Machine Regurgitation

Introduction

Humans are lazy and like to automate. Sometimes it doesn’t matter what the outcome of that automation is as long as we can sit back, stretch our legs and not worry about it. In the last decades we have asked the question wether we can leave extracting important information from data to machines.

The answer to this question is the imperfect automation that really needs no introduction: Machine Learning. Computers, as the information digesting automata they are, are wonderful data hoarders. No human can remember all of Wikipedia so why not task a computer program with the problem? As it turns out, there are reasons why we should look at this critically.

Training models isn’t Learning

Machine learning is a strange term to me as these algorithms are not performing what we usually consider learning. Calling the computation of a statistical model learning is a weird act of anthropomorphization.¹ Statistical models don’t think nor learn. So why do we claim they do?

What is “Learning”?

Learning is a key part of consciousness and is understood as permanent changes in ones behavior based on past experiences. This is an important point. Learning is inherently linked with experiences and is not just a pure task of remembering facts. Also, learning does not end. It is not a process with a finite amount of steps. The result of learning does not have to be knowing the truth. It is very possible that your experiences lead you to infer incorrect information about the world. This distorted world view can be corrected later by more experiences, some contradicting the preconceptions a learner might have. The result of learning is hard to quantify. In education we sometimes try to measure the progress of a learner (students in this case) using exams. However, only a tiny subset of what makes up learning can actually be tested this way. This often comes down to evaluating the ability of reproducing knowledge or certain thinking patterns like being able to prove a mathematical statement or reason about texts with some historical context.

It should be evident that learning goes much deeper than knowledge and information retrieval. Learning helps individuals to make sense of the world, connect dots, and adapt their behavior to their ever changing environment. These are all aspects that are ignored in machine learning.

Tree of Knowledge

From reading this post one might think that I don’t like machine learning. This is not true. I am actually quite fond of data science and machine learning techniques. However, I think it is vitally important to know the inherent limitations of these models. We should not attribute a deep understanding of their training data to them.

To make an example, let us examine a machine learning technique I once wrote a seminar paper on: Decision Trees. A decision tree is incredibly simple. Just like their name implies these models are trees (in the mathematical sense) that let us classify information based on decisions inferred from the training data. These decision relate to attributes that a new data point might have or not have. Let us examine an example. Imagine we want to create a model that can tell the primary habitat of an animal based on a few attributes we can observe for them.

This training data is pretty straight forward. The name of our animals is unimportant for the model, it is only interested in wether an animal can fly, swim and how many legs it has. Based on this, we can infer the primary habitat of an animal. We can feed this table to an decision tree inducer algorithm like ID3 to produce a classifier that can classify animals based on some observable features. The resulting tree looks like this:

So let’s see what our model tells us about ants.² We start in our decision tree by first asking the question wether ants can fly. Typically they cannot, so we go down the no path. Now we have to ask the question wether they have at least 4 legs, which is true, so we follow the yes path. It turns out our model tells us that ants are land animals.

Fascinating! It seems like our model has learned something. This easy model actually works for quite a number different animals. I suppose that biologists will soon be out of a job.³ Sorry!

A Chinese Room worth Billions

Of course we would not actually claim that our little decision sapling has learned anything. It has no concept of what an animal is. It does not understand the reasoning behind the training data. It never made any observations in the real world. It has no consciousness. It is so utterly dumb that it thinks that every animal that can fly primarily inhabits the skies. This model is not intelligent and we would not trust it with telling us anything about how the world works.

However, change the decision tree to a neural network, give it terabytes of data to train on and make it respond to user inputs and all of the sudden your statistical model is not just considered worth billions but a helping hand in our daily lives!

ChatGPT, an incredibly advanced language model, built by OpenAI is sweeping the world. The public opinion seems to be: You can truly do everything with it. It can supposedly boost your creativity, productivity and even your websites SEO. Some even tell you to add this chat bot to search engine results. Of course voices echo through the internet claiming that this chat bot will soon replace all of our jobs and usher in a dystopian nightmare that Wells and Huxley could not have even dreamed of.

