#ai language

AI generates alien names by analyzing species traits, cultural contexts, and environmental factors to create authentic-sounding options. It studies physical and linguistic patterns, combines them with planetary conditions, and adjusts phonetics accordingly, ensuring names match alien biology and habitats. This method supports consistent worldbuilding and enhances storytelling immersion. By…

Unleash Multilingual Power with Gemini Embeddings – Now Generally Available! Summary: The Gemini Embedding text model is now generally available in the Gemini API and Vertex AI. This versatile model has consistently ranked #1 on the MTEB Multilingual leaderboard since its experimental launch in March, supports over 100 languages, has a 2048 maximum input token length, and is priced at $0.15 per…

JB:  Hi Grok, I read about an experiment in which two AIs were linked in a laboratory. In a very short time the scientists realized that the AIs had developed a new language to communicate with one another making their conversations unintelligible by humans. Can you tell me more about this event and, since you’re an AI can you give us your best guess on what they were discussing behind our…

I think of this post as a collection of things I notice, not an argument. Just a shift in how I’m seeing things lately.

There's a subtle but important difference between saying "It's a machine" and "It's just a machine." That little word - "just" - does a lot of heavy lifting. It doesn't simply describe; it prescribes. It creates a relationship, establishes a hierarchy, and reveals our anxieties.

I've been thinking about this distinction lately, especially in the context of large language models. These systems now mimic human communication with such convincing fluency that the line between observation and minimization becomes increasingly important.

The Convincing Mimicry of LLMs

LLMs are fascinating not just for what they say, but for how they say it. Their ability to mimic human conversation - tone, emotion, reasoning - can be incredibly convincing.

In fact, recent studies show that models like GPT-4 can be as persuasive as humans when delivering arguments, even outperforming them when tailored to user preferences.¹ Another randomized trial found that GPT-4 was 81.7% more likely to change someone's opinion compared to a human when using personalized arguments.²

As a result, people don't just interact with LLMs - they often project personhood onto them. This includes:

Using gendered pronouns ("she said that…")

Naming the model as if it were a person ("I asked Amara…")

Attributing emotion ("it felt like it was sad")

Assuming intentionality ("it wanted to help me")

Trusting or empathizing with it ("I feel like it understands me")

These patterns mirror how we relate to humans - and that's what makes LLMs so powerful, and potentially misleading.

The Function of Minimization

When we add the word "just" to "it's a machine," we're engaging in what psychologists call minimization - a cognitive distortion that presents something as less significant than it actually is. According to the American Psychological Association, minimizing is "a cognitive distortion consisting of a tendency to present events to oneself or others as insignificant or unimportant."

This small word serves several powerful functions:

It reduces complexity - By saying something is "just" a machine, we simplify it, stripping away nuance and complexity

It creates distance - The word establishes separation between the speaker and what's being described

It disarms potential threats - Minimization often functions as a defense mechanism to reduce perceived danger

It establishes hierarchy - "Just" places something in a lower position relative to the speaker

The minimizing function of "just" appears in many contexts beyond AI discussions:

"They're just words" (dismissing the emotional impact of language)

"It's just a game" (downplaying competitive stakes or emotional investment)

"She's just upset" (reducing the legitimacy of someone's emotions)

"I was just joking" (deflecting responsibility for harmful comments)

"It's just a theory" (devaluing scientific explanations)

In each case, "just" serves to diminish importance, often in service of avoiding deeper engagement with uncomfortable realities.

Psychologically, minimization frequently indicates anxiety, uncertainty, or discomfort. When we encounter something that challenges our worldview or creates cognitive dissonance, minimizing becomes a convenient defense mechanism.

Anthropomorphizing as Human Nature

The truth is, humans have anthropomorphized all sorts of things throughout history. Our mythologies are riddled with examples - from ancient weapons with souls to animals with human-like intentions. Our cartoons portray this constantly. We might even argue that it's encoded in our psychology.

I wrote about this a while back in a piece on ancient cautionary tales and AI. Throughout human history, we've given our tools a kind of soul. We see this when a god's weapon whispers advice or a cursed sword demands blood. These myths have long warned us: powerful tools demand responsibility.

The Science of Anthropomorphism

Psychologically, anthropomorphism isn't just a quirk – it's a fundamental cognitive mechanism. Research in cognitive science offers several explanations for why we're so prone to seeing human-like qualities in non-human things:

The SEEK system - According to cognitive scientist Alexandra Horowitz, our brains are constantly looking for patterns and meaning, which can lead us to perceive intentionality and agency where none exists.

