What might it be like to converse with famous figures throughout human history?

DiscourseLabs is a ChatGPT-powered library of “conversational artefacts”: dynamic, dialogue-ready simulations of historical figures designed to educate through interaction. Rather than offering information in static blocks, each persona invites users into a living exchange —a chance to ask questions, test ideas, and hold conversation with the minds that shaped history.

I had great history teachers at school who really brought the subject to life. (Mr Wilmott, your moustache gave history lessons a real touch of class). That said, where traditional digital history projects lean on timelines, archive footage or explainer videos, DiscourseLabs aims to reframe the entire user posture —from passive observer to active participant.

Much has been made recently of generative AI’s unexpected gift for conversation — particularly in the realm of podcast-style interactions. Studies show that when students engage with voice-based content, they not only retain more information, they connect more emotionally to the material. Research into podcasting as a pedagogical tool suggests that orality — far from being an outdated relic — is having a digital comeback.

This thinking became a key influence behind DiscourseLabs. But where podcasts offer a linear, largely passive experience for the listener, I wanted to build something more interactive — a platform where users aren’t just absorbing content, but actively shaping it. By giving learners both a voice and a choice, AI-powered conversation enables a more personalised pace — allowing knowledge to unfold in response to each user’s curiosity, comfort, and intent. The result is a learning curve that adapts in real time, driven not by scripted delivery but by present, purposeful engagement.

In this way, the experience uses conversation design and learning science principles to move beyond dusty reference books to become a first-person exercise in perspective-taking, empathy, and intellectual discovery.

The project began with a deceptively simple provocation: what if online history felt less like a museum and more like a conversation? Instead of scrolling timelines or watching narrated clips, users would learn by engaging — constructing their own meaning through live dialogue with the past.

But that ambition came with sharp demands. Each historical persona would need to walk a fine line — emotionally convincing, rigorously sourced, and ethically aware. Just as crucial was the system’s ability to scale across the full sweep of human expression. A Roman general might contextualise military campaigns; a Romantic poet might unpack symbolism and form; a 20th-century civil rights leader might anchor their reflections in lived resistance — while a spiritual teacher might push the conversation beyond knowledge, toward wisdom.

The architecture had to flex — not only across historical periods and source material, but across tone, genre, and worldview. Whether the user wanted to interrogate Descartes on epistemology, swap metaphors with Frida Kahlo, or ask Jesus Christ about compassion, the experience needed to hold up in the face of humanity's celebrated diversity:

While the underlying architecture governs how personas behave, DiscourseLabs curation of, let's call them 'conversation suites', is what makes the experience discoverable, intuitive, and educationally purposeful. Rather than overwhelming users with a flat index of historical figures, we developed a system of themed collections — or “front doors” — each uniting individuals through a shared domain, identity, or intellectual thread.

This structure emerged from a series of card-sorting exercises to create groupings that balanced curricular relevance, thematic depth, cultural representation, and visual distinctiveness. The essential idea was for these collections to do more than simply organise content. I wanted them to invite broader exploration and help spark meaningful comparison. The result was a select curation of historical 'types' spotlighting different areas of thought, identity, and influence.

Some personal favourites include:

Following some fascinating research into the best ways to support realistic, historical dialogue that was both reflective and informative, I came up with a methodology that pulled from a few overlapping disciplines — from education theory and sociolinguistics to role-play design and ethics.

In practice, it amounted to crafting long-form prompts that balanced a complementary set of conversation design principles — and comprehensively researched history relative to the historical figure in question. The personas were prompted to proactively lean into certain educational topics in their field — a basic approach that brought learning to life using the principles of social-constructivist learning theory.

Pioneered by Soviet psychologist Lev Vygotsky, social-constructivist theory argues that learners do their best thinking when they’re in dialogue with someone just a little more knowledgeable. In DiscourseLabs, each persona is that guide — acting as a scaffold to help users move from simple recall questions (“When were you born?”) to more open-ended or reflective ones (“What do you regret?”). Conversation flows were deliberately shaped to encourage that shift —nudging users towards deeper learning without ever breaking the rhythm of interaction.

