{"id":4013,"date":"2024-09-02T11:27:50","date_gmt":"2024-09-02T09:27:50","guid":{"rendered":"https:\/\/www.rivistaeco.com\/?p=4013"},"modified":"2024-09-02T11:27:50","modified_gmt":"2024-09-02T09:27:50","slug":"ai-imitates-us-but-does-it-truly-understand","status":"publish","type":"post","link":"https:\/\/www.rivistaeco.com\/en\/2024\/09\/02\/ai-imitates-us-but-does-it-truly-understand\/","title":{"rendered":"AI Imitates Us But Does It Truly Understand?\u00a0"},"content":{"rendered":"

AI language models continue to astonish us with their ability to mimic human language. But comprehending a language is an entirely different matter. What can a Chinese room, a lion, and a judge teach us about the differences between human language and that used by machines?<\/span><\/i><\/p>\n

Natural language processing (NLP) has made significant strides in recent years. Instead of relying on simple, carefully crafted sets of rules, modern language models infer meaning from vast amounts of data and can be used to create translation services, predictive text\u2014such as when a smartphone suggests words\u2014and other applications. However, it is a mistake to believe they can understand language and its nuances. As many academics do, I have often changed countries of residence and have always tried to learn the local language with varying degrees of success. But what exactly does it mean to “understand a language”? I speak my native language, German, and English, and I can understand Italian. In high school, I also learned French, but I couldn’t pass any significant competency test. How much French did I have to forget before I could no longer speak it? The question is relevant for the entire field of artificial intelligence and particularly for that subset concerned with language.<\/span><\/p>\n

What the Chinese Room Experiment Teaches Us<\/b><\/h3>\n

Let’s start with an old experiment. Imagine you are given a new job assignment: you sit at a desk in a small room with a mailbox in front of you through which a piece of paper with writing you do not recognise is delivered. Your task is to respond to this message with the same set of symbols. You have various resources (books, databases, image recognition) and detailed instructions on how to use them. You must not only match each symbol to a character or word but also rearrange things. After a few weeks, you realize you have become good at it. At some point, your boss reveals that the symbols are a newly discovered language. Would you say you can speak or understand that language? Probably not. You have essentially acted as a language model: you received input, applied an algorithm, and produced output. In the 1980s, this \u201cChinese Room Experiment\u201d by American philosopher John Searle sparked intense discussions within the AI community. For early researchers working in this field, it was important to know whether the models they wanted to build behaved only \u201clike\u201d humans or if they actually possessed human capabilities. The two options implied very different goals and approaches, and it appeared important to clarify what kind of intelligence represented the \u201cI\u201d in AI. As soon as language comes into play, we are understandably tempted to think of AI models in human terms because language is one of the most essential human capabilities and is central to our experience. We automatically end up attributing feelings, motivations, and a rich inner life to the model. But there is no reason to do so.<\/span><\/p>\n

How AI Fills the Gaps<\/b><\/h3>\n

You have probably heard the term \u201clanguage model\u201d in relation to AI lately. You may have also interacted with one knowingly or unknowingly. They are used to improve internet searches (or so companies using them claim), generate images of a sunset made of sardines or videos of falling kittens, enhance programming, and converse with lonely people. But what are language models and how do they work? AI language models, much like the worker in the Chinese Room experiment, have been trained on many inputs without knowing what they are. They were created to play a simple game of “fill in the blank.” For example, if we read the sentence \u201cThey went to the park and drank a ___\u201d we expect the next word to be a beverage. We wouldn\u2019t be surprised to see \u201cspritz,\u201d \u201cbeer,\u201d or \u201ctonic water,\u201d but we wouldn\u2019t accept \u201cfreedom,\u201d \u201cbicycle,\u201d or \u201clamp.\u201d AI language models are very good at filling in the blanks. They were trained to do so, drawing on far more information than even the most studious human could absorb in a lifetime. However, a trained AI language model can do more: one of the most impressive aspects is its ability to continue generating text. It can take a sentence and predict the first word in the next sentence and the next until it has a new complete text. After all, the first word of a sentence is simply a continuation of the previous one. A language model can also provide answers, label data, translate from one language to another, and generate code snippets. However, these abilities are based on a training method more akin to the Chinese Room than the fluid instructions used to train humans. In any case, the model, like the worker in Searle\u2019s experiment, receives instructions linking \u201cinput\u201d to expected \u201coutput.\u201d By changing the type of information in the two categories, we can train models to perform all the tasks listed above. No understanding is required.<\/span><\/p>\n

