THE PARROT AND THE PROPHET

For a moment, consider this idea. What if you could bring together every piece of text humanity has ever created and pile it up? This pile would be unbelievably tall, stretching past our atmosphere, beyond the moon, and further than any spacecraft. It would hold both wise teachings and foolish talk, romantic poems and hateful writings, scientific facts and creative dreams. In short, every word we have ever written down or saved would be there.

After that, imagine giving a machine the ability to read all the material in the pile. The machine would not only read, but it would also find common patterns. It would learn word sequences, and thus become a sort of statistical ghost of human language. Can we define this machine? What knowledge would it possess? What would it think is true? Most importantly, what happens when it speaks back to us?

This is not a thought experiment. This describes our actual work over recent years. We’ve been building this quickly, without enough careful thought. These so-called large language models, which run many tools today, are precisely that: a statistical copy of human talk. A big part of their knowledge comes from being trained on that immense pile of information. Because of this, they understand the outlines of our common way of expressing things. And now, they are talking to us. Their voices are starting to feel eerily like a real person’s. But there’s a bigger issue here. What is the machine’s message? And who is the real source of that voice? The phrase some experts use is “stochastic parrots.” It’s a catchy name that really explains what’s going on. It’s not exactly a compliment, but it’s an accurate description. After all, a parrot can be trained to talk in a way that sounds quite real. It can put words in an order that seems to make sense. They associate words with contexts, so a trained parrot might greet you at the door and bid you farewell as you leave. However, the parrot has no idea what its words mean. The bird has no private thoughts, no aims, and no real sense of what its words are supposed to mean. The process is simple repetition and recombination, following the frequency of sounds it memorized.

The term “stochastic” introduces an important extra detail. In simple terms, stochastic means random or unpredictable, like a roll of the dice. The stochastic parrot isn’t a simple echo machine; it doesn’t just say what it heard, exactly as it heard it. Instead, it makes new sentences. It creates fresh word combinations based on the patterns it learned. For any prompt, it figures out the most probable next word. Then it does the same for the following word, and the one after that. This method can produce writing that reads well, makes sense, and seems inventive.  However, the basic process is about finding patterns, not about gaining true comprehension.

This difference is important. The thing is, our brains aren’t designed to notice it. Throughout human history, language has always come from another person’s thoughts. We naturally assume that when someone speaks, they are expressing their private thoughts. Our automatic reaction is to think there’s a conscious individual behind the words. This idea is so deeply built into our thinking that we can’t just switch it off. We do it all the time. We talk to our pets like they understand us. We get mad at cars that won’t start. We say the weather is being mean. Naturally, we extend this assumption to any machine that can talk to us clearly.

The experts who study these machines also have a term for this habit. They call it the tendency to mistake fluent output for meaningful communication. And they warn that this mistake is not harmless. If we believe we’re talking to a real intelligence, we become vulnerable. The threat isn’t from a bad actor, but from the flawed information the parrot has memorized. So what did it learn? It learned everything in that huge pile of data. And this is where things get more complex and worrying than we first thought.

Let us consider what actually goes into training one of these models. The process begins by scraping the web for text, pulling in billions of pages from every corner of the public internet. This raw material is then filtered, cleaned, and organized into a dataset that can be fed into the learning algorithm. They are unimaginably large, containing hundreds of billions of words. Initially, it feels like a very democratic way of building knowledge. The idea is that the AI absorbs the full range of human expression, from every corner of the globe. How could we possibly be more inclusive than that?

But the appearance of inclusivity is deceptive. The Internet content is just a slice of global life, not the whole picture. The online world is defined by factors like affordability, available technology, and dominant power. Think about the people who actually create the online content that ends up in these training sets. Access to the web is unbalanced. For example, there are far more young users from wealthy nations than elderly users from developing areas. Focusing on the English-language content, which is the largest portion, the demographic bias is stronger. Data from platforms such as Reddit, often used for training, is dominated by a young, male, North American crowd. On Wikipedia, another key training source, the people writing the content are mostly men.

This doesn’t mean that everyone else is totally silent. They are included, but they are a minority in a dataset where a few perspectives are the loudest. And statistical dominance matters because these models learn probabilities. The models learn to recognize frequent patterns and ignore rare ones. They learn that when the training data talks about scientists, it most often pairs that word with male pronouns. They learn that when it discusses certain countries, it most often does so in the context of crisis and conflict. So the model learns a data-based picture of the world. This image doesn’t show the real variety of human life. Instead, it shows the shape of who is loudest online.