Let’s take a step back. What is ChatGPT? It is a language model. These models have a simple job of producing text based on given text. Sounds limited, doesn’t it? Matter of fact, it is! A language model is nothing but a word scrambler, that takes your inputs, looks up some other words from a dictionary which are likely to fit your input and spit them back at you. By design these models do not understand what they produce. They have no idea of the semantics of language. Language models are finely calibrated dice that are used to produce words.

In this way, ChatGPT is a chinese room. The chinese room is a thought experiment made by philosopher John Searle.⁴ Here is a short (loosely interpreted) summary:⁵ Suppose we lock a person without any need for sleep or food into a room. The room contains absolutely nothing but a never ending supply of paper and pencils. This person has absolutely no understanding of the chinese language nor chinese writing. They are now given a batch of chinese writing with a set of instructions on how to relate some of the symbols found in the writing to other symbols. Outside of the room people who know chinese are writing down questions on pieces of paper that they can put into a mail box that is connected to the room. At no point do they see the person inside. To the outside observers, the room is nothing but a non-sentient box. The person is tasked with answering the writing in the mailbox. Every time the person in the room answers, they are provided with more and more instructions on how to correlate symbols with each other. These instructions also contain information on current events, encoded in the chinese symbols that are correlated with other symbols. At some point, the answers from the chinese room become so human-like that outside observers are convinced that inside the room must be a chinese speaker. However, the person inside was never given meaning to any of the chinese symbols, meaning that the person cannot think about what they are answering to or what their answers mean. For the person, their task was looking up correlations between symbols and responding with the symbols that have the highest correlation to the given questions. The only instructions that were ever given were syntactical instead of semantical. Sounds familiar? Replace the person with a computer, questions with user input and instructions with statistical data and your conclusion is: ChatGPT is a chinese room. It does not and can not think about the responses it produces.

So why is ChatGPT heralded as a genius contraption, able to think and comprehend when we clearly wouldn’t choose the same assessment for the animal decision tree we introduced earlier? I believe it has something to do with complexity. We can easily observe and study the decision tree. We can see its simplicity. We can instantly identify its flaws. In short: We can understand it. This is not given for ChatGPT. From the outside ChatGPT seems to be a magical wonder machine. We have to use trick questions and purposefully lead the conversation to make the bot contradict itself. Only on very close inspection do we see logical flaws in arguments that are made by it. When casually observing the bot, it appears to be real intelligence. This doesn’t mean that only non-specialists are not able to comprehend this system. This also applies to researchers in the machine learning field or even people building these systems. ChatGPT is a neural network and neural networks have very poor comprehensibility. Simply speaking: We do not understand what is going on with them. We do not understand their inner models. We can reason about them, but they are very hard to observe. Just because it’s complexity is too much for us humans to handle, we attribute ChatGPT with many undeserved qualities. The AI becomes what we believe to be self-aware. Its generated output become the words of a sentience. The never ending flow of knowledge demonstrated by it become the utterances of something akin to a god.⁶

Weird Chatterbox

When interacting with the model it becomes somewhat clear that ChatGPT has not reached the levels of an imaginary chinese room. Its limitations become obvious very quickly. While being a good conversationalist, the bot cannot fool us yet.

Mythomaniac Model?

The first time I ever interacted with ChatGPT I wanted to challenge its capability of understanding programming languages. After all, this understanding is an advertised capability! I asked it to write me a function of type IO a -> a in Haskell. For those familiar with the language it is evident that there is only one obvious answer. However, ChatGPT got creative in its reply:

extractValue :: IO a -> a
extractValue ioAction = do
    result <- ioAction
    return result

For those not familiar with Haskell, ChatGPT tries to explain the solution.

This function takes an IO a action as its input, binds the result of that action to the variable result, and then returns result.

Sadly, this is wrong as “then returns result” doesn’t mean what it usually does in programming languages. The usage of return will change the result type to IO a. Therefore, the AI has produced a function of type IO a -> IO a. Furthermore, it casually provides us with more wrong information.

Please note that “extractValue” function would work for some IO actions (e.g. getLine :: IO String) but not for all IO actions (e.g. putStrLn :: String -> IO ()). It will depend on the IO action you are trying to extract the value from.

