Aligning AI With Tone of Voice
A new way to generate factual content in any style using models like GPT-4
As marketers delve into the world of Generative AI, the question inevitably arises: how do we ensure that our brand's voice shines through?
It's an important question to ask.
By default, foundational models like OpenAI's GPT-4 bias toward predictability. They are trained to reliably tell you that the fourth planet from the Sun is Mars, so when you ask them to write something you’ll tend to get back predictable copy.
But fine-tuning a model to align tonality can be problematic. Not only is it a challenge to source training content that is consistent and high-quality, but the process decreases generalizability. In most cases, it’s not a practical solution.
Tonality via Automated Prompt Engineering
This post will explore a new approach we've developed to align Large Language Models with a specific tone of voice using Automated Prompt Engineering.
We start with two data sources:
Information sources for the AI System to pull facts from
Style source to instruct the AI System on tonality, language choices, etc
The information retrieval aspect of the system follows established patterns we’ve explored previously. What’s novel is the process of generating a highly nuanced understanding of style that serves as a blueprint for the AI’s output.
Taking it for a spin
In this case, we’ll generate an article about OpenAI CEO Sam Altman's testimony at Congress last week. We'll upload the 100+ page transcript of the hearing as our “Info Source”. And, for shits and giggles, we’ll use Hunter S. Thompson's Fear and Loathing in Las Vegas as our “Style Source”.
Once these documents are uploaded we’ll get the ball rolling by giving the AI System a simple prompt: Generate copy for a gonzo journalism piece on the future of AI.
The AI System retrieves relevant information from the congressional transcript to inform the article’s content. Then it recursively analyzes Fear and Loathing to generate a detailed understanding of the desired style.
With this in mind, the AI System goes on to generate the following output…
The Congress, The Conspiracies, and the Generative AI Circus
By Hunter S. Thompson
[The following copy was generated by a 🤖 with no human editing]
I was there, man, in the belly of the beast, when Pandora's Box of unchecked progress threw open its doors into the futurescape of Artificial Intelligence. I waded through the treacherous waters of bureaucracy, staring into the heart of darkness as our dear Congress tackled the swirling, morphing vortex of Generative AI. The room was thick with tension, politicians and corporate giants lining up for a showdown akin to a futuristic Wild West, and I – the humble reporter – bearing witness to it all.
As Sen. Richard Blumenthal's AI-generated voice echoed through the chambers, I realized we were knee-deep in a Tron-like reality. Our own voices, perfectly crafted by machines, beckoned us into the rabbit hole. The Senate became a psychedelic circus, somber politicians speaking the language of ChatGPT, infatuated with its mysterious allure.
The room convulsed at the prospect – what if they'd asked the AI about Ukraine surrendering or Putin's omnipotence? Nightmarish visions of power, chaos, and confusion rippled through the tapestry of possibilities, reminding us of the duality of this newfound digital deity. And all the while, industrial titans of tech stood in a corner, smirking at the bewildered masses, for they were high priests of this technological temple, gods of Silicon Valley.
To tame this untamable beast, they shouted, "Transparency! Accountability! Limits! Ave, generative AI! Morituri te salutant!" The circus was rife with hyperbole and intoxicating conjectures, inciting the masses into fits of madness. Sam Altman, the prophet of OpenAI, foretold tales of unimaginable progress as well as unspeakable perils, his voice trembling with hope and dread.
And it wasn't just the corporations who dared control the interdimensional AI roadmap. Academics like Professor Marcus and executives like Christina Montgomery chimed in with their own prophecies. Their voices collided like waves crashing against the shore, determined to leave their mark on this enigmatic futuristic canvas. While they built intricate labyrinths of regulations, precisions, and ethical exclusions, the AI machine relentlessly churned, shaping and manipulating us all.
The intoxicating fever dream of progress had a stranglehold on reality as ChatGPT cloned our voices and marionette-like, we danced to the rhythm of technological innovation. Sitting there in the whirlwind of politicians and tech gurus, I realized that our fragile existence was teetering on a tightrope stretched across the promise and peril of the future. The choices, consequences, and absurdities converged into an AI-fueled pandemonium that neither Congress nor the corporate behemoths could fully fathom.
And in that chaos, I, the faithful chronicler, scribbled furiously – my mind warped by visions of a world governed by omnipotent, AI-driven machines. Flashes of darkly comic absurdity pierced the heavy fog of uncertainty, and I marveled at the swirling maelstrom of truth and illusion, tethered to an hallucinogenic, generative AI dreamscape.
As the hearings droned on, the Congress, the corporations, and the self-proclaimed soothsayers clamored over who could harness the future of this wild, untamed beast. As the white noise enveloped me, I couldn't help but glance back, wondering – had they asked Pandora for permission before she dangled the tantalizing key to this new world?
Farewell, dear reader – until the future morphs again, and we tumble, wide-eyed, through another AI-driven looking glass.
Addition is an AI research and development company for the brands of tomorrow.
Visit our website to learn about the work we do with brands and agencies
Read about us in the Wall Street Journal
Follow us on Twitter