Trying to filter the data often backfires, making the bias even stronger. Developers filter the data by removing specific offensive terms from a list. That approach appears logical. This cleaning process harms groups who have turned abusive terms into symbols of pride, because their speech gets caught in the filter. A tool designed to remove bad language winds up censoring the voices of marginalized groups.

And then there is the question of what the data contains beyond simple demographics.  What’s online is not an objective history of human thinking. The internet reflects human society, so it carries all our prejudices and power differences.  It contains all the dark corners of the web, from racist attacks to wild conspiracy theories. It includes the language of hate alongside the language of love, the language of ignorance alongside the language of wisdom. The algorithms of social media favor engaging content, which is often negative. This means hate and ignorance are overrepresented. The machine learns all of this. For the machine, learning stereotypes is no different from learning how to form a proper sentence. It learns that certain groups are associated with certain stereotypes because those associations appear frequently in its training text. It learns that women are more often described in terms of appearance than achievement, that Black people are more often associated with crime than with creativity, that disabled people are more often discussed as problems than as people. These patterns are not added deliberately; they emerge naturally from the statistical structure of the data. And then the machine begins to speak.

When researchers test these models for bias, the results are consistently troubling. When the models get a prompt mentioning, say, race or gender, their output is full of the same old prejudices, sometimes worse. Given a sentence about a doctor, the AI’s default is to use male pronouns; for nurses, it’s female pronouns. Asked to generate stories about different groups, they produce narratives that vary systematically in tone and content depending on the group named. Given a prompt about disability, the AI’s response is likely to be about tragedy and burden.

These are not bugs that can be easily fixed. They are a direct result of how the models are designed to learn. The models have learned that in their training data, certain words tend to appear together. Their knowledge reflects real-world language use, including its flaws and unfairness. To get rid of these biases, we’d have to change the basic math inside the model. We would need to retrain it on a perfect, made-up dataset. In theory, this could work. But in reality, it’s incredibly hard to do.  The size and complexity of the data make it nearly impossible to untangle. Let’s say it is possible. Then we have to ask: who decides what the “correct” patterns are? Where do we draw the line between harmful bias and acceptable viewpoint? which set of human values will be the new standard?

These aren’t problems for engineers to solve. They are political, moral, and philosophical issues that just look like tech problems. While we debate this, the AIs are still being released and used. These models are used everywhere: in online search, writing aids, and teaching apps. Their output is everywhere, from blog posts to news stories. The internet, now filled with AI text, becomes the training ground for the next AI. This is where the loop closes, and where the stakes become truly consequential. Imagine a future version of these models, trained not just on human-generated text but on text that earlier models have produced.

This isn’t a wild guess. It’s happening right now. AI-written text is all over the internet. You can find it in blog posts, product ratings, and social media. This AI content is now part of the web that is scraped for future training sets. The stochastic parrots are beginning to learn from each other. This creates a feedback loop that researchers are only beginning to understand. Each new model not only keeps the old biases but also adds its own spin, amplifying them. Associations that were weak become strong.  Each generation of AI moves further from real human expression, making it more distorted. And because the models produce text that is fluent and coherent, it is difficult to detect this degradation.

The output still looks like language. It still makes sense on a basic level. But this language is very different now. It’s been through many rounds of AI processing. Its shape comes from AI patterns, not from real human life. So, what will happen if we start using these AIs as our main source of knowledge? What happens when the text they generate shapes how we understand history, culture, and each other? How will it affect us if the biases are in the systems that manage our work and play?.

These questions are not hypothetical. They are the questions we should be asking now, while there is still time to shape the trajectory of this technology. The researchers who first raised these concerns in a systematic way did so in a paper that became famous for reasons they did not anticipate. The paper, presented at a major academic conference in 2021, laid out the risks associated with large language models with unusual clarity and force. It discussed how training these AIs hurts the environment through massive energy use and pollution, and how this unfairly affects poor communities. It pointed out that the data over-represents certain groups, contains hidden biases, and is too big for any person to fully understand. It warned about the risk of treating a clever mimic as a real intelligence.

The paper made recommendations. It urged researchers to consider environmental impacts before embarking on large-scale projects. It recommended we spend resources on curating fair datasets, not hoovering up the whole internet. It encouraged developers to step back and consider if their methods serve their purpose and respect community values. Finally, it urged researchers to explore other paths, not just building bigger and bigger AIs.