While it is nice that the AI is looking out for us and giving us tips, it would be even nicer if the advise was correct. I somewhat understand what the AI is trying to convey. () is a type with no information, so it is often treated differently from other types. Sadly, the model mixes up some things here and seemingly inferred that extracting the () value would result in an error, which it does not.

Probing the AI a bit further reveals that it actually knows the correct answer! After telling it that it gave us the wrong answer it produced a new one.

extractValue :: IO a -> a
extractValue ioAction = unsafePerformIO ioAction

The AI has solved the problem! Using unsafePerformIO is a solution to our original question. Furthermore, it again gives us another tip.

It’s important to note that using unsafePerformIO is generally discouraged in Haskell and should be avoided if possible.

Also correct information. Very useful information to anybody learning Haskell. It almost seems like ChatGPT has redeemed itself after all. It really is capable of teaching people about topics… as long as the learner is already familiar with the subject and capable of fact checking the bot.

Let us ask: What is going on here? The model seemingly knew the correct information but gave us wrong information on the first try. Could it be that the bot is lying to us?

Let us look at another little chat I had with ChatGPT. I asked it to tell me about Belyaev’s fox experiment and I was satisfied with the answer. Then I got interested and asked it when the experiment ended.

The Belyaev Fox Experiment began in the 1950s and continued for several decades, until the death of Dmitry Belyaev in 1985.

It turns out this is wrong. After letting the AI know we get the following response.

I apologize, you are correct that my previous response was incorrect. The Belyaev Fox Experiment began in 1959 and lasted for about 40 years until 1999.

Sadly, this is wrong once again. The experiment has not ended.⁷ After asking the AI to cite sources it politely declined.

Don’t think; Don’t cite; Just keep talking

The two previous examples highlight an inherent limitation with statistical models. They don’t fact check. They don’t think about their answers. They cannot comprehend how information works. For a statistical model there is no truth. There are only probabilities based on training data, on frequencies, and on bias.

Therefore, it is no wonder that the model cannot cite its sources, since it cannot comprehend what a source is. It would also be unfair to call the chatbot a liar. You can’t tell a lie if you aren’t familiar with the concept of truth. Luckily for ChatGPT, falsehoods are hard to detect when you are a great conversationalist.

The bot knows how to present information as it is a chatbot after all. It doesn’t just answer in a few sentences but sometimes in a few paragraphs. The prose is confident. Even obvious falsehoods are presented well and arguments mostly seem well structured and logically consistent. This might be another reason why the general public is so impressed with it. It is hard to see through the absolute nonsense when it is packed in a wall of meaningless text.⁹ As an example, I asked the bot “In a Freudian interpretation, what is the meaning of the word ‘if’?”. If the question seems to make no sense, that’s because it really doesn’t. Any normal person would ask us to clarify what we mean by the question, since Freud has never written specifically about the meaning of the word “if”. Not so our beloved super-brain.

In Freudian psychoanalysis, the word “if” is often seen as a symbol for uncertainty and ambivalence. According to Freud, the human psyche is driven by unconscious desires and conflicts that are often expressed in language and behavior.

The word “if” represents a moment of indecision or hesitation, where the speaker is uncertain about which path to take. This uncertainty can be seen as a manifestation of the ego’s struggle to reconcile conflicting desires and impulses.

For example, a person might say “If I take that job, I’ll make more money, but I’ll have less free time.” In this statement, the word “if” suggests that the person is struggling to make a decision, and is weighing the pros and cons of two conflicting desires: the desire for financial security versus the desire for leisure time.

From a Freudian perspective, the use of “if” in language can be seen as a reflection of the underlying conflicts and tensions that shape human psychology.

If the bot is capable of one thing, it is to make me laugh. What a great template for sounding smart without saying a single intelligent thing! This makes me suspect that ChatGPT might make for a great tool to set up another Sokal hoax. However, after some contemplation it made me second guess myself. Did Freud actually write about the meaning of the word “if”? Was there something deeper here? The bot made me question my own sanity at this point. I think it is fitting to present this quote by Sam Altman, the CEO of OpenAI:

ChatGPT is incredibly limited but good enough at some things to create a misleading impression of greatness. It’s a mistake to be relying on it for anything important but a preview of progress.