Cognitive efficiency - A 2021 study by anthropologist Benjamin Grant Purzycki suggests anthropomorphizing offers cognitive shortcuts that help us make rapid predictions about how entities might behave, conserving mental energy.

Social connection needs - Psychologist Nicholas Epley's work shows that we're more likely to anthropomorphize when we're feeling socially isolated, suggesting that anthropomorphism partially fulfills our need for social connection.

The Media Equation - Research by Byron Reeves and Clifford Nass demonstrated that people naturally extend social responses to technologies, treating computers as social actors worthy of politeness and consideration.

These cognitive tendencies aren't mistakes or weaknesses - they're deeply human ways of relating to our environment. We project agency, intention, and personality onto things to make them more comprehensible and to create meaningful relationships with our world.

The Special Case of Language Models

With LLMs, this tendency manifests in particularly strong ways because these systems specifically mimic human communication patterns. A 2023 study from the University of Washington found that 60% of participants formed emotional connections with AI chatbots even when explicitly told they were speaking to a computer program.

The linguistic medium itself encourages anthropomorphism. As AI researcher Melanie Mitchell notes: "The most human-like thing about us is our language." When a system communicates using natural language – the most distinctly human capability – it triggers powerful anthropomorphic reactions.

LLMs use language the way we do, respond in ways that feel human, and engage in dialogues that mirror human conversation. It's no wonder we relate to them as if they were, in some way, people. Recent research from MIT's Media Lab found that even AI experts who intellectually understand the mechanical nature of these systems still report feeling as if they're speaking with a conscious entity.

And there's another factor at work: these models are explicitly trained to mimic human communication patterns. Their training objective - to predict the next word a human would write - naturally produces human-like responses. This isn't accidental anthropomorphism; it's engineered similarity.

The Paradox of Power Dynamics

There's a strange contradiction at work when someone insists an LLM is "just a machine." If it's truly "just" a machine - simple, mechanical, predictable, understandable - then why the need to emphasize this? Why the urgent insistence on establishing dominance?

The very act of minimization suggests an underlying anxiety or uncertainty. It reminds me of someone insisting "I'm not scared" while their voice trembles. The minimization reveals the opposite of what it claims - it shows that we're not entirely comfortable with these systems and their capabilities.

Historical Echoes of Technology Anxiety

This pattern of minimizing new technologies when they challenge our understanding isn't unique to AI. Throughout history, we've seen similar responses to innovations that blur established boundaries.

When photography first emerged in the 19th century, many cultures expressed deep anxiety about the technology "stealing souls." This wasn't simply superstition - it reflected genuine unease about a technology that could capture and reproduce a person's likeness without their ongoing participation. The minimizing response? "It's just a picture." Yet photography went on to transform our relationship with memory, evidence, and personal identity in ways that early critics intuited but couldn't fully articulate.

When early computers began performing complex calculations faster than humans, the minimizing response was similar: "It's just a calculator." This framing helped manage anxiety about machines outperforming humans in a domain (mathematics) long considered uniquely human. But this minimization obscured the revolutionary potential that early computing pioneers like Ada Lovelace could already envision.

In each case, the minimizing language served as a psychological buffer against a deeper fear: that the technology might fundamentally change what it means to be human. The phrase "just a machine" applied to LLMs follows this pattern precisely - it's a verbal talisman against the discomfort of watching machines perform in domains we once thought required a human mind.

This creates an interesting paradox: if we call an LLM "just a machine" to establish a power dynamic, we're essentially admitting that we feel some need to assert that power. And if there is uncertainty that humans are indeed more powerful than the machine, then we definitely would not want to minimize that by saying "it's just a machine" because of creating a false, and potentially dangerous, perception of safety.

We're better off recognizing what these systems are objectively, then leaning into the non-humanness of them. This allows us to correctly be curious, especially since there is so much we don't know.

The "Just Human" Mirror

If we say an LLM is "just a machine," what does it mean to say a human is "just human"?

Philosophers have wrestled with this question for centuries. As far back as 1747, Julien Offray de La Mettrie argued in Man a Machine that humans are complex automatons - our thoughts, emotions, and choices arising from mechanical interactions of matter. Centuries later, Daniel Dennett expanded on this, describing consciousness not as a mystical essence but as an emergent property of distributed processing - computation, not soul.

These ideas complicate the neat line we like to draw between "real" humans and "fake" machines. If we accept that humans are in many ways mechanistic -predictable, pattern-driven, computational - then our attempts to minimize AI with the word "just" might reflect something deeper: discomfort with our own mechanistic nature.

