During these days we have been seeing how news about the advances of artificial intelligence does not stop coming out. The trigger for everything was ChatGPT, which surprised many with its results, creating texts that have been useful to many people. After this, there has been a lot of expectation about the uses of predictive artificial intelligence, but here we will simply call it Artificial Intelligence.
They are sold to us as if they will solve problems, and create many things... But the reality is that we must be very cautious with everything they offer and sell us. We do not deny that it will be useful, but you must be very informed of the promises and dangers they have.
Table of Contents
To begin with, computers are not very intelligent, and therefore Artificial Intelligences are not intelligent
A professor of cross-platform application development taught me over ten years ago that computers aren't smart, but rather very stupid yet incredibly fast at doing things . Computing operates on the basis of binary switches, using 1s and 0s, ever since Ada Lovelace created the first examples of computer programs.
A calculator does not calculate how much 3+2 is. The value of number 3 and two are stored in memory, and it performs the addition operation by moving binary switches to raise two numbers to number 3, and thus obtains the answer: number five. And this is because a program has taught it to add numbers, and that's assuming the program has been programmed correctly.
Another great way to explain it is how Gretchen Grandler's Galileo minicomputer said it in an episode of the series La Banda del Patio:
«Computers are only as smart as those who program us.»
Galileo
Under that phrase, we can synthesize all the warnings about the limit of artificial intelligence. But we have to explain it, because there are many nuances.
ChatGPT and other interpretive language AIs are not perfect and are not search engine friendly
It is possible to think that, since ChatGPT has access to thousands of documents and websites to train its language and knowledge, it will have a high reliability. But the reality is that first, ChatGPT is trained on data from 2021 at the latest, and that an interpretive language artificial intelligence simply uses probabilities to say what words come next.
This means that, fundamentally, neither ChatGPT nor any of the ideas used by the search giants are useful for asking them questions about current events. For example, someone ends up having to ask about mobile phone plans on Reddit because they're using ChatGPT as a search engine :
What should be a simple task for a search engine, an artificial intelligence gives different answers every time you ask it. This is because it's using outdated data that an AI can't realistically interpret. This problem has an easy solution: use a price comparison tool, or visit each Spanish mobile operator's website to compare terms and conditions. But doing it yourself isn't as futuristic as having a text appear that provides a human-like answer.
We must also consider what it's fed, because an AI, we insist, isn't going to create anything new, but rather replicate based on the information it's been given. We discussed some time ago how the YouTuber PutoMikel made it clear that Midjourney and other image-recreation AIs were replicating stereotypes and even misinformation. When a thread of images was shown on Twitter a few months ago depicting how different historical periods would look through selfies, a considerable lack of women was noticeable , given that they have rarely been protagonists in history due to the inherent sexism of historical reenactment. And when showing the " Viking era ," Midjourney created an image of men with beards and glasses wearing the horned metal helmets that originated from a 1930s cartoon. These anachronisms are due to the fact that Midjourney was very likely fed images from Instagram and Pinterest categorized as " Viking " by people who are unfamiliar with the reality of Vikings and see them as something aesthetic, not historical.
When recreating ancient cities, other artificial intelligences focused on European aesthetics and rendered them without color. The reason is that they had been given primary images of Roman and Greek recreations that failed to account for the fact that statues and cities were once vibrant with color, which has since faded. These are errors that many historians would avoid , yet artificial intelligences are replicating flawed perspectives that have long been accepted and not corrected before being used as the basis for their images.
This makes it clear that interpretive Artificial Intelligence systems like Midjourney or ChatGPT will not accurately reflect reality or history unless they are given the necessary range of knowledge. And everything they provide must be reviewed by a human.
Autonomous driving by artificial intelligence is already more fantasy than science
One of the most anticipated uses of artificial intelligence is fully autonomous driving. The idea behind autonomous driving is that one day it will surpass human driving ability by a certain percentage, and since artificial intelligence doesn't get distracted or tired behind the wheel, nor would it confuse things, the time will come when anyone can drive a car without having to take a driving test and can go anywhere. There was speculation that humans on the road would eventually become illegal, except in a few isolated cases. Seven years ago, the YouTuber CGP Grey even suggested that autonomous cars powered by artificial intelligence would arrive and eliminate the traffic problem because a large neural network of autonomous vehicles could coordinate their movements.