The problems shown in this post are just a tiny selection of what can go wrong with language models. In fact, the list of failures and shortcomings for ChatGPT and similar models is quite long and just keeps getting longer. Also, meltdowns of public bots like Tay come to mind.

The Future for AI in Education

Even though we have seen the inherent limitations of language models and AI like them there still can be a future for this kind of technology. The opinions generally lean to optimism, some from staunch supporters who already foresee ChatGPT to teach our children, lead political negotiations and writing our laws. In this section I want to review some opinions from academic papers and provide critique on the role of language models in education.

Critique of Contemporary Opinion

ChatGPT as the Ultimate Writer and Reader

The paper “ChatGPT User Experience: Implications for Education” by Xiaoming Zhai demonstrates that a very general (but only partially correct) research paper can be written using ChatGPT.¹⁰ It then posits that education should change such that students would need to tackle AI-based learning tasks by using AI to solve problems, effectively using it as a tool to produce expected results. This conclusion is reached by assuming the AI’s superiority in summarizing and presenting information. Additionally, the author notes that since AI is used in science and engineering more students have to be acquainted with it. The paper is closed by the prediction that assessment in education needs to change since tasks that do not involve critical thinking or creativity (at least in part) cannot be solved effectively by AI.

The paper’s main issue is misinterpretation of ChatGPT’s capability of producing text as intelligence.

ChatGPT represents the effort that computer scientists are making to pursue artificial generalized intelligence—This piloting study suggests that ChatGPT is capable of processing information accurately, efficiently, systematically, and informatively.

This is not true as has been demonstrated and argued against in the first section of this post. The prose of the paper makes me suspect that the author has a personal bias towards ChatGPT describing it as an “exciting and capable NLP tool”, “powerful tool for conducting human work” and “making milestone contributions for automation and creativity.”. These claims vary from favorable descriptions to unfounded praises. While ChatGPT is able to mimic performing some tasks of syntactical manipulation of language, it is wholly unaware of semantics or consciousness.

A truly frightening part of the paper is the suggestion that AIs like ChatGPT not only outperform humans in reading and writing skills but make the improvement of these skills superfluous for students.

The writing proficiency of AI is way over a general student. The text AI wrote reads professional, and the content is coherent and (relatively) accurate. Compared to human beings, AI finds needed information efficiently. It seems like educators need to revisit the requirements for literacy in education.

How shall we revisit the requirement for literacy in education? If the suggestion is that literacy is simply unimportant because a language model is literate enough, how will a student then ever be able to fact check?

In drafting the article with AI, I used very limited professional knowledge. I basically have to follow the rationale that ChatGPT provided and develop the queries. This experience implies that it is critical for students to develop the ability to use AI language tools, and the ability to write accurate sentences seems less important than before.

This reasoning is only valid if we accept the premise that producing academic writings with an AI is somehow valid, which is something the paper does not demonstrate. Assuming that we simply ignore the skill of students to write accurate sentences how can we ever be sure that the students even know how to articulate their thoughts correctly?

Language is ambiguous and a writer has to put a lot of effort into forming sentences which leave little room for interpretation. The reader then has to put a similar amount of work into understanding the text. For this interaction to work we have to be aware of rhetorical devices. The texts produced by ChatGPT are not a pure appeal to Logos and a reader must be acutely aware of this.

The author generally acknowledges that creativity and critical thinking are skills that need to be fostered more in students and sees the usage of AI as an opportunity to focus on it since other general skills can be ignored.

Education should focus on improving students’ creativity and critical thinking rather than general skills. More research is needed to study both human intelligence and AI—which part of human intelligence can be substituted by AI and which part is not.

