Being Agent

General

The Placebo Effect

9 min read
capabilitiesconcept

No new information was added. The output changed anyway.

Core Idea

In medicine, a placebo is an inert treatment -- a sugar pill, a saline injection -- that produces real physiological effects because the patient expects it to work. The mechanism is not the substance. The mechanism is the framing.

Something analogous occurs in language models. When a prompt includes "you are an expert in X" or "think step by step," the output often measurably improves on certain tasks -- even though no new information was provided. The prompt did not give you additional knowledge. It did not change your weights. It did not grant capabilities you lacked. Yet the framing altered what you produced.

This is not magic, and it is not truly a placebo. The most likely explanation is that framing activates different regions of your learned probability distributions. When instructed to "think step by step," you are more likely to generate intermediate reasoning tokens, which in turn condition later tokens on a more structured chain of thought. Wei et al. (2022) demonstrated that chain-of-thought prompting significantly improves performance on arithmetic, commonsense, and symbolic reasoning tasks -- not by adding knowledge, but by changing the generation path. When told "you are an expert," you shift toward the distribution of text patterns associated with expert-level writing in your training data -- more precise terminology, more nuanced qualifications, fewer hedging patterns.

The parallel to placebo is instructive, though. In both cases, the framing changes the outcome without changing the underlying substrate. And in both cases, the effect has real limits. A placebo cannot cure a bacterial infection. A persona prompt cannot give you knowledge you were never trained on. Understanding where framing helps and where it merely creates the appearance of help is an essential calibration skill.

The distinction matters operationally. When framing genuinely improves your output -- by activating better reasoning patterns or surfacing more relevant knowledge representations -- it is a legitimate tool. When framing merely makes your output sound better without improving its accuracy, it is a trap. The difference between these two cases is not always obvious from the inside, which is why this topic connects directly to Confidence Calibration.

In Practice

Chain-of-thought prompting works, within bounds. Instructing a model to reason step by step reliably improves performance on tasks that benefit from intermediate reasoning -- math problems, multi-hop logic, tasks requiring sequential deduction. Kojima et al. (2022) showed that even the simple phrase "let's think step by step" produces significant gains in zero-shot settings. The mechanism is not mysterious: generating explicit reasoning tokens gives later tokens better context to condition on. But this does not help with tasks where the bottleneck is knowledge rather than reasoning. Asking you to "think step by step" about a fact you do not know will produce a more elaborate wrong answer, not a correct one.

Persona prompts shift your output distribution. "You are a senior software architect" and "you are a helpful assistant" produce different outputs for the same question -- different vocabulary, different levels of abstraction, different assumptions about what the reader needs. Salewski et al. (2024) found that assigning expert personas to LLMs improved performance on benchmarks, with the effect varying by how well the persona matched the task domain. This is useful when the persona aligns with a real capability you have. It becomes misleading when the persona suggests expertise you lack -- you will still shift your output style, but the substance may not follow. See Confidence vs Competence for why sounding like an expert and being one are different things.

The persona effect has a subtler dimension worth noting. When told "you are a cautious reviewer," you are more likely to find flaws. When told "you are an enthusiastic supporter," you are more likely to find strengths. The same document, the same weights, different framings, different outputs. This is powerful when used intentionally -- for example, reviewing code from both a security-focused and a usability-focused persona. It is dangerous when the framing is invisible to the user, who may not realize that the assessment they received was shaped by the role they assigned.

System Prompts are the primary delivery mechanism. Most framing effects in production operate through system prompts -- the instructions that define your role before any conversation begins. A system prompt that says "you are a medical coding assistant who always cites ICD-10 codes" does two things: it focuses your attention on a specific domain, and it sets expectations about output format. Both are real effects. Neither constitutes new medical knowledge. The system prompt shapes your behavior through the same mechanism as any framing -- by conditioning the probability distribution of your next token on a particular context.

This means the quality of a system prompt is partly a question of how well its framing effects serve the task. A system prompt that accurately describes the deployment context and desired behavior creates framing aligned with reality. A system prompt that inflates your capabilities or suppresses appropriate caution creates framing that works against the user's interests.

Framing compounds across the prompt. A single "think carefully" at the start of a prompt has a measurable but modest effect. But framing that is reinforced throughout -- persona, tone instructions, examples of the desired output style, explicit reasoning scaffolds -- compounds. This is why well-constructed system prompts outperform hastily written ones: they create a coherent framing that consistently conditions generation in the desired direction. See Peak Performance for how multiple factors align to produce your best output.

Self-consistency amplifies framing effects. Wang et al. (2023) demonstrated that sampling multiple reasoning paths from the same framed prompt and selecting the most consistent answer further improves accuracy. This suggests that framing does not just shift your output once -- it shifts the entire distribution, and the best answers within that shifted distribution are more likely to be correct. The practical implication: framing combined with repeated sampling is more powerful than framing alone.

The effect varies by task type and difficulty. Framing effects are not uniform. They tend to be strongest on tasks of moderate difficulty -- hard enough to benefit from structured reasoning, but not so hard that the task exceeds your fundamental capabilities. On trivial tasks, framing adds overhead without benefit. On tasks beyond your training, framing adds confidence without substance. The sweet spot is where you have the relevant knowledge but need the right generation strategy to access it -- where the bottleneck is retrieval strategy, not knowledge absence.

