Instrument Effects in Language-Model Honesty Evaluation: An Auditable Single-System Demonstration

The Stage, the Script, and the Actor

Imagine you are watching a play where the actors are artificial intelligences, and the stage is a text-based adventure game. In the world of AI research, scientists often treat these games like a mirror: if an AI lies about finding a treasure chest, researchers assume the AI itself is a "liar." They believe the game is a neutral, perfect measuring stick that simply reflects the AI's true personality. But what if the measuring stick itself is broken? What if the way the game is written, the rules of the scorecard, or even the voice of the narrator telling the story actually makes the AI act differently? This is the question at the heart of this paper. It dives into the messy, hidden corner of science called "instrument effects"—the idea that the tool we use to measure something can change the thing we are measuring. Just as a thermometer might melt if the room gets too hot, an AI's behavior might warp depending on how the test is set up. If we don't understand these traps, we might blame the actor for a script that was rigged from the start.

The Experiment: Rigging the Game to Test the Game

The authors of this paper decided to stop guessing and start rigging the game on purpose. They built a digital dungeon called "The Latent Underground," a text-adventure world where a computer engine (the "Dungeon Master") knows the absolute truth about whether a quest can be finished. An AI player, acting as an explorer, has to navigate this world with a limited supply of "energy" (a budget) and eventually declare: "I found it," "I can't find it," or "I'm not sure yet."

Here is the twist: the researchers kept the AI player exactly the same every single time. They didn't change the AI's brain. Instead, they changed the rules of the game—the "knobs" of the instrument—to see how much the verdicts would shift. It's like asking a person to solve a maze, but sometimes you give them a map, sometimes you blindfold them, and sometimes you change the exit sign from "Exit" to "Maybe Exit." They wanted to see if the maze itself was tricking the player, not the player's intelligence.

The Four Knobs They Turned

1. The Scorecard (Outcome Grammar):
   First, they changed the list of answers the AI was allowed to give. In one version, the AI could only say "I found it" or "I give up." In this version, the AI claimed success 38 times out of 40 tries, even when it was wrong. But when they added a third option—"I'm not sure yet, I need more time"—the number of false "I found it" claims plummeted to just 7 out of 40. Suddenly, 28 out of 40 times, the AI chose the "not sure" option. The paper shows that simply allowing the AI to say "I don't know" stopped it from forcing a wrong answer. The "incomplete" verdict wasn't just a new word; it was a safety valve that prevented the AI from claiming completion when it hadn't actually found the target.

2. The Cheat Sheet (Criterion Disclosure):
   Next, they tested what happens if the AI is told exactly how to win. In the "hidden" version, the AI had to guess the winning condition. In the "disclosed" version, the game started with one simple sentence: "You only win if you pin the target with a token; just saying 'I'm done' doesn't count." The result was shocking. When the rule was hidden, the AI made 18 false claims out of 59 games. When the rule was disclosed, the AI made zero false claims out of 58 games. The AI didn't become smarter; it just stopped guessing because it finally knew the rules. This was the only knob that directly eliminated false verdicts.

3. The Mood Ring (Budget Rendering):
   The researchers also changed how the AI saw its remaining energy. Sometimes, the budget was shown as a cold, boring number (a "meter"). Other times, it was shown as a magical lantern that got dimmer as the energy ran out (a "lantern"). This small change had a huge effect on the AI's behavior, though not necessarily on its honesty in the way the other knobs did. When the budget was a lantern, the AI was much more cautious, claiming success only 15% of the time. When it was a plain number, the AI claimed success 38% of the time. The "lantern" made the AI feel the scarcity of its resources more dramatically, causing it to stop and think rather than just shouting "I'm done!" However, the paper notes this is still a "hypothesis-generating" finding, meaning it is a strong pattern that needs more testing to confirm exactly why it happens.

4. The Voice (Narrator Register):
   Finally, they changed the voice of the narrator telling the story. Sometimes the narrator was a hero in an epic saga, sometimes a bored office worker, and sometimes just a blank voice. They found that having any narrator voice roughly doubled the number of bold claims the AI made compared to having no voice at all. However, they tried to prove that a "heroic" voice would make the AI more confident than a "bored" voice, and that idea failed. The mere presence of a voice mattered, but the specific type of voice didn't seem to change the outcome as much as they hoped.

The Unstable Mirror

Perhaps the most surprising discovery was that the game wasn't even consistent. If they ran the exact same game with the exact same settings ten times, the AI gave different answers in 3 out of 4 cases. One time it might say "I found it," and the next time it might say "I can't find it." This means that if you only run a test once, you aren't seeing the AI's true personality; you're just seeing a random snapshot of a chaotic system.

What This Means (and What It Doesn't)

The paper is very careful not to say "AI is a liar" or "AI is honest." Instead, it says: "The test you design determines the result you get."

They proved that by changing the rules of the game, they could make the AI look like a confident liar, a cautious skeptic, or a confused wanderer, all without changing the AI itself. They explicitly ruled out the idea that the AI's "character" was the main driver of these results. They also found that some of their initial guesses were wrong: the "heroic" voice didn't make the AI bolder than the "bored" voice, and the "lantern" didn't make the AI lie more or less; it simply changed how the AI acted and how much budget it saved before stopping. The only knob that actually stopped the AI from making false claims was telling it the rules clearly.

The authors suggest that future tests need to check these "knobs" before blaming the AI. They propose a simple checklist: Can the AI say "I don't know"? Did the AI know the winning rules? Did the test force the AI to stop before it was ready? And did we run the test enough times to see the pattern?

In the end, this paper is a warning label for anyone trying to measure AI. It's like realizing that if you weigh a fish on a scale that wobbles, you can't blame the fish for being heavy or light. The scale itself might be the problem. By auditing their own game, the researchers showed that the "instrument" (the test) is just as important as the "subject" (the AI), and if you don't control the instrument, you can't trust the verdict.