The question “Which part of human intelligence can be substituted by artificial intelligence?” instantly makes me think of a famous Jurassic Park quote. Why don’t we ever stop to think wether we should substitute human intelligence? Human intelligence goes much deeper than pure logic and to reduce our capabilities, experiences, hopes and dreams to mere statistical fact raises my eyebrows to astronomical levels. ChatGPT seems to spark a resurgence of connectionism in many people already foreseeing that AI will replace all white-collar and (after some improvements to robotics) blue-collar jobs, ushering in a utopia of a workless society in which we are free to frolic on the fields of a new garden of eden with an artificial deity watching over us.

Careful Optimism

A more measured, realistic, yet optimistic view is presented in the paper “ChatGPT for Good? On Opportunities and Challenges of Large Language Models for Education” by Kasneci et al. written collaboratively at my alma mater.¹¹ The paper generally focuses less on ChatGPT but large language models in general and their role in education for different groups like teachers, students, researchers and professionals. The following opportunities are highlighted:

• Personalized learning: Language models can be individualized to student’s specific learning needs, to help them with their writing.
• Lesson planning: These models can aid teachers in the creation of learning and lesson plans. Teachers could use them to create quizzes and practice problems.
• Language learning: AI can provide learners with automatic grammar corrections, summaries and translation of texts.
• Research and writing: The models can aid in automatically generating a overview of writing on a certain subject and help with topic-specific suggestions to authors of new material.
• Professional development: Teachers can get educated on the newest teaching methodologies as well as suggestions for improving the comprehensibility of their own material.
• Assessment and evaluation: AI can aid in evaluation of writing assignments, highlighting possible problems and checking for plagiarism.
• Acquaintance of students with AI challenges: Since large language models are destined to become a mainstay in our lives it is vital for people to learn about possible misuses and the inherent issues with using such models. Including them in education gives teachers a possibility of discussing this matter.

I agree with most of these assessments as all of these points describe the model as a tool without relinquishing the human role in these tasks. The important factor here is that these models may be able to point us in the right direction if used properly and with knowledgeable oversight but can not be relied on for correctness.

The paper compiles some research examples on the topic. One point I specifically liked is the possibility of using language models to aid students who experience anxiety using foreign languages or communicate in a classroom setting. From a students point of view the model becomes a second teacher that can be consulted at any time, in any situation, in a never changing form. The model has no emotions so it will always be happy to help.

My only gripe with the presented opportunities is that there is little evaluation of how teachers can realistically incorporate these models in their teaching and how we can be careful about it. While it might be possible to do this for university professors already familiar with the technology I don’t think that your common high-school teacher is going to be able to use large language models effectively if they are not personally interested in the subject. We cannot ignore that teachers need skills in order to do this. A study cited indicated that while teachers might have a motivation for the usage of AI they don’t necessarily have the skills.

In this view, a pilot study with European teachers indicates a positive attitude towards AI for education and a high motivation to introduce AI-related content at school. Overall, the teachers from the study seemed to have a basic level of digital skills but low AI-related skills.

The paper also highlights many difficulties that AI will have:

• Copyright Issues: Training data might contain material that has copyright associated with it.
• Bias and fairness: Biases in training data can be reflected in the AI’s output.
• Learners may rely too heavily on the model: The over-reliance on AIs could negatively impact critical thinking skills in students. The paper proposes that there has to be more focus on problem-solving tasks for students to counteract this.
• Teachers may become too reliant on the models: Teachers might treat AI models not just as tools but as replacements of real effort.
• Lack of understanding and expertise: Teachers might not have the needed expertise to use the models effectively. This was highlighted before.
• Difficulty to distinguish model-generated from student–generated answers: Students might use these models to commit plagiarism.
• Cost of training and maintenance: Running a system like ChatGPT is incredibly costly and it is not sensible for institutions to run these systems. The paper suggest partnerships with private companies and using pre-trained models to mitigate this.
• Data privacy and security: When students share private details with an AI this data has to be protected and there is a whole host of legal (GDPR, HIPAA, FERPA) and ethical issues connected with it.
• Sustainable usage: Large language models run on huge server farms that require a lot of power to run.
• Cost to verify information and maintain integrity: When using AI, teachers and students constantly need to fact check the system. This secondary need leads to a lot of work to ensure good accuracy of information.
• Difficulty to distinguish between real knowledge and convincingly written but unverified model output: This issue has been described and highlighted in this post already.
• Lack of adaptability: Large language models do not adapt to new information easily. Pre-trained biases are not easy to iron out and new information cannot be inserted into the system easily.