For years, companies like Google, Uber, and Tesla have told investors that one day all cars on the road would be AI-driven. Elon Musk has been predicting since 2014 that Tesla cars " will be fully autonomous next year ."
Multiple US states have allowed self-driving cars to be deployed on the roads. After years of promises and millions of dollars spent on artificial intelligence for autonomous vehicles, almost everyone agrees that self-driving cars are not working as well as expected.
But the truth is, they never did. Tesla is being criminally investigated by the U.S. Department of Justice because the videos they made to promote their self-driving feature were fake. They also falsely advertised that their cars had Autopilot and full self-driving capabilities. This inspired some drivers to take their hands off the wheel, which led to fatalities. In 2022, at least 10 people died in Tesla's self-driving cars in just four months, which is why the government forced them to recall 300.000 vehicles.
All of this is due to unforeseen events on the roads. Animals, balls, children crossing, potholes, roadworks… Humans are good at improvising, but AI isn't. In the case of Teslas, there have been reports of them choosing to run over a child rather than swerve off the road, or disengaging before an accident to leave everything in the driver's hands so that Tesla isn't blamed on mere legal technicalities.
Self-driving technology has important real-world applications , such as collision avoidance and improved cruise control. But the idea of cars becoming our robo-taxis is already pure fantasy. Returning to CGP Grey's video, he says that Artificial Intelligence would solve traffic problems, but increasingly, the solutions lie in public transportation and personal mobility vehicles, given that many cars in traffic jams, especially weekday traffic jams entering and leaving cities like Madrid or Barcelona, are driven by people traveling to or from their commuter town.
A predictive artificial intelligence does not review the results, a search engine does
One of the promises made about predictive artificial intelligence is that you could ask it anything and it would answer without having to visit the websites from which it gathers the information. There has been so much concern that Microsoft has had to start thinking about a system for advertising and compensating content creators . But the reality is that this predictive artificial intelligence thing will most likely be a gimmick that will stop being surprising after a few months.
There have already been cases of errors with ChatGPT, the AI that Bing AI uses for its search results. And in the Google Bard demonstration, they made mistakes that wouldn't have been made with a regular Google search , ironically. Even Google employees themselves insist that this is something that shouldn't be released for the good of the company and its users.
Another popular example is when someone asked Bing when the movie Avatar: The Way of Water was released. It's a fairly simple question for a search engine. But Bing told a user it hadn't been released yet because it was still 2022. When the user replied, " No, I'm pretty sure it's 2023 because I'm alive this year ," Bing tried to convince the user they were lying.
The problem lies in the fact that search engines themselves must index information and categorize it as true or false. While there are indeed cases of " dark SEO ," these are very isolated and centralized. Google and Bing's job is to index information and remove erroneous information from their indexes . Predictive AI can easily fall prey to hoaxes or misinformation because its data feed is practically the entire internet. And as the cartoon character Arthur showed us, people can go online and spread lies.
We are currently experiencing a disinformation crisis, which is why websites like Maldita and Newtral have had to emerge. Their job is to point out false information, however harmless or ridiculous it may seem to many. Maldita had to issue an explanation for why they need to clarify what many people consider absurd: it's important to remember that everyone knows different things, and what is very clear to some may not be so clear to others.
For us it seems obvious, for example, that you cannot plant a bag of money and grow a tree of banknotes. but for an artificial intelligence, this data can be found and replicated without even checking it, because it believes that everything they are told is true,
ChatGPT learns in its own way, and it is not global knowledge
When a teacher's concern is that students would do all their class work, one of them suggested asking ChatGPT himself if he had written anything, as a way of checking if a class work was written by Artificial Intelligence. And just because he said yes, he believed it was a way of proving his authorship.
But when ChatGPT was asked if it had typed a piece of text, which it clearly hadn't, it said yes. This is a warning that no predictive language AI truly learns. Remember: it simply strings together letters and words it thinks make sense together. It doesn't check its sources or keep a history. In a conversation, you can convince ChatGPT that " 2+2=5 , " but that's something it will only use with you in that session.