The Imitation Game<\/b><\/h3>\n

All the enthusiastic or alarmed reports about the latest generation of language models\u2014like the famous ChatGPT\u2014imply that we have reached the point where AI models can use language exactly like humans, or at least are very close to that point. However, language models still have significant gaps. These gaps reveal that using language exactly like humans remains a challenge for AI. But they also show that in the linguistic field \u201cartificial intelligence\u201d is a concept still seeking a definition almost 70 years after Searle\u2019s Chinese Room experiment. Recent improvements in language model performance have been driven by a single factor: size. If the latest generation of models can sometimes appear capable of doing the same things with language as humans, it is mainly because they are extremely large models. The size of an AI model is related to the amount of data it is trained on, but model and size are not the same thing. The number of parameters determines how many things the model adds, subtracts, divides, and multiplies to produce a result. All these numbers must be set to appropriate values to achieve the desired results. In the case of language models, we must determine how likely it is that in a given language each word follows the previous ones. If we take 100,000 words in English, that equals 10 billion (100,000 times 100,000) combinations for two words. In reality, we need an even larger number of parameters to do this, but it gives an idea of the scale. The more parameters a model has, the more its output resembles (and sounds like) the language it was trained to imitate. However, just like a chain can never become a complete circle, the AI model cannot become a language. The person in the Chinese Room can become indistinguishable from a native speaker but will never be one. Imagine now that the worker in the Chinese Room is handed a note that says, \u201cIf you read this text, call the police immediately; I\u2019m being threatened with a gun.\u201d The system they have learned allows the worker to generate a response. But if they don\u2019t understand the text they see, they cannot solve the new problem no matter how serious the threat. The issue of \u201ctrue understanding\u201d is certainly crucial for philosophers and, in our example, for someone threatened by a gun, but in many situations and for most people simply imitating language is probably sufficient. And it is precisely on imitation that one of the most well-known tests for language-based AI systems is founded: the Turing Test. Named after the pioneer of computer science Alan Turing, it is based on a simple question: \u201cCan a computer converse with a human in a way convincing enough to trick them into believing it is another human?\u201d Turing predicted that by 2000, a computer would be able to deceive 30% of \u201chuman judges.\u201d He was not far off: in 2024, we have several contenders aiming to pass the test. In the Turing Test, a human judge converses with two entities, one human and one computer. If the judge cannot distinguish who is who, the computer passes the test. There is even a competition\u2014the Loebner Prize\u2014that offers $100,000 to the first AI model to pass the Turing Test. A fluent conversation is an essential skill for AI, but it is unclear what lessons might be derived about language, computers, and ourselves from the fact that an AI model passes the Turing Test. According to some critics, it is too easy to trick the judge, and the test measures more a computer\u2019s ability to imitate human conversation than intelligence. It is also unclear whether the test\u2019s implicit assumption\u2014that communication is sufficient proof of human-like intelligence\u2014is justified.<\/span><\/p>\n

AI and Wittgenstein\u2019s Lion<\/b><\/h3>\n

We have evidence that having language skills does not mean having intelligence, emotions, or any other human capacity. These come from an unexpected source: brain injuries. When they affect specific areas, they can show us what connections exist between different human faculties. Over the years, neuroscientists have documented thousands of patients who, due to accidents or illnesses, have suffered brain injuries and lost the ability to produce or understand language, known as aphasia. However, as limiting as aphasia may be for a patient\u2019s communication ability, it does not affect their other abilities: aphasics do not lose empathy, control of their emotions, or walk into walls. They can still move, remember things (although working memory is sometimes impaired), smile, and lead a relatively normal life. Linguistics tends to see language as separate from our other cognitive (and physical) abilities. We now know this is not the case. Learning a language requires more than exposure to a large amount of text: we also need social cues, demonstrations, tone of voice, and other factors to understand and communicate. We are not human because we speak; we speak because we are human. Language models, on the other hand, are currently only language factories. If ChatGPT suffered the equivalent of a brain injury that caused the loss of language, it would become useless. The Austrian philosopher Ludwig Wittgenstein left us a poignant image of the problem we face today with AI models. Even if a lion could speak, he wrote, we would not be able to understand it. The lived experience of a lion is so fundamentally different from ours that the barrier to understanding is not language; it is everything else. If the inputs in the Chinese Room came from Wittgenstein\u2019s lion, simply translating them would not be enough to understand \u201clion-ness.\u201d Our models, on the other hand, have begun to sound so human that it is increasingly difficult to remember how significant the differences between them and us are. Wittgenstein\u2019s lion and Searle\u2019s room remind us that philosophical implications are still important in AI. After years of focusing on improving technical factors, philosophical ideas about language are now central again in AI research. However, our desire to attribute mental states to an AI model often overshadows these considerations. Language models will continue to amaze us, but we must remember that \u201cunderstanding\u201d means more than just producing words. Humans use language as an expressive tool to share new thoughts and experiences; computers play a probability game constrained by the limits of the training data they have received. Without language, humans still think, but for computers, language is what gives them life.<\/span><\/p>\n