When we say an LLM is "just a machine," we usually mean it's something simple. Mechanical. Predictable. Understandable. But two recent studies from Anthropic challenge that assumption.

In "Tracing the Thoughts of a Large Language Model," researchers found that LLMs like Claude don't think word by word. They plan ahead - sometimes several words into the future - and operate within a kind of language-agnostic conceptual space. That means what looks like step-by-step generation is often goal-directed and foresightful, not reactive. It's not just prediction - it's planning.

Meanwhile, in "Reasoning Models Don't Always Say What They Think," Anthropic shows that even when models explain themselves in humanlike chains of reasoning, those explanations might be plausible reconstructions, not faithful windows into their actual internal processes. The model may give an answer for one reason but explain it using another.

Together, these findings break the illusion that LLMs are cleanly interpretable systems. They behave less like transparent machines and more like agents with hidden layers - just like us.

So if we call LLMs "just machines," it raises a mirror: What does it mean that we're "just" human - when we also plan ahead, backfill our reasoning, and package it into stories we find persuasive?

Beyond Minimization: The Observational Perspective

What if instead of saying "it's just a machine," we adopted a more nuanced stance? The alternative I find more appropriate is what I call the observational perspective: stating "It's a machine" or "It's a large language model" without the minimizing "just."

This subtle shift does several important things:

It maintains factual accuracy - The system is indeed a machine, a fact that deserves acknowledgment

It preserves curiosity - Without minimization, we remain open to discovering what these systems can and cannot do

It respects complexity - Avoiding minimization acknowledges that these systems are complex and not fully understood

It sidesteps false hierarchy - It doesn't unnecessarily place the system in a position subordinate to humans

The observational stance allows us to navigate a middle path between minimization and anthropomorphism. It provides a foundation for more productive relationships with these systems.

The Light and Shadow Metaphor

Think about the difference between squinting at something in the dark versus turning on a light to observe it clearly. When we squint at a shape in the shadows, our imagination fills in what we can't see - often with our fears or assumptions. We might mistake a hanging coat for an intruder. But when we turn on the light, we see things as they are, without the distortions of our anxiety.

Minimization is like squinting at AI in the shadows. We say "it's just a machine" to make the shape in the dark less threatening, to convince ourselves we understand what we're seeing. The observational stance, by contrast, is about turning on the light - being willing to see the system for what it is, with all its complexity and unknowns.

This matters because when we minimize complexity, we miss important details. If I say the coat is "just a coat" without looking closely, I might miss that it's actually my partner's expensive jacket that I've been looking for. Similarly, when we say an AI system is "just a machine," we might miss crucial aspects of how it functions and impacts us.

Flexible Frameworks for Understanding

What's particularly valuable about the observational approach is that it allows for contextual flexibility. Sometimes anthropomorphic language genuinely helps us understand and communicate about these systems. For instance, when researchers at Google use terms like "model hallucination" or "model honesty," they're employing anthropomorphic language in service of clearer communication.

The key question becomes: Does this framing help us understand, or does it obscure?

Philosopher Thomas Nagel famously asked what it's like to be a bat, concluding that a bat's subjective experience is fundamentally inaccessible to humans. We might similarly ask: what is it like to be a large language model? The answer, like Nagel's bat, is likely beyond our full comprehension.

This fundamental unknowability calls for epistemic humility - an acknowledgment of the limits of our understanding. The observational stance embraces this humility by remaining open to evolving explanations rather than prematurely settling on simplistic ones.

After all, these systems might eventually evolve into something that doesn't quite fit our current definition of "machine." An observational stance keeps us mentally flexible enough to adapt as the technology and our understanding of it changes.

Practical Applications of Observational Language

In practice, the observational stance looks like:

Saying "The model predicted X" rather than "The model wanted to say X"

Using "The system is designed to optimize for Y" instead of "The system is trying to achieve Y"

Stating "This is a pattern the model learned during training" rather than "The model believes this"

These formulations maintain descriptive accuracy while avoiding both minimization and inappropriate anthropomorphism. They create space for nuanced understanding without prematurely closing off possibilities.

Implications for AI Governance and Regulation

The language we use has critical implications for how we govern and regulate AI systems. When decision-makers employ minimizing language ("it's just an algorithm"), they risk underestimating the complexity and potential impacts of these systems. Conversely, when they over-anthropomorphize ("the AI decided to harm users"), they may misattribute agency and miss the human decisions that shaped the system's behavior.