People remember the controversy around the authors more than the paper’s content. Some of them were employed by major technology companies at the time. After the paper was submitted for publication, those authors were asked to remove their names or face consequences. Some chose removal. Others left their positions. A note in the final paper revealed that some authors were ordered by their employers to remove their names. This story shows us a key part of the bigger picture we’re talking about. The technology is being developed primarily by large corporations with enormous resources and competitive pressures.

These corporations have incentives to move fast, to deploy widely, to capture markets. They have much weaker incentives to pause, to reflect, to consider long-term consequences. Researchers who raise concerns find themselves at odds with the institutions that fund their work. We face a massive problem that demands a unified response, but that’s very difficult to achieve. We are building a technology that will shape how future generations understand the world, and we are doing so with remarkably little public conversation about what we are building and why.

Let us return to the image of the enormous pile. It claims to contain all, but that’s far from the truth. Whole parts of what it means to be human are absent. It contains almost nothing from the billions of people who do not use the internet in languages that dominate online spaces. It contains almost nothing from the oral traditions that have preserved knowledge for millennia without writing. It contains almost nothing from the private conversations, the unrecorded thoughts, the unwritten stories that make up the bulk of human life. What it does contain is overstuffed with certain kinds of content: the output of media industries, the debates of online forums, the arguments of social media, the published records of institutions that have historically had the power to document and preserve. It contains the voices of the powerful far more clearly than the voices of the powerless. It contains the patterns of bias and prejudice that have shaped human society for centuries. The result is a technology that can produce fluent, convincing text about almost any topic. Ask it about history, and it will summarize what the history books say—the books that made it into the pile. Ask it about science, and it will explain what the scientific literature documents—the literature that was published in dominant languages and journals. Ask it about culture, and it will describe what the cultural record contains—the record that was preserved by those with the resources to preserve.

Ask it about the botanist in Bangladesh whose work was never internationally recognized. Ask it about the botanist in Bangladesh whose work was never internationally recognized. Ask it about the women who wrote under male pseudonyms because that was the only way to be published. Ask it about the indigenous communities whose knowledge was never written down because their traditions were oral. Ask it about the activists whose newsletters were printed on cheap paper and never archived. And the machine will have nothing to say. Not because it is hiding anything, but because those voices were never in its training data. Their silence is built into the system. It’s a part of how the technology is made.

This is the danger that the concept of the stochastic parrot illuminates. The parrot speaks, but it only speaks what it has heard. If what it has heard is a partial, biased, distorted sample of human expression, then that is what it will reproduce. And because its speech is fluent and confident, we will be tempted to take it as authoritative. We will forget that behind the fluency lies an absence—the absence of all the voices that never made it into the pile.

Philosophers studying this sometimes use the term “value lock.” The idea is simple: when we embed certain values and perspectives into technology, we make them harder to change. They become locked in, built into the systems that shape our daily lives. And because those systems are difficult to modify once deployed, the locked-in values persist even as society’s values evolve.

Language models are a powerful mechanism for value lock. They are trained on data from the past—often the recent past, but sometimes data going back centuries. This data contains the values of those past eras, including values we have since recognized as harmful. When the models are deployed, they reproduce those values in the present. A model trained on twentieth-century texts will generate text that reflects twentieth-century assumptions about gender, race, and power. Those assumptions become locked into the technology, even as the culture moves on.

Consider how social movements change language. The Black Lives Matter movement transformed how many people talk about race and policing. The #MeToo movement shifted conversations about sexual harassment and consent. Marriage equality changed how same-sex relationships are discussed in public discourse.

These changes happened through collective effort, through organizing and advocacy, through people pushing back against dominant narratives. But a language model trained on data from before these movements will not reflect these changes. It will reproduce the older language, the older assumptions, the older biases. And if that model is widely used—in search engines, in writing assistants, in educational tools—it will have the effect of slowing down linguistic and cultural change.

It will act as a kind of cultural brake, preserving older ways of thinking in the very systems that are supposed to help us think anew.

This is not inevitable. Models can be fine-tuned on newer data. They can be adjusted to reflect evolving norms. But these adjustments require effort, resources, and intention. They require recognizing that the default state of the technology is to reproduce the past. And they require making choices about what values to encode—choices that are fundamentally political. The technology companies building these models are not eager to advertise this fact.

They present their systems as neutral tools, as mirrors of reality, as objective sources of information. They do not highlight the ways in which their models are locked into particular value systems, particular historical moments, particular perspectives on the world. To do so would be to acknowledge that they are making choices, exercising power, shaping the future.