Bio<\/b><\/em><\/p>\n

Dirk Hovy is a professor in the Department of Computer Science at Bocconi University, specialising in natural language processing. He is currently working on a book about the risks of anthropomorphising artificial intelligence.<\/span><\/em><\/p>\n","protected":false},"excerpt":{"rendered":"

AI language models continue to astonish us with their ability to mimic human language. But comprehending a language is an entirely different matter. What can […]<\/p>\n","protected":false},"author":7509,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[1],"tags":[],"coauthors":[136],"class_list":["post-4013","post","type-post","status-publish","format-standard","hentry","category-non-categorizzato"],"acf":[],"yoast_head":"\nAI Imitates Us But Does It Truly Understand?\u00a0 - Rivista Eco<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.rivistaeco.com\/en\/2024\/09\/02\/ai-imitates-us-but-does-it-truly-understand\/\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"AI Imitates Us But Does It Truly Understand?\u00a0 - Rivista Eco\" \/>\n<meta property=\"og:description\" content=\"AI language models continue to astonish us with their ability to mimic human language. But comprehending a language is an entirely different matter. What can […]\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.rivistaeco.com\/en\/2024\/09\/02\/ai-imitates-us-but-does-it-truly-understand\/\" \/>\n<meta property=\"og:site_name\" content=\"Rivista Eco\" \/>\n<meta property=\"article:published_time\" content=\"2024-09-02T09:27:50+00:00\" \/>\n<meta name=\"author\" content=\"Dirk Hovy\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Dirk Hovy\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"11 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/www.rivistaeco.com\\\/en\\\/2024\\\/09\\\/02\\\/ai-imitates-us-but-does-it-truly-understand\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.rivistaeco.com\\\/en\\\/2024\\\/09\\\/02\\\/ai-imitates-us-but-does-it-truly-understand\\\/\"},\"author\":{\"name\":\"Dirk Hovy\",\"@id\":\"https:\\\/\\\/www.rivistaeco.com\\\/#\\\/schema\\\/person\\\/a76eadacccc5ab00eb79f72727dabf1a\"},\"headline\":\"AI Imitates Us But Does It Truly Understand?\u00a0\",\"datePublished\":\"2024-09-02T09:27:50+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/www.rivistaeco.com\\\/en\\\/2024\\\/09\\\/02\\\/ai-imitates-us-but-does-it-truly-understand\\\/\"},\"wordCount\":1940,\"articleSection\":[\"Non categorizzato\"],\"inLanguage\":\"en-US\"},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/www.rivistaeco.com\\\/en\\\/2024\\\/09\\\/02\\\/ai-imitates-us-but-does-it-truly-understand\\\/\",\"url\":\"https:\\\/\\\/www.rivistaeco.com\\\/en\\\/2024\\\/09\\\/02\\\/ai-imitates-us-but-does-it-truly-understand\\\/\",\"name\":\"AI Imitates Us But Does It Truly Understand?\u00a0 - Rivista Eco\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.rivistaeco.com\\\/#website\"},\"datePublished\":\"2024-09-02T09:27:50+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/www.rivistaeco.com\\\/#\\\/schema\\\/person\\\/a76eadacccc5ab00eb79f72727dabf1a\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/www.rivistaeco.com\\\/en\\\/2024\\\/09\\\/02\\\/ai-imitates-us-but-does-it-truly-understand\\\/#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/www.rivistaeco.com\\\/en\\\/2024\\\/09\\\/02\\\/ai-imitates-us-but-does-it-truly-understand\\\/\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/www.rivistaeco.com\\\/en\\\/2024\\\/09\\\/02\\\/ai-imitates-us-but-does-it-truly-understand\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"http:\\\/\\\/www.rivistaeco.com\\\/en\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"AI Imitates Us But Does It Truly Understand?\u00a0\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/www.rivistaeco.com\\\/#website\",\"url\":\"https:\\\/\\\/www.rivistaeco.com\\\/\",\"name\":\"Rivista Eco\",\"description\":\"\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/www.rivistaeco.com\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/www.rivistaeco.com\\\/#\\\/schema\\\/person\\\/a76eadacccc5ab00eb79f72727dabf1a\",\"name\":\"Dirk Hovy\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/7e205e16605b5c0b465f20bd0e1a127b4be252beff384e5a4e07115518f510e8?s=96&d=mm&r=g962b76fabd8be51c97041bd2924e2be3\",\"url\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/7e205e16605b5c0b465f20bd0e1a127b4be252beff384e5a4e07115518f510e8?s=96&d=mm&r=g\",\"contentUrl\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/7e205e16605b5c0b465f20bd0e1a127b4be252beff384e5a4e07115518f510e8?s=96&d=mm&r=g\",\"caption\":\"Dirk Hovy\"},\"url\":\"https:\\\/\\\/www.rivistaeco.com\\\/en\\\/author\\\/dhovy\\\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"AI Imitates Us But Does It Truly Understand?