Understanding people is the missing context in AI - and why human input is needed for AI inference, not just training

It's established that human input for training models is critical.

We rightly spend huge amounts of time - and money - on human input and expertise in AI:

• RLHF for training models

• Human feedback for fine-tuning

• Humans-in-the-loop for evals

• Custom workflows designed by domain experts

All of this makes sense. Human input is foundational to building good models.

But then something strange happens.

Once a model is trained, human input suddenly stops.

When it comes to inference, we expect AI models and agents to generate content, make decisions, personalise experiences, and influence real people - without ever speaking to them to get their thoughts, feedback and insights.

We've optimised for model intelligence but ignored real-time audience understanding and insight.

Relevance is The Next Frontier

When AI can generate content in seconds, what matters is not speed, volume or even cost. What matters most is relevance.

For AI agents to produce relevant and targeted outputs - ads, recommendations, products, strategies - they need to understand people. And the best way to understand people is to speak to them.

Scraping the web, analysing historical data, or summarising old research can only take you so far. That data is static, generic, and disconnected from the context an AI is working on right now.

What's missing from the AI stack isn't more data that already exists.

It's real-time human context right at the point of inference.

Understanding People is the Missing Context in AI

As we enter the era of Agents and context engineering, AI needs the ability to instantly discover what people think, understand why they think it, and act on those insights immediately.

Instead of guessing what an audience might care about, AI agents should be able to ask - and get answers from - real people, in real time.

To create outputs that actually resonate, an agent needs context about the exact thing it's working on - right now:

• What does this specific audience care about?

• What language resonates with them?

• What problems are they actually trying to solve?

• What would make them stop and pay attention?

That's the missing layer in the AI stack.

Real Human Input for AI Inference

And that is what we're building: the infrastructure for how AI models and agents can interact and understand real people.

The challenge? 

Traditional research methods weren't built for the AI era - they're too slow, static and expensive.

Other alternatives have popped up, including AI personas and simulated audiences. These solutions are fast and cheap, but they actually do the opposite of human input. They replace the need to speak to or involve real people - instead scraping data from the web, old survey data and research reports.

Although it might not seem like it, these approaches actually move even further away from the need to hear directly from real people.

We thought about this deeply and designed a third way: one that fuses the power of AI with direct human input.

1:1 Digital Twins - Each Trained and Validated by Real People

Our approach is to leverage AI to amplify and augment, not replace, real people.

We built a consumer application, Twineo, that lets anyone create and train a Digital Twin - an AI version of themselves that answers questions and represents their views.

Over 20,000 real people have now trained a Twin.

Key to our approach is that real people are always and directly in-the-loop, continuously training, updating, and validating their Twin.

It's real people, amplified by AI.

As we like to say: if you want to follow up, we can get the actual person behind any Twin on a Zoom call.

What it Means For AI Inference

Via API or MCP, OriginalVoices then gives AI models and agents access to this network of Digital Twins right at the point of inference to:

• Describe any audience in natural language

• Ask about any topic or idea

• Get instant answers

• Use those insights to give your AI the human context it needs to make better decisions, validate ideas, and generate relevant, audience-aligned outputs

https://www.loom.com/share/0b8502982d3b4a52b6681f55180003ba 

Instead of generating generic ad copy or marketing creative, or instead of leaving your product descriptions or app store copy to AI guesswork, an AI agent can instantly learn what truly matters to a specific audience - and use those insights to directly shape messaging, content and decision-making that's relevant, targeted, and grounded in real human understanding.

AI That Interacts and Understands Real People

Human input is critical for AI.

And it should be critical for AI throughout the stack - not just for training but for inference too.

Our vision is to build the infrastructure for human input to move up the AI stack so that data and insight is available at the exact moment it's needed.

We're building the human input infrastructure layer for AI inference.

We call it the human intelligence layer for AI.

Would love to hear your thoughts - comment below or get in touch directly.