AI Isn’t Your God—But It Might Be Your Intern

Was this newsletter forwarded to you? Sign up to get it in your inbox.

We've been told that AI will either save or destroy humanity…yet here we are watching LLMs hallucinate, image generators struggle to draw realistic human hands, and AI coding agents get stuck in loops without supervision.

This gap—the one between the promise of AI and the present reality—isn’t about AI progress slowing down. It’s a product of how we’re fundamentally misunderstanding the technology. 

We’ve cast AI in the role of a God-like entity, when we should be thinking of it more like an intern: an intern who is linguistically capable, sometimes makes decisions in ways we don’t quite understand, and perhaps most importantly—if we put in the time to work with them—has the potential to surprise us.

Why we’re disillusioned by AI

When Ingenuity, NASA’s autonomous helicopter that operated on Mars from 2021 to 2024, took its last flight, the team behind it recorded a video to bid the craft farewell, with one of them describing it as “a plucky little helicopter that just defied everybody's expectations.” According to the first line on its Wikipedia page, the helicopter even had a nickname, Ginny. Humans have a somewhat irrational tendency to anthropomorphize technology—even the most scientifically oriented of us. 

AI has lent itself to being portrayed in science fiction as an overlord, benevolent at times and terrifying at others. It has the reputation of being an intangible, all-knowing, mysterious “thing”—a reputation that broadly fits many people’s perception of God, or other similar higher power they have faith in. I think we’ve come to believe this heightened narrative around AI because:

• LLMs have the ability to speak and understand natural language, and we intuitively recognize language as a sign of intelligence.
• The inner workings of AI models are a black box, and we don’t fully grasp the complexities of how they function. 

These blind spots in our thinking breed irrationally high expectations of AI—and an inevitable feeling of being underwhelmed by it. Understanding them more deeply brings us closer to the truth, so we can set realistic expectations of AI and develop more productive ways of working with the technology.

Language as a sign of cognition

Language plays a big role in making our surroundings comprehensible to us, and our capability to use language to understand and make ourselves understood is one of the most remarkable manifestations of human intelligence. As a result, we instinctively measure the intelligence of other “things”—human and non-human—based on their linguistic capabilities. 

Our notions around what constitutes intelligence have strong ties to language, and I’ve noticed this play out in my own life: I attend Spanish language classes with five other students to whom I exclusively speak in Spanish, both in and out of the classroom. I often catch myself making small, secret judgments about each of their intellects based on how good their Spanish is, even though I know that they’re probably infinitely more articulate in their native languages. 

This concept is reflected in the Turing Test, which we've historically used to evaluate if a machine is approaching intelligence. The test involves a human interrogator asking text-based questions of two hidden subjects—one human and one computer program—with the intent of determining which one is human. A number of different people play the role of interrogator and respondent, and if enough interrogators are fooled into thinking the computer program is human, it is said to exhibit intelligence. 

The Turing Test was developed in 1949, but our inclination to regard language as a sign of intelligence can be traced back as early as 1726, when a simplified version of the Turing Test appeared in Jonathan Swift’s novel Gulliver's Travels. Gulliver was brought to the court of a king who suspected he was “a piece of clockwork…contrived by some ingenious artist” and had been taught him “a set of words” to make him “sell at a better price.” The king was only satisfied that Gulliver wasn’t a machine when he got rational answers to several questions put to him. 

LLMs, of course, can understand and speak natural language—and that explains our tendency to attribute intelligence to them. I’m continually surprised (and delighted) by Claude and ChatGPT’s ability to understand my inarticulate, typo-filled prompts—often describing it as “uncanny”—even though I’m rationally aware of the broad strokes of how LLMs actually work. 

Whether LLMs truly “understand” and “speak” to us is a more philosophical question. The Chinese room argument proposed by philosopher John Searle in 1980 is a thought experiment involving a person who follows instructions to manipulate Chinese symbols without understanding their meaning but, to an outsider, could appear to “understand” Chinese. The experiment makes a distinction between machines following rules to arrange words or symbols (syntactic ability) and truly understanding their meaning (semantic ability). Either way, our subjective experience of using LLMs is that their language capabilities are similar to ours. Consequently, we perceive their general intelligence to be at least comparable to our own. 

What lies inside an LLM?

We know that the underlying logic of LLMs is next-token prediction, a process where the model predicts the next word by choosing the one that’s most likely to follow based on patterns learned from their training data. But next-token prediction doesn’t fully explain questions like how the model internally represents and organizes knowledge, or how these representations influence its output. The precise inner workings of LLMs evade us. 

Anthropic has conducted research that maps examined neuron activations—the internal state of an LLM before it generates a response—and mapped patterns of them, called features, with real-world concepts. In Claude, for instance, Anthropic was able to identify the feature that represented the concept of the Golden Gate Bridge and released a temporary experimental model where this feature was amplified, leading the AI to reference the bridge more frequently, even in unrelated contexts. This early research is promising, but still developing. The sense that we don’t quite understand AI systems feeds a sense of mysticism around the technology, leading us to overestimate its capabilities.

The intern, not the omniscient oracle 

When we perceive AI as an intangible overlord—something bigger than us—we expect the technology to do the heavy lifting on any task we give it. We begin each interaction with an LLM under the assumption that it’s powerful enough to ace it with little to no effort on our end—it is, after all, more capable than us, right? This way of thinking is only compounded by the fact that most modern consumer technologies are predictable and easy to use. 

We’ve grown used to technology that nearly always gets it right. When AI doesn’t live up to these expectations, our disappointment sours into the risk of abandoning the technology altogether. 

Instead, turn that on its head: Think of AI as an intern. Go in with the expectation that you will have to work with the technology for it to be valuable to you. Internalize that its first attempt may not be the best one, and be intentional about how much direction you give it—more for a specific deliverable, less for an open-ended brainstorm. Get into the habit of making notes about what you liked in the AI response and what you didn’t—articulating your feedback to the LLM goes a long way. 

Just like any good intern, AI’s greatest value lies in our willingness to engage with the technology—in the space between what it offers and what we can accomplish working together with it.

Rhea Purohit is a contributing writer for Every focused on research-driven storytelling in tech. You can follow her on X at @RheaPurohit1 and on LinkedIn, and Every on X at @every and on LinkedIn.

We also build AI tools for readers like you. Automate repeat writing with Spiral. Organize files automatically with Sparkle. Write something great with Lex.