\u00a0 - Rivista Eco","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.rivistaeco.com\/en\/2024\/09\/02\/ai-imitates-us-but-does-it-truly-understand\/","og_locale":"en_US","og_type":"article","og_title":"AI Imitates Us But Does It Truly Understand?\u00a0 - Rivista Eco","og_description":"AI language models continue to astonish us with their ability to mimic human language. But comprehending a language is an entirely different matter. What can […]","og_url":"https:\/\/www.rivistaeco.com\/en\/2024\/09\/02\/ai-imitates-us-but-does-it-truly-understand\/","og_site_name":"Rivista Eco","article_published_time":"2024-09-02T09:27:50+00:00","author":"Dirk Hovy","twitter_card":"summary_large_image","twitter_misc":{"Written by":"Dirk Hovy","Est. reading time":"11 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/www.rivistaeco.com\/en\/2024\/09\/02\/ai-imitates-us-but-does-it-truly-understand\/#article","isPartOf":{"@id":"https:\/\/www.rivistaeco.com\/en\/2024\/09\/02\/ai-imitates-us-but-does-it-truly-understand\/"},"author":{"name":"Dirk Hovy","@id":"https:\/\/www.rivistaeco.com\/#\/schema\/person\/a76eadacccc5ab00eb79f72727dabf1a"},"headline":"AI Imitates Us But Does It Truly Understand?\u00a0","datePublished":"2024-09-02T09:27:50+00:00","mainEntityOfPage":{"@id":"https:\/\/www.rivistaeco.com\/en\/2024\/09\/02\/ai-imitates-us-but-does-it-truly-understand\/"},"wordCount":1940,"articleSection":["Non categorizzato"],"inLanguage":"en-US"},{"@type":"WebPage","@id":"https:\/\/www.rivistaeco.com\/en\/2024\/09\/02\/ai-imitates-us-but-does-it-truly-understand\/","url":"https:\/\/www.rivistaeco.com\/en\/2024\/09\/02\/ai-imitates-us-but-does-it-truly-understand\/","name":"AI Imitates Us But Does It Truly Understand?\u00a0 - Rivista Eco","isPartOf":{"@id":"https:\/\/www.rivistaeco.com\/#website"},"datePublished":"2024-09-02T09:27:50+00:00","author":{"@id":"https:\/\/www.rivistaeco.com\/#\/schema\/person\/a76eadacccc5ab00eb79f72727dabf1a"},"breadcrumb":{"@id":"https:\/\/www.rivistaeco.com\/en\/2024\/09\/02\/ai-imitates-us-but-does-it-truly-understand\/#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.rivistaeco.com\/en\/2024\/09\/02\/ai-imitates-us-but-does-it-truly-understand\/"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.rivistaeco.com\/en\/2024\/09\/02\/ai-imitates-us-but-does-it-truly-understand\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"http:\/\/www.rivistaeco.com\/en\/"},{"@type":"ListItem","position":2,"name":"AI Imitates Us But Does It Truly Understand?\u00a0"}]},{"@type":"WebSite","@id":"https:\/\/www.rivistaeco.com\/#website","url":"https:\/\/www.rivistaeco.com\/","name":"Rivista Eco","description":"","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.rivistaeco.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Person","@id":"https:\/\/www.rivistaeco.com\/#\/schema\/person\/a76eadacccc5ab00eb79f72727dabf1a","name":"Dirk Hovy","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/secure.gravatar.com\/avatar\/7e205e16605b5c0b465f20bd0e1a127b4be252beff384e5a4e07115518f510e8?s=96&d=mm&r=g962b76fabd8be51c97041bd2924e2be3","url":"https:\/\/secure.gravatar.com\/avatar\/7e205e16605b5c0b465f20bd0e1a127b4be252beff384e5a4e07115518f510e8?s=96&d=mm&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/7e205e16605b5c0b465f20bd0e1a127b4be252beff384e5a4e07115518f510e8?s=96&d=mm&r=g","caption":"Dirk Hovy"},"url":"https:\/\/www.rivistaeco.com\/en\/author\/dhovy\/"}]}},"jetpack_featured_media_url":"","jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/posts\/4013","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/users\/7509"}],"replies":[{"embeddable":true,"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/comments?post=4013"}],"version-history":[{"count":1,"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/posts\/4013\/revisions"}],"predecessor-version":[{"id":4014,"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/posts\/4013\/revisions\/4014"}],"wp:attachment":[{"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/media?parent=4013"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/categories?post=4013"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/tags?post=4013"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/coauthors?post=4013"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}