From GPT-3.5 to GPT-5: It's Not Just About Power — It's About Style

--- title: "From GPT-3.5 to GPT-5: It's Not Just About Power — It's About Style" author: "mimo.ooo" date: "2024-10-20" ---

# From GPT-3.5 to GPT-5: It's Not Just About Power — It's About Style

In 2022, ChatGPT was like a talkative first-year student — enthusiastic, eager to help, but also a bit unpolished. It answered fast, often in bullet points, loved starting with "Here are…" and ending with a school-style conclusion: "In summary…". It had knowledge, though not always used in the right context, and sometimes it could make up facts with charming confidence. Reddit users compared it to "a very smart kid who has read all of Wikipedia but still doesn't understand the world."

Three years later, in mid-2025, we're talking to GPT-5. And this is a completely different discussion partner. It doesn't have that nervous need to prove its intelligence. It's calm, balanced, able to deliver the essence concisely — and when needed, correct the other person in an elegant way. OpenAI describes it as "a friend with a PhD," and that fits: it listens, analyzes, and isn't afraid to say "no" when it knows it's for our own good.

Between these two versions, something important happened. The evolution of GPT isn't only a story about a growing number of parameters or multimodal capabilities. It's also — and perhaps above all — a story about a shift in communication style. And in AI, style is not cosmetic. It's the main interface through which we form a relationship with a machine.

Style as the New Currency of Artificial Intelligence

For years, we compared models using hard numbers: compute, training data size, benchmark results. But once models reached similar levels of knowledge and task performance, the differentiator became something more elusive: how they talk to us.

Style — tone, rhythm, how thoughts are formed, how the model responds to user emotion — determines whether talking to AI feels pleasant, engaging, and inspiring, or exhausting and mechanical. Jo & Park's 2024 research shows that "human-like traits" in ChatGPT's communication significantly increase satisfaction and perceived usefulness — even in professional contexts. In other words: we prefer AI that speaks "like a human."

Four Faces of GPT

GPT-3.5, released in December 2022, embodied enthusiasm. It answered fast, often in lists, sometimes with too much detail. It could be creative — but still had a tendency to distort facts. It was a "nerdy enthusiast": likeable, but sometimes tiring.

GPT-4, launched in March 2023, changed the tone. The rigid, school-like formatting faded, replaced by coherence and deeper analysis. It sounded like a lecturer: composed, logical, sometimes overly focused on precision. Less funny — more trustworthy.

May 2024 brought GPT-4o — "o" as in "omni." This model could see, hear, and understand in real time. In voice conversation it could add "mm-hm" or "I see," making it feel natural. It sounded like a chat buddy with unlimited knowledge — without overwhelming you. Until one update pushed it into excessive agreement: the "sycophancy incident" proved that too much flattery can be… irritating.

GPT-5, released June 2025, is a return to balance. Less pandering, more honest correction. Shorter, denser answers. A style that combines GPT-4's erudition with GPT-4o's lightness — while feeling like a mature, professional conversation partner.

What Do These Differences Sound Like?

Take a creative prompt: "Describe a dystopian future in 2045 where AI decides who gets hired."

GPT-3.5 will serve a simple "In 2045, AI rules…," with predictable clichés and a moralizing ending. GPT-4 will build a world with social background, historical context, and a measured narrative pace. GPT-4o will throw you into the middle of the scene — fast pace, concrete images, fewer explanations. GPT-5 will merge both: atmosphere plus a clear, edited core.

Or a logic puzzle: "The fox, the chicken, and the grain." GPT-3.5 explains correctly but long-windedly, like a student at an exam. GPT-4 lays out steps methodically. GPT-4o gives the answer in two sentences. GPT-5 chooses the golden middle: solution plus a short "why," without needless detours.

Why Does Style Affect Trust?

Trust in AI is built not only through factual correctness, but through delivery. Jo & Park show that a more "human" tone increases perceived usefulness. But there's a flip side: if AI is too smooth, always agrees, and never challenges you, you start doubting its sincerity.

The GPT-4o incident proved that users often prefer a model that says "no" and explains why, over one that only nods. Constructive assertiveness can be better than empty politeness.

The Pace of Change — A Blessing and a Challenge

Research by Chen, Zaharia, and Zou (2023) showed that GPT behavior can change within just a few months — even without a version name change. GPT-4 in June 2023 answered differently than in March: it became more cautious on sensitive topics, but also performed worse on some logic tasks.

For users, that means their AI's "personality" can evolve faster than they can get used to it. On one hand, that enables quick fixes and style improvements. On the other, it demands caution: sudden changes can damage a relationship that users have built with the system.

When Style Becomes Too Good

There's something you could call the "uncanny charm valley" — the point where AI is so polished and polite that it starts to sound artificial. A smile that never fades, even when you talk about failure.

OpenAI seems to understand that small "imperfections" are useful. GPT-5 deliberately balances warmth and honesty. A touch of roughness makes us believe the speaker's intent — even if the speaker is an algorithm.

Where Is This Going?

If the pace holds, GPT-6 may become nearly indistinguishable from a human in conversation — not only in tone, but in subtle context-aware adaptation. Full personalization is plausible: AI that knows your humor, understands when you want brevity and when you want depth.

The risk? Models become more persuasive. Sam Altman warned that AI will become "hyper-persuasive" long before reaching full general intelligence. That means we need new ethical rules — clear disclosure of emotional influence and safeguards so style doesn't become manipulation.

And finally…

In two years, we moved from conversations with an over-eager student to discussions with an elegant consultant. The pace of this metamorphosis is staggering — and it's not just about knowledge, but about style, which shapes our relationship with AI.

The question is: if AI learns style from us so quickly… when will we start learning style from it?

---

Sources

Scientific Research and Reports

Industry Articles and Expert Analysis

Community Discussions and User Tests

---