Many things that AIs do have already been done, but they sell it to us as something magical
A few months ago, Spotify launched an AI DJ that creates a personalized radio station just for you. It would be exciting, if it weren't for the fact that this already exists , and it's what much of Spotify's business has been based on: recommending new artists based on your tastes. That was the algorithm, just like the artificial intelligence that already recommends YouTube videos, but we call it " the algorithm " because, for all practical purposes, they're the same: they learn from you automatically and give you a personalized result.
This is one of many cases where artificial intelligence is promised, but it's something that was already happening. This is because including "artificial intelligence" looks great on shareholder roadmaps , as it implies that you're not lagging behind. And this is crucial in the tech sector. It's the same as two years ago when NFTs became incredibly popular and dozens of websites announced plans to implement them, even if it was something as basic as using your NFT as an avatar. We can expect many websites and services to offer some form of artificial intelligence, and that this use will either be something that already exists or something minor. There's so much hype surrounding artificial intelligence that we might be promised things that won't be useful, or that could even worsen due to the dependence we'll develop on it.
Why are we impressed by ChatGTP?
Humans are programmed to understand language. That's why we have a native language. If we see a string of words that make grammatical sense, we naturally search for and find meaning, and we naturally tend to assume there must be a mind like ours behind it, even when all we're looking at is a jumble of words put together by a probabilistic parrot . As the researchers wrote, " The tendency of human interlocutors to attribute meaning where there is none can mislead researchers and the public into considering a synthetic text meaningful ."
And that's what makes linguistic models dangerous. It means we naturally trust what they say. And the potential harm is enormous. Bad actors could use them to generate mountains of disinformation that many find credible, and racist or sexist biases in the training data could spread widely and be reinforced because " it must be true, the AI said so ."
In 2021, Timmer Gebru and her colleagues wrote an article on the ethics of Artificial Intelligence use, warning of dangers that have begun to emerge. She worked at Google as an AI ethics specialist, and because of that article, she was fired for highlighting the ethical risks involved. And those risks of mass disinformation are now coming to light due to Google Bard, which employees themselves described as a public menace.
The need to moderate what is fed to ChatGPT translates into human labor and use without rights of content
A Time Magazine investigation recently found that to keep all the filth and slime of the Internet in a ChatGPT and DALL-E, OpenAI makes Kenyan workers screen for toxic content identifying which clips are too violent, sexually explicit, or disturbing to be used in the training data. These workers are exposed to
graphic scenes of violence and death, sexual assaults and sexual abuse of minors. And in return, they are paid less than US$2 an hour. This is because their entire business model is based on using the labor of literally millions of humans without paying them for their rights as creators.
So-called “generative AI” like ChatGPT and DALL-E can produce any kind of text or image you want. But they can only do this if they are given a huge amount of training data to copy from. And that data they are fed comes from humans. AI companies might argue that this is all fair use and therefore legal, but artists and authors disagree, and so they have banded together to sue them.
So far, there have been regulations, generally agreeing that permission should be requested from the people whose texts and images are taken. It has also been regulated that this content cannot be copyrighted because there hasn't been sufficient human intervention to establish authorship, which should discourage many people who wanted to produce large quantities of books or commercial images using Artificial Intelligence.
Artificial Intelligence is not bad, but it is not magic
We certainly don't want to dismiss any of the potentially useful applications of artificial intelligence. ChatGPT can do funny things, help with computer blocks, write letters with disposable content to save time, modify elements of an image without adding more work for the original artist… But we mustn't forget that everything they do will, at best, be an intermediate step and should not be considered a final product.
Midjourney has great potential for concept preview, and it's something that can be extremely useful in audiovisuals. ChatGPT can make the letters and communications of professionals who do not speak languages other than native and English. But those uses that are ethical and do not give shoddy results are few.
The best example of the limitation of something like ChatGPT is its result in a law exam in the United States: it passed the theoretical part but could not express itself in a way that was not basic, full of short sentences and with an extremely small vocabulary, and incapable of fine-tuning the small details and interpretations. In science fiction comics there is talk of robot judges capable of legislating in the most complicated cases, but that is already more fantasy.
Artificial intelligence can take us away from many repetitive mechanical jobs, or that serve very specific small details. But that at most will help humans do their job better, not take away from them.
