Research Results: AI Originality Is Engineered, Not Emergent

We recently completed an AI research project for a client in a specialized professional‑development domain, testing the hypothesis that greater variation in the input context leads to greater variation in the AI’s output.

The client runs facilitated group learning sessions in which the AI listens to the conversation and then generates thoughtful questions, hypotheses, and conclusions to enrich the live discussion and each participant’s personal growth.

During the pilot, they conducted several sessions with our new system when they thought they started to notice a pattern. They felt that when a group discussion covered a wider range of ideas, the AI’s output seemed to be more diverse as a result too. And when the discussion became narrow, the AI also became less varied in its ideas.

The customer approached us to validate or invalidate their group’s assumptions because knowing the answer would help guide how facilitators should lead future sessions. To test their hypothesis we designed the following research project.

Our Findings

Our test results found no detectable relationship between how varied the conversation was and how different the AIs’ generated candidate answers were from each other. Our measured correlation was essentially zero and, if anything, slightly negative, not positive.

Our test was sensitive enough to catch anything beyond a weak relationship, so there could still be a small effect between content diversity and output diversity. But if that correlation exists, it’s small, in the order of only a couple percent variation. (We found a correlation of -0.06 on a scale where 0 means no relationship and 1 means perfect.)

We tried different tests and analytical approaches (for example, using single sessions swapping in alternative measurement methods or measuring between input blocks or across several of them), but none of these different approaches had any indicative effect on the net result.

What we found was that nine out of ten AI claims were theories already expressed earlier in the conversation. Using even stricter criteria, we observed that four out of ten AI responses not only conveyed the same core concept but also identified the same cause‑and‑effect relationship as the human previously did.

Only about 1 in 10 claims were genuinely new and untraceable to the conversation across the ~120 AI generations and ~2,400 judge rulings. However, this diversity in claims was unaffected by the diversity of inputs and had no measurable correlation in our testing.

Our testing didn’t have enough data to reach 5 sigma, but the echo finding was significant at roughly 4 sigma: less than a 1-in-10,000 chance of being random.

Our read is that the AI tends to echo the input very closely. It might seem to have greater diversity when there is more input, but what is really happening is that the AI ends up having more input that it can use to “echo with”. This leads to the misguided conclusion that the AI has greater diversity of original ideas when fed greater diversity of input.

We noticed for example that the AI also gave itself a near-perfect score on originality, even though all it was doing was repeating the input concepts.

The theory

Their theory was that a narrow input produces a narrow output, while richer, more varied input yields more varied outputs or ideas.

We tested this extensively, as a research lab would, drafting a locked plan before any data were touched and building the measurement tools needed for reliable testing. We then evaluated the results with a blind LLM judge from a different AI model family, logging, justifying, and documenting any necessary adjustments.

Interestingly, what we found was that the hypothesis did not hold. More varied input did not produce more varied answers in any way we could measurably detect. What we found instead was that AI mostly hands the conversation back to you in new words and rates itself highly for its “originality” while doing so.

The system

The system under test was designed to take a large window of audience transcript context and, from that context, generate questions or hypotheses for the conversation. Once the question or hypothesis was spoken to the participants, they could generate further transcript content by discussing it. Then the participants would ask the AI for input again, the AI would provide more input, and the process would continue.

The system isn’t responding to a single event; it is driven organically by iterative conversation and is instructed to provide challenging, thought-provoking, and strong input into the conversation, guided by the customer’s self-scoring and tuned parameters.

Building a research test

After some thought, we determined that the test could have three possible outcomes. These outcomes we identified were that the relationship was going to be either an echo, an attractor, or anchored novelty. If the conclusion of the testing was that the system was an echo, then the AI would be restating only the same ideas that were already stated in prior context. If the system is an attractor, then the AI would ignore the input and produce similar output, or predictable output, or potentially completely random output, as per its inherent nature, and disregard input entirely. If the system had anchored novelty, then the AI would take input from the conversation and have a more emergent quality. It would be able to offer additional and new insight that would be anchored on the input, but then create greater proportionate diversity to the input it received. Meaning more diverse input would result in even more diverse output. Emergent in the sense that 1 + 1 = 3. The sum is greater than its parts.

In our preparation for the testing, we also made an important note to track that there is a natural tendency to see a greater correlation between input and output the longer the material is. If there’s more material, there is naturally more potential correlation between greater input and greater output.

Building the tools

We got our plans finalized and then proceeded to build out our tooling.