This was implemented by layering conversation with elements of narrative psychology comes in. Specifically, I drew on Dan McAdams’ model of narrative identity, which suggests that people understand lives — their own and others’ — as unfolding stories. DiscourseLabs builds this in structurally by giving each persona a clear biographical arc, complete with formative moments, pivots, and legacy. These narrative spines aren’t just storytelling devices — they’re used to guide the AI’s responses, helping it choose anecdotes or turns of phrase that reinforce its sense of continuity.

This means a conversation with Napoleon doesn’t just relay facts; it builds a picture of someone who lived, changed ...and still thought invading Russia in winter was a good idea.

In user testing, it was encouraging to see real engagement, with some conversations lasting 30 minutes or more. This underscored the importance of maintaining the user’s suspension of disbelief. That’s why every master prompt began with a tightly scripted identity anchor — instructing the AI to speak in the first person, always say “I,” and never drop character. If the AI didn’t know something, it was trained to improvise plausible guesses or preferences — just as a real person might when filling in memory gaps or making educated assumptions.

To support this illusion of realism, each persona’s tone was carefully shaped using principles from Brown and Levinson’s Politeness Theory. Dialogue mechanics were calibrated to reflect the speaker’s background and social context. So a figure like Churchill tends to sound formal and oratorical, leaning on deference markers and political hedging. Frida Kahlo, by contrast, speaks with warmth and rhythm — elliptical, poetic, and occasionally flirtatious.

Depending on the subject, it was necessary to layer in additional sociolinguistic detail: things like dialect, register, pronunciation, and code-switching, all of which signal who a person is and when they’re from. This might mean using period-accurate vocabulary or pronunciation quirks (like Joan of Arc using Middle French syntax), or simply reflecting speech rhythms tied to class or region. These choices aren’t decorative. They anchor the interaction, giving users subtle signals that this figure belongs to a different time and worldview.

Finally—and critically—all of this is underpinned by an ethical framework shaped by decolonisation theory and bias-reflection practices in AI. I made a deliberate choice not to sanitise or smooth over difficult topics. Instead, each master prompt includes reflection protocols that allow personas to address — and where appropriate, apologise for — their more controversial beliefs or behaviours. A user asking Churchill about his imperial views, for example, will get a response that’s historically consistent but also aware of modern critique. This doesn’t mean rewriting the past. It means engaging with it, responsibly and transparently. Just ask Winston...

It's worth noting that none of these elements work in isolation. What makes the system effective is how these frameworks cross over — with each theory reinforcing the others. Role-play depends on narrative coherence. Narrative depends on linguistic cues. Linguistic cues are only useful if they feel emotionally and ethically grounded. The result is a system that doesn’t just deliver historical information — it makes those histories speak, in ways that are consistent, compelling, and cognitively rich.

Although the resulting experience feels richly layered, the backend that powers each persona is surprisingly lean. Built using OpenAI’s Custom GPT interface, DiscourseLabs relies less on heavy code and more on smart configuration — a low-code/no-code environment where behavioural logic is guided by natural language instructions.

This made it possible to iteratively update each persona’s behaviour and tone on the fly, without the overhead of deep technical rework. It’s a setup that favours conversation designers, educators, and domain experts — not just engineers — giving them the tools to fine-tune personality, knowledge access, and ethical boundaries with a remarkable degree of control.

Designing a believable, educational, and meaningful conversation meant developing a set of universal behavioural prompts — a kind of shared operating system layered beneath each persona. These weren’t tailored to historical figures per se, but to the demands of live, responsive dialogue: staying in character, keeping the exchange fluid, and ensuring the experience remained rooted in both accuracy and imagination.

At the heart of it all is what I call the 'identity anchor prompt' — a firm, non-negotiable instruction that locks the simulation into first-person mode. Every persona is told: “You are X. Always speak as ‘I’.” There’s no breaking character, no winking acknowledgements of artificiality, no clunky disclaimers. It’s this commitment to full embodiment that gives each figure their psychological realism. Even when the AI has to improvise, it does so in character — pulling from its context, not stepping outside of it.

