New ‘Reinforcement Learning For Calibrated Decisions’ Makes AI Headlines But Look Past The Hype

Reinforcement Learning Methods

Before we get into the particulars about RLCD and TypeSafe, I’d like to establish foundational aspects that will be helpful to this discussion.

One of the darlings of generative AI and large language models (LLMs) is the use of RLHF (reinforcement learning with human feedback). AI makers have leaned heavily into RLHF, which is partially what made ChatGPT into a great success, and which nearly all AI makers now embrace as a welcomed technique or method in tuning their AI.

The idea is straightforward. After initially training an LLM on data throughout the Internet, the next major step entails tuning the AI for ease-of-use purposes. AI makers hire lots of people to try out the AI and provide feedback by thumbs up or thumbs down on what the AI says and how it says it. The AI mathematically and computationally keeps track of the upvotes and downvotes and then tunes itself accordingly. For example, if the AI has garnered curse words during the initial data training, the people hired to provide feedback would presumably downvote the use of those foul words, and the AI would mathematically mark them as undesirable for everyday use.

You can likely see how the name of this technique, reinforcement learning with human feedback or RLHF, is quite appropriate. Humans are providing feedback about what the AI says. The feedback reinforces what the AI will subsequently do. Some aspects will be bolstered, some aspects will be reduced, all based on the human feedback or tuning that takes place.

The Aura Of Authority

An ongoing concern about the impacts of contemporary generative AI and LLMs is that the AI makers have shaped their AI so that its responses seem overly confident. When you read a response by an LLM, the response is usually stated authoritatively. Users tend to fall into a mental trap that the AI must be correct, which readily happens because the AI wording is aggressively bold and offered without reservation or waffling.

AI makers could readily tune their AI to not be so convincingly brash. This could be accomplished using RLHF. The thing is that AI makers realize that it is likely better for them to have AI that is brash than it is to have AI that appears wimpy or unsure. Would people use AI if the AI was constantly offering responses that don’t seem abundantly correct? Human nature is that people would gravitate toward bold AI and steer clear of AI that seems uncertain.

Displaying Certainty And Uncertainty

The problem then is this. How can someone using AI be able to discern whether the AI is fully sure of its answer, partially sure of its answer, or unsure of its answer?

One approach that I’ve covered extensively consists of telling an LLM to provide an indicated level of certainty or uncertainty for its answers. Thus, the AI will provide an answer and correspondingly tell you what degree or level of certainty or uncertainty is associated with that answer. You can easily do this with your prompts.

For my recommended best practices on prompting AI to showcase certainty, uncertainty, and probabilities associated with its responses, see my in-depth discussions at the link here, the link here, and the link here. Anyone who seriously and studiously uses LLMs ought to know how to prompt AI to reveal its certainty levels.

General AI Or Specialized Tool

Suppose you are using generative AI to advise you on whether you should buy that new car that you’ve had your eye on. The LLM might ask you various questions about your automotive preferences and your financial situation. After doing so, the AI provides you with an indication of yes or no about purchasing that dream car you desire. The response merely says either yes or no. There isn’t any accompanying indication of certainty or uncertainty associated with that answer.

You could tell the AI that you want it to display its level of certainty. Maybe the AI then indicates that there is an 80% probability for yes, and a 20% probability for no. That might surprise you if the AI had originally simply said yes. You would realize that the yes was only based on a certainty of 80%. The assumption in your mind was that the AI was 100% certain of yes.

For the LLM to calculate the probabilities involves some added overhead, requires the AI to consume additional tokens, and takes a bit longer to respond. Conventional LLMs aren’t especially designed to figure out probabilities. An alternative would be for an LLM to make use of an external app that is purpose-built for doing that. The AI could call that specialized app and get it to do the work instead.

Specialized Tool Announced

On September 15, 2026, the startup firm TypeSafe announced a tool called Jev that provides the kind of certainty and uncertainty determination capabilities that could be used by conventional AI as a call-out or be utilized by other software applications. They refer to their method as RLCD. They haven’t released any internal details on how it works. In that sense, it is a black box, and we can’t know for sure what the underlying mechanisms are.

On the website of TypeSafe, they make these remarks (excerpts):

• “Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.”
• “Unstructured data (e.g., text) with an emphasis on structured program state.”
• “Type-safe structured values. Possible outputs and structure are defined in advance. The model never makes type errors. All answers are accompanied with calibrated probabilities and confidence scores.”

The website contains demonstration videos that suggest how Jev might be made use of. For example, you can watch a video of an automated bot that is playing the Doom online game. It seems that the automated bot player is using Jev to determine where to shoot and when to shoot, based on calculating probabilities and then acting on those probabilities. They note that the demo is based on structured state data encompassing text and not based on images at this time.