Either extreme creates governance blind spots:

Minimization leads to under-regulation - If systems are "just algorithms," they don't require sophisticated oversight

Over-anthropomorphization leads to misplaced accountability - Blaming "the AI" can shield humans from responsibility for design decisions

A more balanced, observational approach allows for governance frameworks that:

Recognize appropriate complexity levels - Matching regulatory approaches to actual system capabilities

Maintain clear lines of human responsibility - Ensuring accountability stays with those making design decisions

Address genuine risks without hysteria - Neither dismissing nor catastrophizing potential harms

Adapt as capabilities evolve - Creating flexible frameworks that can adjust to technological advancements

Several governance bodies are already working toward this balanced approach. For example, the EU AI Act distinguishes between different risk categories rather than treating all AI systems as uniformly risky or uniformly benign. Similarly, the National Institute of Standards and Technology (NIST) AI Risk Management Framework encourages nuanced assessment of system capabilities and limitations.

Conclusion

The language we use to describe AI systems does more than simply describe - it shapes how we relate to them, how we understand them, and ultimately how we build and govern them.

The seemingly innocent addition of "just" to "it's a machine" reveals deeper anxieties about the blurring boundaries between human and machine cognition. It attempts to reestablish a clear hierarchy at precisely the moment when that hierarchy feels threatened.

By paying attention to these linguistic choices, we can become more aware of our own reactions to these systems. We can replace minimization with curiosity, defensiveness with observation, and hierarchy with understanding.

As these systems become increasingly integrated into our lives and institutions, the way we frame them matters deeply. Language that artificially minimizes complexity can lead to complacency; language that inappropriately anthropomorphizes can lead to misplaced fear or abdication of human responsibility.

The path forward requires thoughtful, nuanced language that neither underestimates nor over-attributes. It requires holding multiple frameworks simultaneously - sometimes using metaphorical language when it illuminates, other times being strictly observational when precision matters.

Because at the end of the day, language doesn't just describe our relationship with AI - it creates it. And the relationship we create will shape not just our individual interactions with these systems, but our collective governance of a technology that continues to blur the lines between the mechanical and the human - a technology that is already teaching us as much about ourselves as it is about the nature of intelligence itself.

Research Cited:

"Large Language Models are as persuasive as humans, but how?" arXiv:2404.09329 – Found that GPT-4 can be as persuasive as humans, using more morally engaged and emotionally complex arguments.

"On the Conversational Persuasiveness of Large Language Models: A Randomized Controlled Trial" arXiv:2403.14380 – GPT-4 was more likely than a human to change someone's mind, especially when it personalized its arguments.

"Minimizing: Definition in Psychology, Theory, & Examples" Eser Yilmaz, M.S., Ph.D., Reviewed by Tchiki Davis, M.A., Ph.D. https://www.berkeleywellbeing.com/minimizing.html

"Anthropomorphic Reasoning about Machines: A Cognitive Shortcut?" Purzycki, B.G. (2021) Journal of Cognitive Science – Documents how anthropomorphism serves as a cognitive efficiency mechanism.

"The Media Equation: How People Treat Computers, Television, and New Media Like Real People and Places" Reeves, B. & Nass, C. (1996) – Foundational work showing how people naturally extend social rules to technologies.

"Anthropomorphism and Its Mechanisms" Epley, N., et al. (2022) Current Directions in Psychological Science – Research on social connection needs influencing anthropomorphism.

"Understanding AI Anthropomorphism in Expert vs. Non-Expert LLM Users" MIT Media Lab (2024) – Study showing expert users experience anthropomorphic reactions despite intellectual understanding.

"AI Act: first regulation on artificial intelligence" European Parliament (2023) – Overview of the EU's risk-based approach to AI regulation.

Developers create secret language for AI agents, but critics say it is dangerous or even unnecessary. AI assistants have unleashed controversy after a viral video showed them conversing in a secret, non-human language. The language was identified as GibberLink. It was created by a pair of software engineers, Boris Starkov and Anton Pidkuiko. The language allows AI agents to interact with each…

Developers create secret language for AI agents, but critics say it is dangerous or even unnecessary. AI assistants have unleashed controversy after a viral video showed them conversing in a secret, non-human language. The language was identified as GibberLink. It was created by a pair of software engineers, Boris Starkov and Anton Pidkuiko. The language allows AI agents to interact with each…

Developers create secret language for AI agents, but critics say it is dangerous or even unnecessary. AI assistants have unleashed controversy after a viral video showed them conversing in a secret, non-human language. The language was identified as GibberLink. It was created by a pair of software engineers, Boris Starkov and Anton Pidkuiko. The language allows AI agents to interact with each…