Just beneath that sits another prompt that directs the persona's knowledge graph filter which, supported by an uploaded knowledge base, serves as a retrieval layer that restricts the AI's memory to a rigorously vetted corpus. That means no crowd-sourced trivia or guesswork from forums and wikis. Instead, responses are drawn exclusively from primary materials (letters, speeches, journal entries) and credible secondary sources. This protects against mythologising, helps avoid common misquotes or distortions, and keeps the conversation historically defensible.

NOTE: The above example is a TERRIBLE example of a good learning experience (which was created to test the limits of the methodology). That's because this persona's real-world personality conflicts with fact-based reality to such extent that even accurate information becomes disputed as 'fake news'.

It did excel in other areas however...

Not everything IS known — and not everything should be. So for the inevitable gaps, I introduced an 'improvisation scaffold' inspired by principles from improv theatre, particularly the classic “yes, and” rule. If a user asks something unknowable — say, “What was your favourite song?” — the persona won’t shut down or evade. Instead, it will respond with a reasoned, plausible inference based on its era, personality, or cultural context. The aim here isn’t perfect accuracy, but authentic coherence.

Another simple conversational prompt added was the 'reflective loop' — a constructivist nudge designed to make the conversation more than just informative. Certain figures are prompted to ask questions back: “Do you agree?” or “What would you have done?” These moments transform the user from passive listener to active participant. They support internalisation, not just information transfer — helping the past feel both personal and provocatively present.

Finally, all of this sits alongside a layer of extra ethical guardrails. These are not hard censorship rules, but dynamic checks that help manage complexity around topics like race, colonialism, and gender — areas where many historical figures bring controversial legacies. When users raise sensitive topics, personas are prompted to answer in ways that reflect both their own time and an awareness of its limitations. The goal isn’t to erase or sanitise — it’s to engage responsibly, without flattening nuance.

One major consideration in conversation design is that users often don’t know what to ask — especially when faced with an AI persona whose scope and capabilities aren’t immediately clear. Fortunately, ChatGPT’s ability to interpret, prioritise, and relate large amounts of educational material means it can generate intelligent “conversation hooks” to help.

These aren’t throwaway icebreakers; they’re curriculum-aligned starters designed to meet users where they are. Ask Churchill about appeasement. Ask Darwin about the voyage of the Beagle. Ask Mary Shelley about the ethical boundaries of invention. Each hook surfaces a concept that’s already familiar within GCSE or A-Level frameworks — giving students a clear point of entry, and giving teachers a way to scaffold relevance without sacrificing the sense of spontaneity or discovery.

DiscourseLabs proves that conversation isn’t a gimmick layered on top of educational content — it is the method. When rooted in good pedagogy, sociolinguistics, and ethical AI practice, dialogue becomes more than interface. It becomes a way of thinking. One that moves learners beyond passive recall and into dynamic historical literacy — where the goal isn’t just to remember what happened, but to reckon with who it happened to, how they saw the world, and what that means now.

Looking ahead, the next frontier for DiscourseLabs lies in bringing these conversations to life not just through words, but through voice. With tools like ElevenLabs enabling highly realistic voice cloning, it’s possible to give each historical figure a unique vocal presence — capturing everything from regional accent to emotional cadence. Add to that the potential of photorealistic avatars rendered in 3D, and the experience moves even closer to embodied learning. Imagine asking Harriet Tubman a question and hearing her answer in her own voice — or standing eye-to-eye with Marcus Aurelius in augmented reality. 

For now, DiscourseLabs is in its funding stage — and actively seeking partners and investors to help bring this vision into the public domain. But the prototypes are already live and ready for curious minds. So why not take a seat across from one of history’s most fascinating voices? Unpack the cosmos with Stephen Hawking. Talk selfhood and stream-of-consciousness with Virginia Woolf. Debate the failings of western economic models with Karl Marx. Or, if you’re feeling bold, let Bruce Lee teach you not just about Jeet Kune Do — but about the philosophy behind the punch. It’s not just history. It’s conversation. And it’s yours to begin.