⚠️Human-in-the-loop over emphasis for our insecurity or a real need?
Human-in-the-loop (HITL) is used in many different contexts. In the context of machine decision-making, it is when humans are overseeing the results and are able to refine, adjust, or override the outcomes of an automated decision model for any reason.
The HITL concept has become more popular and has been discussed more since the popularity of the #GenAI and #LLMs algorithms and their related software applications.
I can understand. Again, let me emphasize that the so-called hallucination is not a defect of #LLMs but rather the by-product of their underlying transformer architecture. This hallucination caused a global panic at mass regarding the trustworthiness of the outcome of the decisions made by machines.
When we consider the context of #LLMs and use cases organizations wanted to use, it makes sense that the HITL is a big element in ensuring that no hallucination becomes the final decision's outcome.
💣However, I want to raise some questions and concerns here.
1️⃣ AI is not all LLMs — The #GenAI movement from big techs not only did not help the AI but also damaged the reputation and trust in the decisions made by machines. There are many other techniques in #AI that are 100% accurate and reliable and have nothing to do with #LLMs and therefore there is no hallucination issue (feature?) posed to the decision's outcome.
2️⃣ Humans are way too slow and smart — The other challenge I see with an overemphasis on Human-in-the-loop is that we are limiting ourselves to overseeing the outcomes. Imagine a machine that takes hundreds of thousands of decisions in a minute. I would like to hear from any of the HITL advocates how they imagine putting humans to validate and refine this high volume of decisions.
3️⃣ Automated validation — I believe what we generally think of refining, adjusting, and validating the outcomes of decisions made by machines is either based on human intuition or deep in the form of domain knowledge. That domain knowledge (in a symbolic form) must shape the “Intelligence Core” of automated decisions, which, in this case, there is no need for a human-in-the-loop for the outcomes of automated decisions. Also, the “Intelligence Core” should be used to automatically validate the decision's outcome, which, again, “automatic” means no human-in-the-loop.
4️⃣ The Bias – and specifically human biases such as confirmation bias, cognitive bias, self-interest, power-seeking interest, etc. are all nothing we are surprised by in today's life. I'm sure you have had experiences or know a person who was harmed based on the wrong decisions of some medical staffs (no offense to anyone, but that's the reality). These kinds of issues would not have happened and or could have been entirely avoided if we (humans) had Robot-in-the-loop (RITL), and by robot, I mean automated decisions based on pure domain knowledge.
Posted here.
Published September 23rd, 2024 at 07:30 am
