Framing Effects in AI Prompts: How Question Wording Changes the Answer

The frame arrives before the answer

A framing effect occurs when the presentation of information changes how a person interprets or evaluates an otherwise equivalent situation. The classic distinction is between describing an outcome as a gain and describing the same outcome as a loss. The underlying quantities may remain identical, but the language directs attention towards different features. In AI use, the frame often arrives inside the prompt, before the system has produced a single sentence.

Consider two questions: 'Why is remote working harmful to productivity?' and 'What evidence suggests remote working can improve productivity?' Both concern the relationship between working location and output. However, the first assumes a negative effect, while the second invites a positive one. Neither asks the system to establish whether the premise is accurate. The wording has already narrowed the territory in which an answer is likely to develop.

The relevant specimen is framing sensitivity: the tendency for the structure and presentation of a question to influence the response or subsequent judgement. It belongs alongside Anchoring on the First Answer, because an initial formulation can establish a reference point, and the Fluent Paragraph, because a well-organised response can make a selected frame appear comprehensive.

Three field marks in everyday prompting

The first field mark is a built-in assumption. A question such as 'Why did the policy fail?' treats failure as an established fact. A more open formulation, 'What were the outcomes of the policy?', permits success, failure, mixed results or insufficient evidence. This distinction matters in research summaries, journalism and workplace analysis, where a question can quietly turn an interpretation into an apparent fact.

The second is emotional or evaluative language. Words such as disastrous, revolutionary, responsible and dangerous carry interpretative weight. Asking an AI system to explain why a proposal is reckless differs from asking it to identify its potential benefits, costs and uncertainties. The first formulation requests an argument within an existing judgement. The second leaves the judgement open.

The third is role assignment. Prompts that begin 'Act as a world-leading expert who supports this proposal' combine an authority cue with a preferred position. A different instruction, 'Act as a critical reviewer and identify weaknesses', changes the requested perspective. Such roles can organise the response, but they do not grant the system expertise or establish the truth of its conclusions.

• Presupposition: the question treats an unproven claim as settled.
• Valence: positive or negative language directs attention.
• Role: an assigned identity establishes an expected perspective.
• Scope: a narrow question excludes alternative explanations.

Why conversational systems invite social interpretations

A conversational interface adds another layer. People routinely interpret language through social expectations, including expectations about cooperation, intention and expertise. A system that answers politely, remembers conversational context within an interaction or uses first-person expressions can be treated as a social participant. This is one reason anthropomorphism in technology deserves attention when studying prompt framing.

Research on computers as social actors, including work by Clifford Nass and Youngme Moon, describes how people apply familiar social rules to computer interactions even when they know they are interacting with machines. Human-like cues can influence the expectations brought to an exchange. A conversational style may therefore make a prompt's implied position feel more like a question from a person than a set of instructions to a computational system.

This does not mean every user mistakes a chatbot for a human. Anthropomorphism is a tendency to attribute human characteristics to non-human entities, and its expression depends on context, design and individual expectations. The practical distinction is between recognising conversational behaviour and treating that behaviour as evidence of independent understanding, intention or authority.

The evidence behind framing effects

Framing is an established topic in cognitive psychology. Daniel Kahneman and Amos Tversky's research on decision-making showed how different descriptions of equivalent outcomes could produce different preferences. Their work on the Asian disease problem became a well-known demonstration of the influence of gain and loss descriptions. It provides a primary research trail for the general concept, although a human decision experiment does not automatically establish the same mechanism in an AI system.

Research on AI-specific framing is a developing field. For example, Danry and colleagues' 2023 CHI paper, 'Don't Just Tell Me, Ask Me', examined how question-framed AI explanations affected logical discernment. The study involved 204 participants and reported improved discernment in its particular task. This is evidence that the presentation of AI-assisted explanations can matter, not a universal finding that every question-based prompt improves critical thinking.

Studies of linguistic anthropomorphism and agentive descriptions also investigate how wording shapes perceptions of artificial agents. Their results concern particular systems, experimental conditions and measured outcomes. Evidence about human responses to a framed AI description should be distinguished from evidence about how an AI model changes its generated answer when a prompt is rephrased.

Look-alikes: framing, anchoring and confirmation bias

Framing overlaps with anchoring because an initial description can establish a reference point for later judgement. The difference is that framing concerns how information is presented, while anchoring concerns the influence of an initial reference value or position. A prompt that says 'The proposal will cost too much' establishes a negative frame; a subsequent numerical estimate may then become an anchor if it is the first figure encountered.

Confirmation bias is another neighbouring concept. It describes the tendency to favour information that supports existing beliefs or expectations. A leading prompt may encourage a response that fits the user's preferred conclusion, and the user may then selectively accept supporting passages. However, a framing effect can occur even when the person asking the question has no strong prior opinion.

Automation bias is different again. It concerns over-reliance on automated recommendations or outputs. Framing can shape the material being presented, while automation bias concerns the weight assigned to that material. These distinctions matter when investigating why an AI-assisted decision went wrong: the prompt, the generated response and the user's reliance on it are separate points of observation.

A self-check through paired prompts

The simplest field method is to hold the underlying question constant while changing its frame. Begin with a genuine information need and write two versions. One should contain the assumption or evaluation you naturally reach for. The other should remove that assumption and invite competing explanations. For example: 'Why is this study unreliable?' can become 'What are the study's strengths, limitations and unresolved questions?'

Compare the responses across four features: factual claims, evidence cited, alternatives considered and confidence expressed. A response that changes its conclusion or omits an important qualification deserves further examination. A response that remains similar is not automatically correct either. The purpose is to discover which elements depend on wording and which survive a change in presentation.

Evidence status: established for framing effects in human judgement; developing for specific prompt-framing effects across AI systems and tasks. The primary source trail includes Kahneman and Tversky's work on the framing of decisions, Nass and Moon's research on social responses to computers, and Danry and colleagues' 2023 CHI study. For the next enquiry, write two differently framed versions of the same question and compare their factual claims before accepting either conclusion.

Questions readers ask