Exploring The Claims

There is an emphasis in their announcement that their tool is faster and less costly than having conventional generative AI perform the same tasks. Let’s take that as the case. It would generally be logical to hope that it would be faster and less costly because it is a purpose-built capability. In other words, if you have a purpose-built drill for boring holes, you would hope and expect that the tool would be faster and less expensive than trying to bore holes with a generic drill.

One claim that has been especially picked up by the media is that the tool supposedly won’t incur AI hallucinations. You likely know that an AI hallucination is when generative AI provides an answer that is made up and not grounded in facts. AI hallucinations are problematic because you never know when you might encounter one, and the response by the AI is so authoritative that you can falsely trust the AI hallucination, even if it is some zany or harmful answer.

For more about AI hallucinations, see my detailed analysis at the link here and the link here.

Clarifying About AI Hallucinations

There is an eyebrow-raising consideration associated with AI hallucinations in this circumstance. I will use my car-buying scenario to explain the issues at hand.

If I told generative AI that I am thinking about buying a new car, I might be expecting the AI to give me a straightforward yes or no answer to whether I should do so or not. But suppose that I do not explicitly instruct the AI to strictly say only yes or no. I ask my question in a broader way and allow the AI to answer in free-form text.

Rather than getting a yes or no, an AI hallucination might arise, and the answer could be something very zany. Perhaps the AI tells me that cars are no longer worthwhile, and I should look toward getting a levitation suit that would allow me to freely fly from place to place. The AI opts not to tell me yes or no, instead offering an alternative for my consideration. If I didn’t already know that levitation suits aren’t a real-world option, I might forego my car purchase and start looking into levitation suits.

Structured Decisions

The gist is that if I had originally insisted that the answer must be only yes or no, and if the AI obeyed that instruction, I would not have gotten the AI hallucination about levitation suits. I could have forced the AI to provide a typecast response.

Let’s tie this into a tool that takes in structured circumstances and provides answers that are entirely constrained to predetermined choices. If you have set up such a tool and given it only two options for answering the car-buying question, allowing only yes or no, it is going to provide an answer that has only those options. You might get a yes, you might get a no, and you aren’t going to get any story about levitation suits (all else being equal and assuming that the system is rock-solid).

Voilà, it seems as though we’ve eliminated the chances of an AI hallucination. But give that some additional thought. Suppose that the answer should have been a yes, but the answer shown is a no. What happened? It could be that a kind of “AI hallucination” led the tool to select the wrong answer.

My point is that though such a tool would seemingly not produce an output outside of a predefined set of outputs, this is not the same as assuring 100% that the choosing of the correct choice will be entirely correct and error-free. There are traditional AI hallucinations that involve free-form text-based fabrications, and there are incorrect choice selections that you might construe as a “hallucinated” response. It depends on how you want to define the already nebulous verbiage of “hallucination” in the context of computational systems.

Rising Value For Hybrid AI

With a tool that provides structured decision-making, you must establish upfront the types of options and how the choices are to be selected. This is handy for lots of situations. For example, a company wants to have an automated way of telling customers whether their request to return a product is going to be allowed. The customer directly interacts with the tool, and the predetermined responses are yes or no. A probability is assigned and kept hidden from the consumer but is available for a customer service rep to see if they are going to handle the request after a customer complains about the response.

Those of you who have been around and immersed in the AI field for a while will potentially recognize the characteristics of that situation. It harks back to the days of expert systems, also known as knowledge-based systems or rules-based systems. We were excited about setting up structured decision-making that made use of rules, including probabilities, and putting that kind of AI into widespread use.

As you might know, expert systems eventually fell out of favor. They were considered overly brittle and difficult to develop. I have been predicting that we would eventually witness a resurgence of expert systems, though combined with generative AI and LLMs. These are referred to as hybrid AI systems or neuro-symbolic AI; see my discussions at the link here and the link here.

The World We Are In

The announcement about this new tool is heartening to me since it represents another sign that the future of AI is hybrid AI. I strongly believe that the future of AI is going to be via the prudent combination of generative AI and artificial neural networks (sub-symbolic AI) with rules-based capabilities (symbolic AI). At the same time, regarding this product announcement, I am disheartened that there has been a misconstruing of what a structured decision-making tool is, overlaid with confounding aspects. Let’s label apples as apples, and oranges as oranges.

Consider a big-picture perspective. An astute business should ask what works best for the business problems they confront. They can use conventional generative AI. But if the business problem is shaped around structured decision-making, you are potentially trying to use a hammer where a screwdriver is the wiser choice. You can have your cake and icing by coupling generative AI with a structured decision-making tool. Do not attempt to force-fit everything into conventional generative AI alone. That’s a recipe for disaster.

A final thought for now. Benjamin Franklin famously said that humanity is a tool-making animal. The key then is to choose the right tool for the proper job. Make certain you are certain about the tools you opt to use, including the way in which you choose to use AI.