Human in the Loop AI

If we want to remain the “humans in the loop,” it’s crucial that we engage with technology on our own terms. Isn’t this where LLMs and Generative AI can truly shine? Surely, they are the perfect interface for us mere natural language speaking mortals - bridging complex algorithms and everyday users. Perhaps it’s all still very early and we’re just playing like puppies with something with which we have no real understanding. Perhaps the laggards aren't laggards at all, but frustrated futurists with much bigger brains than the rest of us. There may be some truth somewhere in there.

To those constructive sceptics of Generative AI, I say, it's understandable - especially given how much hype has surrounded it and the number of solutions labelled "AI" that often aren't AI at all. That said, real generative AI is built on significant advancements in machine learning, data science, and computer science. Its outputs are based on recognising patterns from large datasets rather than random chance. Taking time to understand the technology, and the way it's implemented, helps demystify, and shows that real AI can be a powerful tool, albeit one that requires the right approach and knowledge to use it effectively.

Fact is anyway with all that hype; the AI parade has moved on. We're past the point of AI being a novelty or buzzword; most buyers now simply expect AI to be part of the systems they use. The conversation has shifted from "Does this have AI?" to "How is AI being applied?" We're now firmly into the phase where practical applications and use cases matter more than the technology itself. It’s all about how AI can deliver real value and solve specific problems, rather than the technology being a selling point on its own.

Speaking of value, it's also fair to say that Gen-AI in the hands of untrained users can lead to mixed or even negative results. The quality of AI output is deeply influenced by the input - where prompting has now become a bit of a user skill in its own right. "Prompt engineering" is essential for training models effectively, but even after initial engineering, at point of use, users need some level of guidance or training to get the best results.

Simply integrating an API into a system without careful design or understanding, can and probably will lead to poor outcomes, confusion, possibly even unintended consequences. Proper training on how to interact with a model, by creating effective prompts, and interpreting responses is crucial for success. Without this, it's like giving a powerful tool to someone without explaining how to use it - it could lead to chaos rather than productivity.

One other point, this one on the importance of training data - which is so poorly misunderstood. While AI models are powerful, the quality and structure of the data they learn from is critical. Models need a well-organised, contextual framework to understand and process information accurately - this is where semantic indexing comes in. It’s essentially organising data in a way that helps the model grasp relationships, context, and meaning, much like the way we use a thesaurus to understand language, connections, and look for synonyms etc. When the data lacks this structure, the AI's output can easily veer off course. Helps explain why responses sometimes miss the mark. Proper data organisation is as crucial as the algorithms themselves.

There’s a very good reason Microsoft called theirs CoPilot and not AutoPilot. It really underscores the importance of human oversight, continuous refinement, and user education when deploying these systems. Careful training, thoughtful design, and quality data can unlock huge potential for value creation, while on the flip side, neglecting these considerations is more likely to result at best in frustration or worse a complete waste of time.

The difference between success, frustration, or worse often comes down to whether we approach AI as a partner or merely a product. As we move forward, let’s remember that if we want to remain the humans in the loop - real value lies in the synergy between human insight and machine intelligence.