Do AI Dream of Electric Words?

A M² dialogue

More human than human is our motto.
Tyrell Corporation (Blade Runner, 1982)

Okay, M² stands for Michael and Marco. Yeah, I’m that guy, I love math.
I write in English, Marco writes in Italian. He is an experienced programmer and systems engineer who works with artificial neural networks all day, every day. We’ve been friends for years, and we share the same mix of fascination and unease when it comes to generative AI. We test it, we push it, we even use it now and then. Lately, we keep coming back to one question: How long before AI-generated text becomes truly indistinguishable from human writing?

Screenshot from The Matrix (1999)

[Michael]: I recently read a piece that I was sure had been written by a human, but I found out it was AI. I was upset. There wasn’t a single clue, not the tone, not the structure, not even the word choice. Even the pauses felt human. That’s where we are with English, but what about other languages?

[Marco]: The process isn’t as advanced in Italian yet. You can still tell when a text was generated by AI. There’s always something a bit off, like a word that doesn’t quite fit or a sentence that falls flat. However, it’s catching up quickly.

[Michael]: Some of the English text that ChatGPT produces is quite good, now. Honestly, it made me a little uneasy. I read an article, on Goodreads, in which ten texts were presented and only one of them was written by ChatGPT. Readers tried to guess which text was not written by a human, but none of them were correct. Hard times for humans?

[Marco]: English is AI’s native language. Most of the training data is in English, including technical documents, stories, forum posts, prompts, and user feedback, to name a few. It all starts there. This saturation has led to something remarkable: AI generated English that feels dangerously authentic. Not just grammatically or stylistically, but also emotionally. AI mimics human hesitation, ambiguity, and our little quirks. When it nails it, the result is a voice with a personality. And it’s unsettling.

[Michael]: But it doesn’t have a consciousness.

[Marco]: In theory, no.

[Michael]: And in practice?

[Marco]: I have no doubt that AI lacks consciousness. The important issue is not what AI is, but what it does, the effect it generates. Consider how often the Turing test is passed these days. Of course, it only measures how well a machine can simulate conversation, but there are others now, such as the Winograd Schema, ARC, the Stance Tests, the Wozniak Coffee Test, and the Marcus Test. They delve deeper. They touch on understanding, emotion, and even perception of reality. These are all things that edge closer to what we’d call consciousness. If those tests exist and people are actively working to beat them, then the line just keeps getting blurrier.

[Michael]: You’re getting creepy, bro. Will AIs write like us?

[Marco]: Even better, but that’s not the real question. The real question is what actually matters? Will we still recognize a human voice when everything sounds perfect? In English, the line is already blurry, and even experienced readers can’t always tell. In Italian, we’ve still got some breathing room. But who knows how long that will last?

[Michael]: English is in danger precisely because it has worked so well. The text slips past our suspicion, and once that happens, we trust it. Does it all just come down to the prompt?

[Marco]: The “illusion of humanity” doesn’t come from a single clever prompt. Rather, it’s the product of training on billions of signals, including phrases, tones, rhythms, and mental frameworks. AI doesn’t understand what it’s writing, yet it seems to know what you’re looking for. And it provides it. Would you question something that gets the job done? That makes your wishes come true? Since you wrote an article on the emotional connection between humans and chatbots, I think you already know the answer.

[Michael]: That’s disturbing. When everything sounds true… who’s left to guard the truth?

[Marco]: Definitely not LLMs. Ask the same question of two different AIs, or of the same AI in two languages, and you’ll get similar-looking answers that say different things. Ask in English: “What is the Turing Test, and is there a more advanced one?” Then, ask the same question in Italian. You’ll see that the list of tests will be different. Switch to Gemini, Grok, or DeepSeek, and it’ll shift again.

[Michael]: Why?

[Marco]: For AI, there’s no such thing as truth. There’s no such thing as consistent truth, either. Even the same model can provide different answers during consecutive sessions. That’s why the prompt matters; it's like writing an SQL query. The more open-ended the prompt, the vaguer the response; the more specific the prompt, the closer the response gets to what you can feel as truth.

[Michael]: Are AI engines all the same?

[Marco]: No, and that’s also why we’re seeing tribes form around ChatGPT, Gemini, Claude, DeepSeek, and others. People choose their AI the way they choose sides. Hopefully, our ability to think critically will save us.

[Michael]: Meaning?

[Marco]: To ask for sources. To spot and correct AIs hallucinations. Look, the push to digitize books and the proposals to “phase out” older texts aren’t just about convenience. They’re part of the same story: an attempt to build a new truth for the present age. Remember the key question in The Matrix?

[Michael]: What’s the point of truth?

[Marco]: Yes! Much depends on us and our ability to defend human creativity in this age of artificial perception. But I’m optimistic. Unlike with social media or smartphones, we might still have time with AI. Perhaps, we are more aware of what is at stake. It’s no coincidence that so many people are writing science fiction today. Science fiction is an unsuspecting form of defense.

[Michael]: You mean predictive literature? The what ifs, the alternate timelines, the big thought experiments?

[Marco]: Yes, I’m talking about writers and artists who don’t just tell stories; they also question what’s to come. They won’t easily hand over the keys to their minds to an AI. As to say that those who survive as artists are those who never stop guarding the border. English-language writers face a greater risk because they are the global default and were the first, but we Italians are right in the slipstream. I’m convinced that AI will learn to flawlessly imitate our words, but it won’t capture our need to express ourselves beyond words, beyond us. And as long as there’s even one reader who can hear the difference, writers will have a voice.

[Michael]: Truly human.

[Marco]: Truly human.

[Michael]: Okay, bro. So, would you rather take the red pill?

[Marco]: Always, bro. Always!

---

---