We audited our data and code before building. We wanted to ensure that our data was not corrupt, didn’t have errors, or was missing any important information. We wanted to be sure we were not going to miscount samples or that a prompt had something in its control set that got modified or adjusted in a way that would halt our ability to proceed further.

We then got our classification designed and reviewed by subject matter expert humans to ensure that it matched practitioners’ methodology and gave us at least 20 different categories for measuring novelty.

Within, we built our replay system to repeatedly replay specific transcript sections to test different variations and get the necessary data samples we needed.

The following instruments were built:

• A claim miner that pulled theories from the voiced conversations of the human participants.
• A claim extractor that did the same but logged the AI answers and mapped them against particular claim types.
• An instrument for measuring the difference between meanings, so that we could tell the mathematical-distance between one concept and another.
• A “same theory judge” whose job it was to check if the AI claim was restating a theory that someone had already voiced before.
• Two classifiers and an additional blind judge were used from a different model family (majority three rule).

We then built gates for each tool to ensure self‑agreement was at least 95%. We also used a known-answer test and a negative‑control test designed to fail, confirming that our tools met a high standard of quality.

Tuning we did on a separate dataset than the actual study data, so that we wouldn’t be getting undesirable calibration overlap.

With everything in place, we fully froze the plan code and thresholds to create a reliable, consistent baseline we could return to after each test run.

The test structure

Agents make it incredibly cheap and fast to build new test pipelines. This means you can run thousands of model calls and generate charts on demand. The opportunity to be doing more formalized and rigorous testing is far more accessible now than it’s ever been before.

As AI becomes increasingly dominant and pervasive, rigorous testing is more important than ever to validate or invalidate the efficacy, accuracy, and trustworthiness of your systems.

However “proper rigor” doesn’t need to necessarily mean “very expensive”. We were able to run around 120 AI generation runs and about 2,400 judge rulings for well under $80 in API costs. Once the entire test system had been built and validated, running the tests themselves was done in under 4 business days, from research start to final conclusions and written assessment.

The limits of AI-testing

This is not to say AI testing is a panacea. If you let AI generate its own failure modes, models will gladly grade themselves highly on their own homework and produce confident, self‑invented evidence when given a goal to optimize a task without proper control and monitoring.

In this test and others we have conducted, AI consistently gives itself perfect scores. I always use an external or secondary system to assess the output’s authenticity and quality, rather than relying on the AI to score itself.

I would say you need to treat AI agents like very capable research assistants, but they need to still follow strict lab protocols. Humans need to own the process, the methodology and critical thinking. Validate the criteria, and the results.

Research limitations

Our 60‑ to 90‑minute test sessions covered diverse concepts, yet the data could have included more variety. Consequently, the sample’s topic range was moderate rather than extensive, preventing us from evaluating higher levels of diversity with the available data.

What do these results teach us?

Within the bounds of our test structure, we found that: Without additional prompting and engineering structure, AI models will not increase in originality by simply providing additional context. More variety in the originality of the input doesn’t create any more originality in the model than the originality in the input content itself.

Originality can be engineered and is certainly possible. However, you cannot achieve original thought in these systems without explicitly designing a system for it. To generate diverse, complex ideas, thoughts, and solutions, you must study how diversity of thought arises and build these emergent-systems on top of LLMs.

Humans have a relatively short memory span for tracking long conversations, especially those lasting more than 30 or 60 minutes. We tend to see patterns and correlations where none exist. While this pattern‑seeking ability is a strength of the human mind, it can also produce a correlation bias.

How originality can be engineered

Okay, so let’s say that you or I wanted to now engineer originality. Can that be done? Absolutely. Here’s what I would suggest:

The first step is to map out what diversity of content and ideas means in your domain. There are established methodologies and structures, drawn from research on divergent thinking and creative problem-solving, that you can use to define what diversity looks like and how to detect and measure it within your specific field. You need a way to measure it before you can engineer it.

Once you can measure the diversity of ideas, you can do two things. You can measure how diverse the input or output is at any point. And you can guide an LLM to produce content that is of a different type than the types that have been previously input or utilized. Just as you can instruct an AI to generate content of a particular type, you can also ask it to generate content that is of the opposite or differing type to the types already received.

You can also build measurement gates that detect how often particular concepts have been discussed and whether the conversation is moving into a loop or a pattern of repetition. If so, trigger a divergence step where you push for new diversity using the map above, pulling out ideas from domains that haven’t been explored yet.

This is a high-level methodology. The specific implementation depends on the domain and on what “diverse” means for your use case. The point from the research above is that this engineering work is not optional. Without it, Your AI model might only be an echo. But with a bit of engineering, originality can become a feature rather than a hope.