As Nathan Jacobson wrote recently, we are very excited to announce the launch of our Disco Research Assistant. What is Disco Research Assistant for? Simply put, its main purpose is to help people learn about intelligent design (ID) through an AI-powered tool. As Science and Culture Today fans are well aware, we have numerous articles on our websites about ID, with more published every day, including everything from popular articles to technical scientific papers. Tens of thousands of these articles, as well as books and other material, have been “ingested” by a large language model (LLM) system that uses retrieval-augmented generation (RAG) to generate outputs in response to user questions. In short, the user can ask questions of the search agent to learn about ID and related subjects, and it will answer. Oh yes — it is free for anyone to use.
If you want to see it in action, use the search bar at the top of Science and Culture Today, or for a more dedicated experience, you can give it a try here: https://disco.discovery.org/
To get a feel for how it works, try asking basic questions like:
- “What is intelligent design?”
- “What does intelligent design say about evolution?”
- “What are the top three lines of evidence for intelligent design?”
One cool aspect of Disco Research Assistant is that you can even chat with it after your initial query — so if you get an answer and you want to dig deeper, then just keep asking questions.
Of course, you can also ask it deeper or more technical questions — if there’s an article in our database then the LLM will provide an answer. Here are a few I tried out today just to see what it produced:
- “Could the type 3 secretory system have served as an evolutionary precursor to the flagellum?”
- “Are there bilaterian animals in the precambrian?”
- “What’s the strongest argument against intelligent design being science?”
You’ll notice that these particular queries pertain to potential arguments against ID raised by critics. Here at Discovery Institute, and especially at Science and Culture Today, we frequently discuss and respond to critiques of intelligent design. What that means is that if there’s a criticism that we’ve addressed, then you can ask the Disco Research Assistant and it will give you an answer. But note that it’s programmed to give answers that reflect the contents it has ingested — so that means you’re going to get the ID-perspective on things.
Disco Research Assistant is impressively versatile in that it will acknowledge what critics say, and then give you the ID take. You’ll learn both sides of the argument, but ultimately you’ll hear what ID says in response. That’s OK because the purpose of this tool is to help people learn about ID — our tool was built under the assumption that this is why you are here: to learn about ID. But it tries to maintain a neutral tone, explaining the ID perspective and the perspective of critics. You’ll be hard pressed to find an AI tool out there that is better informed and more objective.
All AIs Have Limitations
Over the past 10 or so months I’ve worked closely with Winston Ewert, Nate Jacobson, and the developers of this AI tool, managing this project and helping to ensure that it’s giving answers that are accurate and reflect the ID perspective of the materials it’s been trained on. We are generally very happy with the outputs that it provides, with the important caveats that no AI-based tool is perfect and that sometimes AIs can make mistakes. Living, breathing, thinking humans are always encouraged to fact-check whatever an AI tells you. This is a limitation of all AIs, not just Disco Research Assistant.
The point is this: Discovery Institute is very pro-technology and we are eager to adopt new tech that emerges (precisely as we’ve done here). But it’s important to always keep in mind the limitations of technology. Technologies like AI are tools designed to assist the God-given creative abilities of humans. But AIs are not omniscient or omnipotent, and no matter what anyone tells you, they can’t replace humans.
Research Assistants Demonstrate Why AI Isn’t Creative
Earlier this year I helped develop an internal Discovery Institute policy on using AI. The basic gist of that policy explains that Discovery Institute is, as I said, very pro-technology and we believe AI is an extremely useful tool. However, AIs can never replace the genuine creativity of a human being. Those core aspects of our AI policy were repeatedly reaffirmed during the process of creating Disco Research Assistant.
Disco Research Assistant is a fantastic tool for learning about ID. Sometimes I even use it when doing my own writing, to help me dig up links to my own previous writings about particular topics. But working on this project has strongly emphasized to me another major limitation of AIs: they’re not truly creative.
