Human rights principles

From GPT-4 to 4.5: What has changed and why it matters

In March 2024, OpenAI released GPT-4.5, the latest iteration in its series of large language models (LLMs), pushing the boundaries of what machines can do with language understanding and generation. Building on the strengths of GPT-4, its successor, GPT-4.5, demonstrates improved reasoning capabilities, a more nuanced understanding of context, and smoother, more human-like interactions.

What sets GPT-4.5 apart from its predecessors is that it showcases refined alignment techniques, better memory over longer conversations, and increased control over tone, persona, and factual accuracy. Its ability to maintain coherent, emotionally resonant exchanges over extended dialogue marks a turning point in human-AI communication. These improvements are not just technical — they significantly affect the way we work, communicate, and relate to intelligent systems.

The increasing ability of GPT-4.5 to mimic human behaviour has raised a key question: Can it really fool us into thinking it is one of us? That question has recently been answered — and it has everything to do with the Turing Test.

The Turing Test: Origins, purpose, and modern relevance

In 1950, British mathematician and computer scientist Alan Turing posed a provocative question: ‘Can machines think?’ In his seminal paper ‘Computing Machinery and Intelligence,’ he proposed what would later become known as the Turing Test — a practical way of evaluating a machine’s ability to exhibit intelligent behaviour indistinguishable from that of a human.

In its simplest form, if a human evaluator cannot reliably distinguish between a human’s and a machine’s responses during a conversation, the machine is said to have passed the test. For decades, the Turing Test remained more of a philosophical benchmark than a practical one.

Early chatbots like ELIZA in the 1960s created the illusion of intelligence, but their scripted and shallow interactions fell far short of genuine human-like communication. Many researchers have questioned the test’s relevance as AI progressed, arguing that mimicking conversation is not the same as true understanding or consciousness.

Despite these criticisms, the Turing Test has endured — not as a definitive measure of machine intelligence, but rather as a cultural milestone and public barometer of AI progress. Today, the test has regained prominence with the emergence of models like GPT-4.5, which can hold complex, context-aware, emotionally intelligent conversations. What once seemed like a distant hypothetical is now an active, measurable challenge that GPT-4.5 has, by many accounts, overcome.

How GPT-4.5 fooled the judges: Inside the Turing Test study

In early 2025, a groundbreaking study conducted by researchers at the University of California, San Diego, provided the most substantial evidence yet that an AI could pass the Turing Test. In a controlled experiment involving over 500 participants, multiple conversational agents—including GPT-4.5, Meta’s LLaMa-3.1, and the classic chatbot ELIZA—were evaluated in blind text-based conversations. The participants were tasked with identifying whether they spoke to a human or a machine.

The results were astonishing: GPT-4.5 was judged to be human in 54% to 73% of interactions, depending on the scenario, surpassing the baseline for passing the Turing Test. In some cases, it outperformed actual human participants—who were correctly identified as human only 67% of the time.

That experiment marked the first time a contemporary AI model convincingly passed the Turing Test under rigorous scientific conditions. The study not only demonstrated the model’s technical capabilities—it also raised philosophical and ethical questions.

What does it mean for a machine to be ‘indistinguishable’ from a human? And more importantly, how should society respond to a world where AI can convincingly impersonate us?

Measuring up: GPT-4.5 vs LLaMa-3.1 and ELIZA

While GPT-4.5’s performance in the Turing Test has garnered much attention, its comparison with other models puts things into a clearer perspective. Meta’s LLaMa-3.1, a powerful and widely respected open-source model, also participated in the study.

It was identified as human in approximately 56% of interactions — a strong showing, although it fell just short of the commonly accepted benchmark to define a Turing Test pass. The result highlights how subtle conversational nuance and coherence differences can significantly influence perception.

The study also revisited ELIZA, the pioneering chatbot from the 1960s designed to mimic a psychotherapist. While historically significant, ELIZA’s simplistic, rule-based structure resulted in it being identified as non-human in most cases — around 77%. That stark contrast with modern models demonstrates how far natural language processing has progressed over the past six decades.

The comparative results underscore an important point: success in human-AI interaction today depends on language generation and the ability to adapt the tone, context, and emotional resonance. GPT-4.5’s edge seems to come not from mere fluency but from its ability to emulate the subtle cues of human reasoning and expression — a quality that left many test participants second-guessing whether they were even talking to a machine.