What would it mean to build these systems differently? The researchers who have thought most deeply about these questions offer several suggestions, none of them easy. First, they argue, we should slow down.

The race to build ever larger models, ever more powerful systems, is driven by competition and hype rather than by careful consideration of what is actually needed. Before embarking on a new project, we should ask: What problem are we trying to solve? Who benefits from the solution? What are the risks? Are there less resource-intensive approaches that could achieve similar goals?

These questions should be asked at the beginning, not after the model is built and deployed. Second, we should invest in curation, not just collection. The current approach of scraping everything and then trying to filter out the worst is backwards. It assumes that bigger is better, that more data automatically means better models. But as we have seen, more data also means more bias, more noise, more harmful content.

A better approach would be to carefully construct datasets that represent the diversity of human experience, that include voices that have been historically marginalized, that document their contents and limitations transparently. This is slower and more expensive than web scraping. It is also more responsible. Third, we should develop better documentation practices. Every dataset should come with a “datasheet” that describes its contents, its sources, its limitations, and its intended uses. Every model should come with a “model card” that explains its capabilities, its biases, its performance characteristics, and its appropriate applications.

These documentation practices make it possible for users to understand what they are working with, to anticipate problems, to make informed decisions about deployment. Fourth, we should broaden who participates in building these systems. The technology industry is famously homogeneous—overwhelmingly male, overwhelmingly white, overwhelmingly educated in a handful of elite institutions. This homogeneity shows up in the models. When the people building the technology come from similar backgrounds, they share similar blind spots. They do not notice what is missing. They do not anticipate how the technology will affect communities they have never thought about. Bringing more voices into the development process is not just a matter of fairness; it is a matter of building better, more robust technology.

Fifth, we should consider whether some applications are simply too risky to pursue. The ability to generate convincing human-like text has obvious benefits. It also has obvious dangers. Bad actors can use it to spread misinformation, to manipulate public opinion, to harass and intimidate. The technology itself does not distinguish between beneficial and harmful uses. It is a tool, and like any powerful tool, it can be used for good or for ill.

Deciding which uses to enable and which to prevent is a social and political question, not a technical one. The challenge we face is not primarily technological. It is ethical, political, and philosophical. We are building machines that will shape how future generations understand the world, and we are doing so with remarkably little reflection on what we are doing and why.

The concept of the stochastic parrot gives us a way to think about this challenge. It reminds us that these systems do not understand what they are saying. They are not minds with beliefs and intentions. They are patterns of language, nothing more. And the patterns they have learned come from a particular slice of human experience—a slice that is broader than any individual’s experience but narrower than the full diversity of human life.

When we interact with these systems, we are not conversing with an intelligence. We are looking into a funhouse mirror that reflects back a distorted image of our own collective expression. The distortion comes from the biases in the training data, from the limitations of the learning algorithms, from the choices made by the developers, from the pressures of the market. It comes from all the ways in which the technology is shaped by the society that produces it.

The mirror shows us ourselves, but not as we are. It shows us ourselves as we appear in the records we have left behind—records that are incomplete, biased, and shaped by power. It shows us the past, not the future. It shows us what was, not what could be. Yet, we are using this mirror to guide our decisions, to educate our children, to shape our culture. We are treating the stochastic parrot as if it were a prophet, as if its fluent speech were wisdom, as if its confident assertions were truth.

The prophets of old were believed to speak for the divine. Their words carried authority because they came from beyond the human. Our new prophets speak for the human, all too human. Their words carry authority because we have forgotten that they are just parrots, repeating what they have heard. The question before us is whether we can remember. Whether we can look at these fluent, confident machines and see them for what they are: mirrors of our own making, reflecting back a partial and distorted image of who we have been. Whether we can then use that image not as a destination but as a starting point—a way of seeing what is missing, what is silent, what has been left out. And whether we can build, alongside these machines, the conditions for all voices to be heard, not just the loudest and most powerful. This is not a technical problem. It is a human one. And like all human problems, it requires not better machines but better judgment, better attention, better care.

It requires us to become conscious architects of our collective memory, rather than passive consumers of whatever the stochastic parrots produce. The parrots will keep talking. That is what they do. The question is whether we will keep listening—and whether, when we listen, we will remember that there is more to speech than pattern, more to meaning than probability, more to understanding than fluent repetition. The question is whether we will remain human enough to hear the silence behind the sound.

This article was previously published on qatarmoments. To see the original article, click here