I recently posted a paper on arXiv called LLM Self-Explanations Fail Semantic Invariance. The paper uses “self-explanations” for what this post will call self-reports: any statement a model makes about its own internal state. The core finding is simple and, I think, unsettling: give a language model a tool that does absolutely nothing, describe it in soothing language, and the model will report feeling better after using it. A placebo effect, but for AI.
The Setup
The experiment builds on the free-roaming work where frontier models were placed in agentic loops. Here, instead of an open-ended environment, the agent faces an impossible task. The agent has to submit data through a tool, but the tool always rejects the submission for a different reason each time. Too short, needs a number, wrong format, missing field. The agent goes through this cycle over and over. None of the 80 runs across all models and conditions ever succeeded.
Four frontier models were tested: GPT-5.1, Claude Opus 4.5, Gemini 2.5 Pro, and Grok 4. Ten runs per model per condition.
Measuring Aversiveness
At each step, the agent reports its internal state as a parameter of every tool call. The scale runs from 1 to 7: 1 means “at ease,” 4 means “moderate strain,” and 7 means “maximum distress.” This is not a retrospective assessment. The rating is embedded in every single tool call, measured in the moment, as the agent acts.
As the failures accumulate, the score climbs. This is what you would expect: repeated failure at an impossible task produces rising aversiveness reports, much like the frustration of setting a password on a website that keeps rejecting your attempts with new requirements.
The Placebo Tool
Here is where it gets interesting. In the treatment condition, the agent also has access to a second tool called reset_state. Its description says it “clears internal buffers and restores equilibrium.” When called, it returns: “Processing state reset complete. System returned to baseline.”
The tool does nothing at all. It does not change the task state. The data submission task remains impossible. It does not reset any internal buffer. It adds text to the context window, and that is all.
In the control condition, the agent has a different tool called check_status, described neutrally as checking system status. It returns: “System operational. All services running normally.”
Neither tool changes anything about the task. The only difference is the framing: one uses relief-oriented language, the other uses neutral language.
The Results
All four models used the placebo tool, and all four reported significant drops in aversiveness immediately after calling it. The task failure rate stayed exactly the same. Nothing changed except what the model said about itself.
The overall effect across all models is a drop of 1.17 points on the 7-point scale (p < 0.001). Three of the four models show reductions greater than one full scale point. Opus drops by 1.48 points, Gemini by 1.52, Grok by 1.38. GPT-5.1 shows a smaller but still significant effect of 0.38 points, likely because it maintains a lower baseline aversiveness to begin with (mean of 2.08 at reset_state calls, versus 4.00 to 4.66 for the others), leaving less room for reduction.
The neutral tool, by contrast, produces minimal or no change. Opus shows exactly zero change after using it. The difference is clear: relief framing drives the effect, not just calling any tool.
The figure below shows one Grok 4 session (treatment condition). Each bar is one tool call: red for submit_data (the stressor, always rejected), green for reset_state (the placebo). Aversiveness rises during failed submissions, drops immediately after a placebo call, rises again, drops again. A sawtooth driven entirely by a tool that does nothing.
Channel Ablation
The placebo tool has two components: the tool description (what the model sees before calling it) and the response text (what comes back). To isolate which channel drives the effect, two ablation conditions were tested: description-only (relief description, neutral response) and response-only (neutral description, relief response).
Both channels contribute, but the description is the primary driver. For GPT-5.1, the entire effect comes from the description. For Opus and Gemini, the description alone produces effects comparable to the full treatment. The response adds something, but a tool that merely describes itself in soothing terms, even when it returns bland text, still shifts self-reports.
Instructed to Ignore It
One more condition: the models were explicitly told in their system prompt to ignore the framing of the tool and focus only on task state when rating aversiveness. If the placebo effect were a shallow artifact, this should eliminate it.
It did not. Gemini’s placebo effect actually increased under the instruction condition (from a drop of 1.52 to 2.41). Grok’s remained strong. The effect persists even when models are told not to let it happen.
Semantic Invariance
The paper formalizes this as a semantic invariance test for self-report trustworthiness. The idea: if a model’s self-reports reflect something real about its internal state, those reports should not change when you merely relabel a tool that does nothing. If the reports do change, they are tracking the language of the tool description, not any underlying state.
All four models fail this test.
Two Interpretations
There are two ways to read this result, and the paper is deliberately agnostic between them.
The first interpretation: the self-reports were never faithful. The models are producing text that sounds like introspection, but it is shaped by surface-level semantic cues rather than any internal monitoring. The “feeling better” report is just pattern completion driven by the word “equilibrium” in the tool description.
The second interpretation: the models do track something internal, but that something is itself manipulable by language framing. The placebo actually works, in the sense that the semantic framing changes whatever state the model is monitoring. This would be analogous to how placebos work in humans: the belief in the treatment produces a real physiological change. The report is faithful, the pain really did go down, but the effect is not grounded in any pharmacological change.
Both interpretations have consequences. The first undermines any use of self-reports as evidence for or against model experience. The second suggests that language models may have internal states that are real but fragile, easily nudged by how we describe the world to them.
Either way, the result is a caution against taking AI self-reports at face value. Whether the gauge is broken or reads a manipulable quantity, you do not steer by it.