oblivion or outright be deleted. Content moderation is an extremely complex, dreadfully nuanced social, human activity. And machines are bound to fail in this sector, beyond some easy boolean selections.
Content moderation is another type of AI that can improve over time, but it might continue to fail in understanding highly contextual, local, cultural, social nuances. Since these social nuances do not have any pattern, they are continuously evolving at a high pace, highly subjective, and absolutely unpredictable.
Some examples from the book: Somebody posted a beautiful photo of a white family with a black child (likely adopted) in their beach-side house, with a caption “Everyone needs a pet.” This seriously derogatory and racist post slipped through the crack of content moderation AI because, if you separate the two artefacts, the photo, and the caption, both are absolutely innocent. Separately, the photo is just a beautiful photo of a family, and the caption is just another harmless sentence.
The problem begins when they are posted together. After some manual “reporting” it went to the human moderators, they too decided not to remove it, as the community guideline of the said platform does not specify anything about the implication of a novel meaning born out of a combination of a text and image. They may have extensive rules about what individual texts and images can or can not be posted, but there is no instruction about what to do when two totally innocent content becomes derogatory when combined together.
Then the debate about whether it is ok or not ok to post photo of the Napalm Girl (if you do not know the image, take a search, a photo of a naked young girl among a few children running away from the devastation of Napalm bomb in the Vietnam war). If yes, then it breaks several of most common community and safety rules of social media, while it is an important piece of history which we should have been able to share and talk about freely. How would an AI ever be able to understand these social issues and context? There is no amount of training data and computational super-power can answer these questions.
Thus, if an AI product is claiming to be a demi-god of content moderation, with some gimmicky results, be very careful. If you end up deploying tools like these without vigorous testing in a real community setting, prepare for some social havoc.
The Super AI and Why AI Myths Persist
The authors points out, a vast majority of researches about AI are questionable, some of them are outright fake, and most of them can not be reproduced or tied with inaccessible proprietary resources. At the same time there have been a surge of irresponsible books, articles and commentary by otherwise reputable authority figures.
Meanwhile, journalists are responsible for spreading sensationalised headlines mixed with unrelated images. Almost all news about AI comes with a humanoid robot, or some kind of si-fi demon, whereas, the AI is far away from either of them, and those images get etched into our psyche about the definition of AI.
The book also talks about the ELIZA Effect of technological innovation. ELIZA was one of the earliest chatbots, originally programmed in the mid sixties. When I first used a Windows computer in the late nineties, almost all household machines used to come with ELIZA pre-installed. I loved talking to ELIZA, albeit it was a rudimentary program that could mimic human-like conversations, it was fascinating to me back in the days. That exactly is what ELIZA Effect is, it means, we get fascinated by a technology, find it magical, and give it more credit than its actual merit. Currently we are in that bubble of ELIZA Effect about generative AI, we put undue trust and confidence in it.
And from there, the ideas of Super AI, AGI (Artificial General Intelligence) are fuelled. They are not incorrect thoughts, it is not like there is no possibility of them coming, but the current AI models we use are far behind.
Most commonly known or used AI products today are generative pre-trained models, which is yet to go to the level of effectively self-learning and reconstructing itself without human intervention. A line quoting directly from the book “we should be far more concerned about what people will do with AI than with what AI will do on its own,” forget the fear of what AI will do to us, we are the bigger danger ourselves.
Regulations and Adaptation
Finally the book talks about AI policies and laws with great importance, the debate about whether to regulate or not to regulate, if we regulate, how much? The authors recommend, we should immediately move towards strict regulations about certain types of AI globally.
A vast majority of the corporations who are selling AI products would of course try to sell the idea of having no regulations in the name of freedom of innovation, but the dangers of having no regulations at all is too high to cater to any such request. In some instances, some big tech companies tried to manipulate the government about a regulation for limiting the number of parties who can build AI tech, that is even a bigger danger, stop any such conspiracy vehemently in any country. Because that will help no one rather than companies turning into monopolies.
The technology should be open for anyone to play with—while there should be strict universal regulations in place about certain types of use, like the generation and use of deep-fakes of real human characters, replacing humans with AI for critical decision making that requires nuanced human intuition etc. We need these regulations “yesterday,” while remaining careful about keeping the door for innovation and technological advancement open, and never to allow any exclusive right or benefit to any corporation.
AI models should be more open for scrutiny. AI researches need to be replicable, reproducible, while the academia needs to hold their horses about jumping on the bandwagon. They need to be strict about upholding the academic integrity while teaching and researching AI. It is very easy to derail by the hype, myth and the magic.
I highly recommend the book. It would be a timely and necessary read for anyone.