The power of persona: How character shaped perception

One of the most intriguing aspects of the UC San Diego study was how assigning specific personas to AI models significantly influenced participants’ perceptions. When GPT-4.5 was framed as an introverted, geeky 19-year-old college student, it consistently scored higher in being perceived as human than when it had no defined personality.

The seemingly small narrative detail was a powerful psychological cue that shaped how people interpreted its responses. The use of persona added a layer of realism to the conversation.

Slight awkwardness, informal phrasing, or quirky responses were not seen as flaws — they were consistent with the character. Participants were more likely to forgive or overlook certain imperfections if those quirks aligned with the model’s ‘personality’.

That finding reveals how intertwined identity and believability are in human communication, even when the identity is entirely artificial. The strategy also echoes something long known in storytelling and branding: people respond to characters, not just content.

In the context of AI, persona functions as a kind of narrative camouflage — not necessarily to deceive, but to disarm. It helps bridge the uncanny valley by offering users a familiar social framework. And as AI continues to evolve, it is clear that shaping how a model is perceived may be just as important as what the model is actually saying.

Limitations of the Turing Test: Beyond the illusion of intelligence

While passing the Turing Test has long been viewed as a milestone in AI, many experts argue that it is not the definitive measure of machine intelligence. The test focuses on imitation — whether an AI can appear human in conversation — rather than on genuine understanding, reasoning, or consciousness. In that sense, it is more about performance than true cognitive capability.

Critics point out that large language models like GPT-4.5 do not ‘understand’ language in the human sense – they generate text by predicting the most statistically probable next word based on patterns in massive datasets. That allows them to generate impressively coherent responses, but it does not equate to comprehension, self-awareness, or independent thought.

No matter how convincing, the illusion of intelligence is still an illusion — and mistaking it for something more can lead to misplaced trust or overreliance. Despite its symbolic power, the Turing Test was never meant to be the final word on AI.

As AI systems grow increasingly sophisticated, new benchmarks are needed — ones that assess linguistic mimicry, reasoning, ethical decision-making, and robustness in real-world environments. Passing the Turing Test may grab headlines, but the real test of intelligence lies far beyond the ability to talk like us.

Wider implications: Rethinking the role of AI in society

GPT-4.5’s success in the Turing Test does not just mark a technical achievement — it forces us to confront deeper societal questions. If AI can convincingly pass as a human in open conversation, what does that mean for trust, communication, and authenticity in our digital lives?

From customer service bots to AI-generated news anchors, the line between human and machine is blurring — and the implications are far from purely academic. These developments are challenging existing norms in areas such as journalism, education, healthcare, and even online dating.

How do we ensure transparency when AI is involved? Should AI be required to disclose its identity in every interaction? And how do we guard against malicious uses — such as deepfake conversations or synthetic personas designed to manipulate, mislead, or exploit?

On a broader level, the emergence of human-sounding AI invites a rethinking of agency and responsibility. If a machine can persuade, sympathise, or influence like a person — who is accountable when things go wrong?

As AI becomes more integrated into the human experience, society must evolve its frameworks not only for regulation and ethics but also for cultural adaptation. GPT-4.5 may have passed the Turing Test, but the test for us, as a society, is just beginning.

What comes next: Human-machine dialogue in the post-Turing era

With GPT-4.5 crossing the Turing threshold, we are no longer asking whether machines can talk like us — we are now asking what that means for how we speak, think, and relate to machines. That moment represents a paradigm shift: from testing the machine’s ability to imitate humans to understanding how humans will adapt to coexist with machines that no longer feel entirely artificial.

Future AI models will likely push this boundary even further — engaging in conversations that are not only coherent but also deeply contextual, emotionally attuned, and morally responsive. The bar for what feels ‘human’ in digital interaction is rising rapidly, and with it comes the need for new social norms, protocols, and perhaps even new literacies.

We will need to learn not only how to talk to machines but how to live with them — as collaborators, counterparts, and, in some cases, as reflections of ourselves. In the post-Turing era, the test is no longer whether machines can fool us — it is whether we can maintain clarity, responsibility, and humanity in a world where the artificial feels increasingly real.

GPT-4.5 may have passed a historic milestone, but the real story is just beginning — not one of machines becoming human, but of humans redefining what it means to be ourselves in dialogue with them.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!