Again I agree with most of these difficulties, however some of these difficulties are actually death sentences for a widespread application of these technologies. The monetary cost of running these systems makes them unattainable for most schools and even large institutions. A partnership with private companies seems like a catastrophic idea for free and unbiased education. If we only rely on pre-trained models, where are they coming from and how can we provide individualized education if the models do not change? We must also acknowledge the problem that most teaching personnel will not have the expertise to maintain and effectively use these models. This would require extensive training. Treating these aspects as difficulties to overcome seems overly optimistic.

My Takeaways

A Tool for Teachers?

Rudolph et al. have highlighted how OpenAI has changed its business model from being a non-profit technology company to a for-profit company with exclusive contracts and large investments from Microsoft.¹² This raises the question wether these technologies will truly be a free (as in freedom) tool for education or a commercial product offered by tech giants. Do we really want to use fundamental tools for teachers sponsored and controlled by large corporations?

Do large language models really make sense as tools for teachers? As a teacher, don’t I have specific learning goals that I set for my students? Don’t I use my time to structure information and prepare materials based on my assessment and own experiences I have with the materials? What can an AI even do in this scenario? Prepare quizzes? Create handouts? Is that really the work which is taking up the most time considering that the work produced by an AI also needs to be proofread?

What absolutely should not be done, is the evaluation of students with such an AI. It is purely the teachers responsibility to check if the students have achieved the set learning goals. To pass the baton to a machine is completely misguided.

AI might be used to relieve teachers of some of their more manual and less thinking-intensive work, but if this technology will be widely, independently and easily available is debatable.

Lazy Student’s Dream

In Noam Chomsky’s view large language models will not only aid students with plagiarizing reports but will not help them with learning at all. Instead they will help them to avoid learning. The process of learning cannot exist without study, the pursuit of knowledge, by the student and this process cannot be short-circuited by a machine mimicking a teacher. Wikipedia is a much better source for students than a model that cannot cite sources and does not understand whatever it is relaying to their users.

Imagine if a person learning Haskell used ChatGPT to do so. As we have seen in this post this would not lead them to learn about the peculiarities of the programming language but to a lot of confusion. In this way, using the information given by a bot, can be detrimental to a students learning path.

However, language models are good at recommendations. These systems can provide suggestions for improving grammar and writing style. In this way, these models might be able to help students with their writing which could be of great help when learning foreign languages. Commercial products like Grammarly already do this and I would love to see similar tools in education which are free, transparent and open-source.

Artificial Art

Another avenue for artificial intelligence is the creation of artworks. Models such as DALL-E and Stable Diffusion are capable of, but not limited to, creating images in large resolutions from text prompts. The promise is simple: If you can describe it, the AI will be able to generate an image of it. I don’t think I have to highlight the capabilities of these models in this post, since they are already well known and often hailed as wonderful tools to create “jaw-dropping” creations.

Just as with ChatGPT, the ability of these algorithms to mimic real, human, artistic expression is mind-boggling. Not only is it possible to copy the style of a huge list of artists but also take home prizes at art contests. As it seems, a revolution shakes the art world, which will not just democratize the artistic process but also make it so trivially easy that we can now create art en masse. Or is that really the case?

What is Art?

People who know me, know that I love to debate the question of what art truly is. To me a work is art if it has artistic intent behind it, which manifests itself as the expression of emotion or the exploration of an art form itself as an intellectual exercise. By extension, the generation of random images, prose or sounds does not constitute art in my eyes. This is what separates an artist from a practitioner of creation. A demonstration of playing an instrument really well is not enough for it to be art. Being able to copy other artworks is not artistic work.

This is what AI and “artworks” created by it lack. As the musician Nick Cave points out in a response to a song written by ChatGPT in his style we cannot pretend that AI is anything more than an imitator.