Why do I say this? The outputs that Disco Research Assistant provides can be truly impressive and usually closely match the question you asked. But it’s not thinking for itself and it’s not generating anything new. Its answers simply reflect the materials it’s been trained on. To appreciate why this is the case, let’s briefly explore how LLMs work.
An LLM is simply predicting outputs based upon extremely complex matrices of numbers and vectors which represent specific words. Within this highly complex multidimensional grid of vectors, an LLM can “predict” what word should come next based upon the preceding words, or the query that the user inputted. Those vectors are not creative through some intelligent process — the AI doesn’t “know” the meaning of the words. Rather, the vectors are created by analyzing the training materials, and each word is analyzed to determine its likelihood of occurrence in the context. An LLM is just using numbers and vectors to predict what words should be outputted based upon materials that the AI been trained on (“ingested”).
When helping to create Disco Research Assistant, I was able to appreciate this fact in real time. That’s because if we were to add a new article to the AI’s training database, that could immediately change the output in response to a particular query. This is not because the AI got “smarter” or ‘learned” something new. It’s simply because it readjusted its numerically based word-prediction vectors based upon the new material it was “fed.” The process of “feeding” new articles into the AI and watching the output adjust immediately made me appreciate that the AI is not engaged in reasoning and it was not being creative — it’s just spitting back responses based on a complex algorithm that linguistically and numerically analyzed whatever it’s been fed.
“The Prompt”
There’s something else that I learned about the unintelligent nature of AIs while working on this project — and it’s related to something called the “prompt.” The prompt can mean different things in the world of AI, but in this context I’m talking about a back-end aspect of operator control that the front-end user never directly sees. By changing the prompt, the operator or controller of the AI can essentially create “guardrails” that instruct the AI to respond in particular ways in particular situations.
While developing the Disco Research Assistant, I was able to see how modifications to the prompt could instantly change its outputs. Again, this wasn’t because the AI got “smarter” or was able to “reason” in new ways. It’s simply because the content of the materials in the prompt were given a higher priority in the LLM’s predictive matrices, and thus had a statistically much higher ability to influence the words it would output. As impressive as the AI was, it wasn’t reasoning and it wasn’t thinking. It was just modulating its highly complex multidimensional matrices of vectors based upon instructions that I had given it in the prompt — in essence reprogramming what kinds of words it should output in response to queries. The training materials were one set of materials that governed its vectors. The material in the prompt was another. But the priority level given to materials in the prompt really helped me appreciate that it’s just outputting as it’s been told to do.
No AI is Perfectly Objective
This brings me to my last point — the fact that no AI is perfectly objective.
The capabilities of AIs are truly impressive. They seem to “know” a lot because they’ve been trained on vast databases of words, images, sound, video, and other input sources that humans have created. But as I’ve already said, they really don’t know anything — the outputs they give just represent words that have been used to train the predictive vectors in the AI’s complex system of matrices. And what I learned from helping oversee the creation of the Disco Research Agent is that you could easily influence the outputs by changing the training materials, especially the prompt.
But that’s not unique to Disco Research Assistant — it’s true of any LLM-based AI. Yet AIs are so powerful that we may begin to think of them as perfectly objective oracles that can any answer virtually any question we give it. But you are almost guaranteed that there’s a human programmer or team of programmers behind every AI that put guardrails on the AI to modify or manipulate its outputs in certain situations.
Here’s an example: Just as our AI was trained on materials that are pro-ID and was oriented to give the pro-ID perspective, I assure you about many AIs out there that they’ve been instructed to give you the pro-evolution view. Our AI will back up its claims just like any AI will when asked for sources. That doesn’t mean it’s wrong — you can literally trace our AI’s answers back to articles that will show you that what it’s saying is backed by the scientific literature. But the difference is that we’re disclosing what our AI was trained to do, while many other AIs will pretend they’re telling you what they claim “science says.”
AIs do what they’re trained to do. You need to know that next time you think that any AI is a perfectly objective oracle. Fact-check an AI, explore other perspectives, and see if what you’re being told holds up. Ours is no exception.