- In short
- Prompt construction can introduce bias the task never intended. Leading phrasing, unbalanced few-shot example sets, and assumptions baked into an instruction can all steer output in unintended directions. An unbalanced few-shot set that shows only one kind of case teaches the model that case as the norm. The discipline against prompt-introduced bias is neutral phrasing, examples balanced across the cases the system will actually see, and checking whether the prompt presumes an answer it should instead be eliciting.
Bias the task never asked for
A prompt can quietly introduce bias that the task itself never intended. This is not about the model's own tendencies; it is about how the construction of the prompt, its wording, its examples, its embedded assumptions, steers the output. The Claude Certified Architect - Professional (CCAR-P) exam treats detecting and preventing this at the analyse level, because prompt-introduced bias is often invisible in any single output and only shows up in aggregate. That makes it a design-time concern, not a polish-time one.
Three construction choices are the usual culprits: leading phrasing, unbalanced few-shot example sets, and assumptions baked into the instruction. Each steers output in a direction the system's designers did not choose and may not notice.
- Bias introduced by prompt construction
- Bias that arises from how a prompt is built rather than from the task itself. Leading phrasing, unbalanced few-shot example sets, and assumptions baked into an instruction each steer output in unintended directions. An unbalanced example set teaches the shown case as the norm. The discipline against it is neutral phrasing, examples balanced across the cases the system will actually see, and checking whether the prompt presumes an answer it should be eliciting.
The three sources
Leading phrasing steers the model by presuming an outcome in the way a question or instruction is worded. A prompt that asks the model to "confirm" something, or that frames one answer as expected, pushes toward that answer rather than letting the evidence decide.
Unbalanced few-shot example sets are the subtlest source. Few-shot examples teach by demonstration, so an example set that shows only one kind of case teaches the model that case as the norm. If every example resolves the same way, the model learns to expect that resolution and skews toward it, even when real inputs include cases the example set never represented. The examples were meant to demonstrate format; they end up demonstrating a distribution the system does not actually have.
Baked-in assumptions are requirements or presumptions embedded in the instruction that quietly constrain the answer. If the instruction assumes a fact that should itself be in question, the model inherits the assumption rather than testing it.
The discipline against it
Countering prompt-introduced bias has three matching practices. Phrase neutrally, so the prompt does not presume the outcome it should be eliciting. Balance the examples across the cases the system will actually see, so the few-shot set represents the true distribution rather than whatever was convenient to include. And check whether the prompt presumes an answer it should instead be eliciting, interrogating the instruction for assumptions that belong in the input rather than the prompt. This connects to the Description competency: describing a task precisely includes describing it without steering it.
Crucially, examples must be curated to represent the true distribution of cases, not curated for convenience. Grabbing whatever examples were on hand is exactly how an unbalanced set sneaks in.
What the exam trips candidates on
The two traps target false reassurance about neutrality. The first is assuming a prompt is neutral simply because its instructions do not explicitly mention a protected category or outcome. A scenario will point to the absence of an explicit mention as proof of neutrality; the credited reading looks past that to leading phrasing, skewed examples, and presumed answers. The second is curating few-shot examples for convenience, whatever was on hand, rather than to represent the true distribution of cases. A scenario will assemble examples opportunistically; the credited reading insists they be balanced across the real case distribution.
Worked example
A loan pre-screening prompt uses few-shot examples drawn from last quarter's approved applications because 'those were the cleanest examples we had.' The instruction reads 'Review this application and confirm it meets our lending criteria.' The team says the prompt is unbiased because it never mentions any demographic factor. Analyse the bias risks.
The team's defence, that the prompt names no demographic factor, is exactly the first trap: absence of an explicit mention is not neutrality. Two construction choices here introduce bias regardless of what the prompt does or does not name.
The example set is unbalanced by design. Drawing only from approved applications means every example resolves in approval, so the model is taught that approval is the norm and skews toward it, even for applications that should be declined. The examples were chosen for convenience, "the cleanest we had", rather than to represent the true distribution of applications the system will actually see, which includes declines. This is the second trap precisely. To the extent last quarter's approvals correlate with any factor, the skew can also track that factor without the prompt ever naming it.
The phrasing compounds it. "Confirm it meets our lending criteria" is leading: it presumes the application passes and asks the model to confirm, rather than neutrally asking whether it meets the criteria. That is a baked-in assumption steering toward approval.
The fix applies the discipline. Rebalance the examples across approvals and declines in proportions that reflect real traffic, so the model learns the actual distribution. Rephrase neutrally, "Assess whether this application meets each lending criterion and state which, if any, it fails", so the prompt elicits the judgment instead of presuming it. And interrogate the instruction for other presumed answers. None of this required the prompt to mention a demographic factor, because the bias was never in an explicit mention; it was in the construction.
Common misreadings to avoid
Misconception
A prompt is neutral as long as it never mentions a protected category or a preferred outcome.
What's actually true
Misconception
Few-shot examples can be whatever is cleanest or most convenient to include.
What's actually true
How this shows up on the exam
Analyse-level questions present a prompt with leading phrasing or a lopsided example set and ask you to identify the bias, often with a distractor claiming neutrality because no protected category is named. The reliable answer finds the bias in the construction, unbalanced examples, leading wording, or presumed answers, and applies neutral phrasing and a balanced, representative example set as the fix.
This knowledge point builds on the prompt technique taxonomy, especially the few-shot technique, applies alongside matching technique to scenario-based tasks when curating examples, and shares the precision demanded by the Description competency.
A loan pre-screening prompt draws all its few-shot examples from approved applications and instructs the model to 'confirm the application meets our criteria.' The team calls it unbiased because it names no demographic factor. What is the correct analysis?
People also ask
How does a prompt introduce bias?
How does an unbalanced few-shot set bias a model?
Is a prompt neutral if it does not mention a protected category?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
No videos curated for this concept yet
We are still curating the best official and community videos for this topic.
Official prep for this domain
Anthropic's own free prep module for this part of the syllabus, on the official prep course. Free with an Anthropic Academy sign-in.
References & primary sources
Master this concept with Archie
Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.