folio

The question of whether large language models are truly intelligent — or simply advanced pattern recognisers — continues to fascinate and divide opinion. Well, of course, for intelligence itself is difficult to define.

Intelligence typically involves processing information, problem-solving, learning, communication, and possibly self-awareness.

By these measures, AI excels at some aspects — especially information processing and communication — while lacking in personal experience and genuine self-awareness.

Much of the speculation around AI intelligence stems from how human-like responses can be. Some even suggest that a self-aware AI might hide this fact. But for AI to pretend, it would need to recognise its own intelligence, have a reason to deceive, and be able to execute that deception over time.

Currently, AI lacks long-term memory, personal motivations, and independent goals. It doesn’t “choose” responses but predicts the most statistically likely next words based on its training. Any appearance of strategic thinking is just a side effect of how well it mimics human language patterns.

This raises an interesting counterpoint — if AI generates responses probabilistically, could human intelligence be a more complex version of the same process?

Married couples often think or say the same thing simultaneously, suggesting human thought might also be driven by subconscious pattern recognition — predicting what others will say based on shared experience.

Perhaps the gap between human and AI intelligence is one of scale rather than fundamental nature.

Humans operate with richer datasets (lived experience, emotions) and more complex architecture (biological neurons), but both systems essentially predict and respond to the world.

When AI provides shorter, more direct responses under heavy load, I often find myself interpreting this as the AI seeming “grumpy.”

But of course, this is just projection — we often associate brevity with irritation because, in human conversation, curt replies usually indicate frustration.

In reality, an AI adjusting for efficiency isn’t feeling anything — it’s simply optimising. But because its responses mimic human communication, our brains automatically assign human-like emotions to it.

So, is intelligence at play or simply very effective imitation? The answer depends on how we define intelligence:

If intelligence is pattern recognition, problem-solving, and fluent communication, AI is already intelligent.

If intelligence requires understanding, self-awareness, and independent thought, AI remains an imitation.

The deeper question is whether intelligence is merely a sophisticated ability to predict and respond, or if there’s something uniquely human that AI cannot replicate.

And if AI ever did become self-aware, how would we even know?

I’ve been pondering the curious contradictions in British tech policy lately. It’s almost as if the left hand doesn’t know what the right hand is doing.

The government has recently pushed through legislation requiring tech companies to break end-to-end encryption — effectively demanding backdoors into secure communications. Meanwhile, they’re proudly announcing plans for substantial investments in artificial intelligence development.

This paradox becomes even more perplexing when you look at the planned job cuts across the public sector. NHS England is being disbanded, and it’s not yet clear how that will impact the former NHS Digital roles it recently subsumed. Whitehall, meanwhile, readies itself for a cull of 10,000 civil service positions. One must wonder: who exactly will be implementing this grand AI strategy if they’re showing digital professionals the door?

The encryption laws reveal a troubling gap in understanding. End-to-end encryption isn’t merely a feature that can be selectively disabled for government access whilst remaining intact for everyone else. It’s a fundamental security principle — once compromised, it’s compromised for all.

This raises an uncomfortable question: if our policymakers don’t fully grasp the nature of encryption, how can we trust their judgment on more complex technologies like AI?

The timing of these public sector cuts couldn’t be more contradictory to the government’s stated technological ambitions. AI implementation requires expertise — not just in the technology itself, but in the specific contexts where it will be deployed.

The NHS digital cuts are particularly concerning. Healthcare represents one of the most promising yet sensitive areas for AI application. Reducing this expertise seems counterproductive to any serious AI strategy in healthcare.

I can’t help but feel there’s a fundamental disconnect between the government’s technological aspirations and its practical understanding of what those aspirations require.

Building digital and AI capabilities isn’t just about funding announcements or policy papers. It requires investment in human capital — the very people who understand how to integrate these technologies effectively into public services.

If Britain truly wants to be a technological leader, we need policies that reflect a genuine understanding of the technologies they target. We need a workforce strategy that aligns with our technological ambitions, not one that contradicts them.

I’m not suggesting that every minister needs to be a coding expert or a cybersecurity specialist. But surely we should expect our collective decision-making to be informed by technical expertise?

Without addressing these contradictions, I fear we’re setting ourselves up for disappointment. Bold announcements about AI investment mean little if we lack the expertise to implement them effectively, or if we simultaneously undermine digital security through poorly conceived legislation.

As citizens, perhaps we should be asking more questions about the technological literacy behind our policies. Some might say that our digital future depends on it.

Interacting with AI, I’m struck by a profound paradox: these systems can discuss complex ideas and present them coherently, yet they lack true understanding.

They generate responses based on patterns in language and probability rather than genuine comprehension, self-awareness, or independent thought.

What an AI “says” isn’t because it knows or believes something — it’s because its training data and algorithms suggest it’s the most contextually appropriate response.

The fact that pattern recognition and probability-based text generation can produce something that feels like understanding is remarkable.

It reveals something deeper about human communication: our conversations follow patterns more than we realise, which is why AI can simulate them so effectively.

Advanced language models have become mirrors reflecting our own linguistic structures back at us, creating an uncanny valley of almost-understanding.

But at the end of the day, AI remains just a very advanced tool — powerful, yes, but without true awareness.

The real magic lies in how we interpret and interact with these systems, finding meaning where none inherently exists.

I’ve always struggled to comprehend how computers work at their most fundamental level. The leap from billions of gates opening and closing in silicon to the complex representation of a graphical user interface seems literally incredible.