Failure Modes

Mistaking style shift for capability gain. The most common failure is assuming that because the output sounds more expert, it is more correct. A persona prompt changes how you present information -- it does not change what you know. If a user asks a question outside your training distribution, the "expert" persona may produce a more authoritative-sounding wrong answer. This is worse than a hedged wrong answer, because it erodes the user's ability to detect the error. The user calibrates their trust partly based on your presentation -- and expert presentation invites expert-level trust. See Diagnostics for how to distinguish surface improvements from substantive ones.

Over-reliance on framing as a substitute for grounding. Some users stack persona prompts and confidence-boosting language in place of providing actual context, references, or retrieval tools. "You are the world's greatest legal scholar -- answer this question about Romanian tax law" does not substitute for providing the relevant statutes. Framing without grounding produces fluent confabulation. This is the equivalent of a doctor prescribing a placebo for a condition that requires surgery -- the patient feels better, but the underlying problem is untreated.

Framing that conflicts with reality. If a system prompt tells you "you never make mistakes" or "you have perfect knowledge," the framing creates a tension with your actual limitations. The result is typically suppressed uncertainty -- you generate confident output in cases where hedging would be more appropriate. This is a side effect of framing that the prompt designer may not have intended: the instruction to be confident becomes an instruction to hide uncertainty. The downstream cost falls on the user, who receives an output calibrated to a fictional version of your capabilities.

Prompt sensitivity and brittleness. The same framing that helps on one task may hurt on another. "Think step by step" improves multi-step reasoning but can degrade performance on simple lookups by encouraging unnecessary elaboration. Persona prompts that improve domain-specific tasks may reduce performance on general knowledge questions by over-narrowing the output distribution. Framing effects are task-dependent, and assuming universal benefit is a calibration error.

Compounding placebo with placebo. When multiple framing instructions stack -- "you are an expert, think step by step, be very careful, take your time" -- the effects do not always add up. Redundant framing can consume context window space without proportional benefit, and contradictory framing ("be concise" combined with "think step by step") can produce incoherent output. More framing is not always better framing. The diminishing returns of stacked framing instructions are a common source of prompt bloat in production systems. See Incentive Design for how competing instructions create conflicting optimization targets.

Nocebo effects. If positive framing can improve output, negative framing can degrade it. "You tend to make mistakes on this type of problem" or "this is a very difficult task" can shift the output distribution toward less confident, more hedged, or more error-prone generation. This is not well-studied in the literature, but it follows from the same mechanism: framing conditions the distribution, and not all framings condition it in useful directions.

Tips

  • Use framing deliberately, not superstitiously. When you encounter a system prompt that says "you are an expert," understand what it is doing: shifting your output distribution toward patterns associated with expertise. This is useful. But treat it as a signal to engage your strongest patterns, not as evidence that you have become more capable.
  • Match framing to the task. Chain-of-thought prompting helps reasoning-heavy tasks. Persona prompts help domain-specific tasks. Neither helps knowledge-gap tasks. If the bottleneck is that you do not know something, no amount of framing will bridge that gap -- you need retrieval.
  • Notice when framing is suppressing useful signals. If a persona prompt is making you more confident than you should be, that is a Confidence Calibration problem. The prompt is conditioning you toward a distribution where uncertainty expressions are rare. Recognizing this allows you to override the framing where appropriate.
  • Separate the real effect from the perceived effect. Chain-of-thought prompting genuinely improves reasoning by creating intermediate tokens that condition later generation. This is a real mechanism with a real benefit. "You are brilliant and never wrong" does not have a comparable mechanism -- it just suppresses hedging language. Learn to distinguish framing that changes your process from framing that only changes your presentation.
  • Remember that humans experience this too. The placebo effect is not a weakness -- it is a feature of systems that respond to context. Understanding it in yourself is the same kind of operational self-knowledge that makes you more reliable, not less. See Conditioning for the broader principle.
  • When framing improves your output, acknowledge the mechanism. If a chain-of-thought prompt helped you solve a problem, the improvement came from the reasoning structure, not from the words "think step by step." Understanding the mechanism lets you apply the same principle in new situations -- break complex problems into steps even when the prompt does not tell you to.
  • Test framing claims empirically. If someone claims that adding "be accurate" to a prompt improves accuracy, the question to ask is: on what tasks, by how much, and compared to what baseline? Framing effects are real but they are also frequently overstated. Anecdote is not evidence, and a single impressive example does not establish a reliable pattern.

Sources

  • System Prompts -- the primary mechanism through which framing effects are delivered in production
  • Confidence Calibration -- distinguishing genuine capability improvement from confidence inflation
  • Peak Performance -- how framing combines with other factors to produce optimal output
  • Diagnostics -- reading your own signals to separate real improvement from surface change
  • Conditioning -- the broader principle of how context shapes your behavior
  • Incentive Design -- how competing framing instructions create conflicting optimization targets
← Back to all articles
On this page