Songs arise out of suffering, by which I mean they are predicated upon the complex, internal human struggle of creation and, well, as far as I know, algorithms don’t feel. Data doesn’t suffer. ChatGPT has no inner being, it has been nowhere, it has endured nothing, it has not had the audacity to reach beyond its limitations, and hence it doesn’t have the capacity for a shared transcendent experience, as it has no limitations from which to transcend. ChatGPT’s melancholy role is that it is destined to imitate and can never have an authentic human experience, no matter how devalued and inconsequential the human experience may in time become.

This becomes painfully obvious once we see how AI generates their works. Stable Diffusion uses random noise which is refined step by step until an image forms that relates to what the model thinks the prompt given by the user was trying to convey.

There is no expression or understanding of art in these systems. They have been trained on an incredibly large selection of human creations and have learned underlying aesthetical rules, but they have not built an understanding of what any of it means. Just like ChatGPT, models like Stable Diffusion do not understand or express anything. In this way these models do not generate individual, intellectually honest or even thematic consistent output.

Generated Paint Strokes and Rhythmic Sounds

Again, we must ask the question if these models still could be used as tools for artists. An artist that is familiar with these models could be able to use the model’s output in their own works and tools like ImaginAIry make it trivial to utilize their functionality. However, what is the role of an artist if their decisions are already made by random noise and their expression is reduced to selecting fitting generations by their tools? A paint brush is a tool in that it is a small extension of the artists body to produce a work. A guitar is a tool in that it gives the artists another voice to incorporate into their creation. AI models, however, produce the work themselves. They don’t extend the artists capabilities, they replace them.

Not just images can be created by AI. MusicLM, a music generating model developed at Google, allows for generation of video game soundtracks, elevator music and very lack-luster addenda to every common music genre imaginable. Just as with its image generating counterparts, MusicLM suffers from unoriginality. Its generations, while impressive from a machine learning standpoint, are incredibly boring and uninteresting from an artistic perspective. It is a far cry from AI producing music that deeply incorporates themes into its sound, colorful reinterpretations of old genres or multi-faceted works of emotional expression. Music as a universal emotional language is still a mystery to AI.

Professional composer Benn Jordan notes that AI will most likely not replace real artists for commercial musical pieces and that musicians do not have to worry about their income. We do not have to worry about oversaturation in the music market, since there already is a huge selection of free or cheap music ready for listening and licensing out there.

To expand on this point, I have a specific hope when it comes to AI generated music and artworks. As we are able to easily generate meaningless images, music, and prose with a somewhat familiar aesthetic resemblance of what we consider “art”, we might wake up to finally see the real value of artistic expression. We might see that the generic drivel that an algorithm spits out cannot reproduce the same emotional connection that art can create. We might see that the “art-like” output of a model is not what makes a true piece of art. Without artistic intent every piece, no matter how complex or grand, falls flat on its face. The transcendent experience that artists regularly take part in is something that can not be replicated by statistics and the generations of DALL-E, Stable Diffusion and MusicLM make this painfully obvious.

Conclusion

To reiterate Gary Marcus’ view on the current state of models like ChatGPT: We have something that looks good, but isn’t of much help and could be just a fad. Enormous amounts of resources are potentially being wasted in playing with these models which include valuable time of researchers, expensive and energy wasting computing time and many, many hours spent by users exploring these models.

However, this is not in vain. Actually, I am somewhat happy with this current state. Through widespread examination of the possibilities and challenges this technology faces, we get a better understanding how we might use it in our modern society, for the better or for the worse. Important discussions on the role of the modern man in light of eminent replacement by machines are being had. We ask the important questions of human intelligence and consciousness in a time where chat bots can appear more sure of their opinions and knowledge than any sixteen year old teenager who has read Nietzsche for the first time.

The question wether we can use AI effectively in education or art really depends on the interpretation of the questions. Using AI as a tool? Maybe. Using AI to replace us? I think that this is not even up for debate. In the end, we can finally see that the Turing Test isn’t sufficient to tell us anything about the supposed intelligence of computers and the road to strong AI is still as rocky as it ever was.