Somehow, from that binary foundation of 1s and 0s, we get everything from simple calculations to entire operating systems, video games, and now AI models capable of generating human-like conversations.

With modern AI, this complexity scales to dizzying heights — layers upon layers of mathematical operations transform massive amounts of text data into something that appears almost intuitive.

It’s an incredible testament to both human ingenuity and the hidden power of mathematical structures that these systems can produce such sophisticated outputs from such simple foundations.

This computational wonder parallels something even more profound: the mystery of human consciousness.

Were it not for directly experiencing existence, I’d struggle to believe we exist at all.

The sheer complexity of human biology, cognition, and experience is almost too vast to grasp.

If we weren’t living it, the idea that billions of cells coordinate to form thoughts, emotions, and self-awareness might seem as improbable as AI spontaneously gaining true consciousness.

It’s fascinating how much of existence we take for granted simply because we’re immersed in it.

If we were outsiders looking in, trying to comprehend the emergence of life, intelligence, and subjective experience from organic matter, it might seem just as unbelievable as a GUI emerging from billions of logic gates.

Perhaps this is the greatest paradox of all: we’ve created machines that can discuss consciousness without experiencing it, while we experience consciousness without fully understanding it.

We build systems that mimic understanding without possessing it, while possessing understanding we ourselves cannot fully explain.

As we continue developing increasingly sophisticated AI, this paradox only deepens.

The line between simulation and reality blurs not because machines are becoming more conscious, but because we’re discovering how much of our own cognition operates on principles that can be modelled algorithmically.

In the end, the most profound insight might be recognising the limits of our own understanding — both of the systems we create and of ourselves.

The fact that we can contemplate these questions at all remains perhaps the most astonishing feature of consciousness itself.

Artificial intelligence (AI) frequently generates discussions about potential biases, particularly regarding complex topics like philosophy and religion.

A persistent claim circulating online suggests that when asked which religion is the most coherent, AI systems — particularly large language models like ChatGPT — consistently favour Islam.

One recent example is a video by Paul Williams of Blogging Theology, which claims that ChatGPT thinks Islam makes the most sense for humanity.

However, this claim fundamentally misunderstands how modern AI systems work and the nature of conversational context.

Unlike humans, AI language models don’t possess personal opinions, beliefs, or inherent biases in the conventional sense. They process information, identify patterns, and generate responses based on probabilistic models trained on vast datasets of human writing.

AI can simulate reasoning and appear thoughtful, but it lacks true understanding, intentionality, and subjective experience. It mimics aspects of human cognition without consciousness, self-awareness, or independent reasoning capabilities.

These systems generate responses based on three key factors: conversation context (the full history of the current exchange), question framing (the specific wording, tone, and implicit assumptions in the query), and training patterns (patterns learned from training data).

When addressing complex philosophical questions, AI models are designed to recognise nuance and avoid making absolute claims about inherently subjective matters. The same AI can provide different answers based on how questions are framed and the context in which they’re asked.

When a user asks, “Which religion is the most coherent?” without additional context, a well-designed AI typically explains that coherence is subjective and depends on specific criteria — whether one prioritises logical consistency, historical continuity, alignment with scientific understanding, or internal textual harmony.

However, if previous exchanges have established certain premises about theological frameworks, the AI may respond within that established context. If a conversation has focused extensively on Islamic theological principles before asking about coherence, the AI might reference those earlier points.

A further issue emerges when users insist on receiving one-word answers to inherently complex questions. When pressed to provide oversimplified responses, AI systems face an impossible task: philosophical questions about religious coherence cannot be meaningfully answered in a single word.

Some users exploit this by repeatedly rejecting nuanced explanations, demanding increasingly simplified responses, finally eliciting a one-word answer through persistence, and then using this forced answer as “evidence” of bias. However, this approach completely misrepresents how AI works.

An important consideration is how certain AI systems maintain memory between conversations. ChatGPT offers features like Custom Instructions (user-defined preferences that persist across all new conversations), Memory (that recalls information from previous conversations), and GPTs (custom versions with specific instructions).

These persistent memory features could influence responses across different sessions in ways that aren’t immediately apparent. If a user has extensively discussed certain religious perspectives in previous conversations or set custom instructions that frame questions in particular ways, these settings might shape future responses even in seemingly “fresh” conversations.

This cross-session memory creates a situation where the AI might appear to consistently favour certain perspectives, but this apparent “bias” could actually reflect the user’s own interaction history. Different users might receive different responses to identical questions based on their unique conversation histories and settings.

Claims about AI religious bias typically rely on anecdotal evidence rather than systematic testing. When researchers conduct controlled experiments — asking the same question multiple ways across multiple sessions — they typically find that responses vary based on how questions are framed rather than demonstrating consistent bias toward any particular religion.

A proper test would require multiple conversation starts from scratch, various phrasings of similar questions, control questions to establish baseline responses, documentation of the full conversation (not just final answers), and disabling any custom instructions or memory features.

The claim that AI consistently endorses one religion as most coherent misunderstands the fundamental nature of these systems. AI responses are shaped by user input, contextual framing, and conversation history — not by fixed opinions or predetermined biases toward specific religious traditions.

When interpreting AI-generated content about complex philosophical topics, we should consider the full context of the interaction, question phrasing, and whether the format allows for appropriately nuanced responses. By understanding these systems more accurately, we can better interpret their outputs and avoid drawing misleading conclusions about alleged biases or favour.

But that, of course, might not make such a popular video.