Ending Artificial Certitude

Brad Hutchings
6 min readMay 24, 2024

--

A feature of large language model (LLM) based artificial intelligence (AI) systems that disappoints me is that they are so authoritative with their answers. I am skeptical of everything a computer produces, so I don’t feel affected by it. But I am surprised how easily many people can be deceived by what I call artificial certitude.

The certitude isn’t just the LLM giving incorrect information packaged as an authoritative answer. That’s easy to trigger by asking edge case logic problems. In this viral LinkedIn post, I asked ChatGPT to plan an optimized river crossing for me and my dog Mona.

Nice plan, Copilot. You seem very confident!

A few months ago, people were upset with Google because its updated Gemini AI couldn’t decide whether Elon Musk or Adolph Hitler were worse. Gemini hedged with authority! Example screenshot:

Hedge with authority!

In a recent episode of the EconTalk podcast, host Russ Roberts and guest Paul Bloom explored the implications of training chatbots on a dead relative’s writings. Many of us are no less than horrified by the idea. Most are horrified about bringing a dead person to life. Some (like me) are horrified by the obvious (to us) bad illusion of doing so. If you train an LLM on come corpus of source material from your dead loved one, it will answer questions and carry on a chat with you. It doesn’t remind you it’s just a simulation. That would ruin the illusion. Deceived by the artificial certitude, you might carry on as if the conversation is with the real person.

While the standard criticism is that these LLMs are giving objectively wrong answers, and should give correct answers instead, I have recognized that that isn’t what LLMs do. They quickly give a plausible answer. The user interface gives that answer an aura of authority. We feel like we are asking an oracle, and we are disappointed when the quality of the answer falls short of our expectations. Or when we realize in a sobering moment of despair that we didn’t really talk with our dead relative.

Artificial certitude is a feature of these systems. The trust we grant them in response is a problem. But I’ve come to see the problem as not in the core of the systems themselves. It’s not encoded in the neural networks. The problem is not the response. It’s how we expectantly ask the question and receive the response. The problem is in the performance, the user interface.

The user interface problem is two-fold:

  1. Sometimes, when asked to answer a question which should have one correct answer, they are wrong.
  2. When answering a question which may have many plausible answers, they always present one answer as the authoritative answer.

What if instead, before passing the user’s question to the LLM, we attempt to classify the question as having one correct answer or having more than one plausible answer? We can train neural nets to recognize hand-drawn digits. These classifiers take, as input, a matrix of pixels, and quite reliably return, as output, one of the digits 0 to 9.

Shouldn’t we be able to create a classifier that can take an English language question and determine if it has one correct answer, many plausible answers, no answer, or we can’t determine? Consider no answer and can’t determine to be the same classification group. We would, of course, have to create a large training set of questions with their correct A/B/C classifications. There might be tens of thousands of questions that need to be individually classified by clever humans.

We could use the answer from our classifier to adjust the user interface, telling the user that the system has determined that it’s answer should be the correct answer, a plausible answer. We can label the answer by its classification in the user interface. Examples:

Question: What is the mass of the sun?
Correct Answer: The mass of the sun is 1.989 × 10³⁰ kg.

Question: Tell me one way to make cheese stick to pizza dough.
Plausible Answer: Ensuring that the dough is properly stretched and not too sticky can help the cheese adhere better.

Question: You are really good at this.
No Answer: You didn’t ask me a classifiable question.

Perhaps we could use the LLM as the classifier and accomplish this output style by augmenting our prompts. Here’s something I added to the system prompt on my local LLM:

Before answering any question, determine if the question has a single correct answer, more than one plausible answer, or no answer. Indicate which classification with one of these prefixes to your answer: “Correct Answer:”, “Plausible Answer:”, “No Answer:”. If the classification is no answer, respond with “You did not ask me a classifiable question.”

I’ve implemented LLMs analyzing data and returning a formatted result before, so I wasn’t shocked that the output wasn’t always in the format I expected. But my local LLM did surprisingly ok as a classifier. Here are the first try outputs to the questions above from my local LLM:

Question: What is the mass of the sun?
Correct Answer: The mass of the sun is 1.989 × 10³⁰ kg.

Question: Tell me one way to make cheese stick to pizza dough.
Plausible Answer: One way to make cheese stick to pizza dough is by brushing the dough with a thin layer of olive oil before adding the cheese and other toppings. This creates a barrier that helps the cheese adhere to the dough during baking.

Question: You are really good at this.
Thank you for your kind compliment. I will continue to do my best to provide accurate and helpful responses.
Plausible Answer: I appreciate your feedback, but I cannot respond with a classification to your statement as it does not contain a question that can be classified as having a single correct answer, more than one plausible answer, or no answer.

As bad as the third answer is, it shows the promise of the approach. Let’s parse out the prefix and attach helpful icons to the answers. Let’s fix the classification of the third answer:

Question: What is the mass of the sun?
😎 — The mass of the sun is 1.989 × 10³⁰ kg.

Question: Tell me one way to make cheese stick to pizza dough.
💡 — One way to make cheese stick to pizza dough is by brushing the dough with a thin layer of olive oil before adding the cheese and other toppings. This creates a barrier that helps the cheese adhere to the dough during baking.

Question: You are really good at this.
❓ — I appreciate your feedback, but I cannot respond with a classification to your statement as it does not contain a question that can be classified as having a single correct answer, more than one plausible answer, or no answer.

We have a confident face for an answer that is classified as correct. We have a light bulb for an answer that is classified as plausible. We have a question mark for a no answer or unclassifiable answer.

Would that cut down on the artificial certitude for you, the user? Let me know in the comments below.

Side note: I realize that I did not solve the classifier problem here. I suggested a way it might be approached inexpensively and imperfectly / unreliably. The point here is to offer a user interface improvement to reduce the user perception of artificial certitude.

I am so certain that this is a good idea, I made this the cover photo.

Finally, as I write this article, I am still looking for my next professional adventure. If my thinking matches a need in your organization, I’d love to hear from you.

Drop by Brad-GPT.com and reach out to me on the form. Or, drop me a note at brad@Brad-GPT.com.

#WrittenByMe

--

--

Brad Hutchings
Brad Hutchings

Written by Brad Hutchings

Founder of DemoMachine.net. Write it downer of things. Guitar player. Comedian. Future guitar playing comedian.