03 · What You Need to Know
Why Contextual Meaning Is Difficult for Generative AI to Interpret
Understanding Words Is Not the Same as Understanding Their Use
Language communicates more than the literal meanings of individual words. People use expressions to acknowledge others, signal disagreement, maintain relationships, soften criticism, and communicate experiences that may be difficult to state directly.
Pragmatics examines how language is used and interpreted in context. A statement's meaning may depend on who speaks, to whom, under what circumstances, and for what communicative purpose.
Consider the sentence, "That's just wonderful."
Spoken after receiving good news, it may communicate genuine satisfaction. Spoken after another unexpected technical failure, it may express sarcasm.
The words remain identical, but the communicative meaning changes.
Generative AI can infer possible meanings from linguistic and contextual patterns. Yet the model may not have access to the tone, shared history, or situational knowledge that participants and interviewers used to understand the exchange.
What Types of Context Matter in Qualitative Research?
Context is not a single variable that can be added to a prompt. Different dimensions may shape the interpretation of participant accounts.
| Contextual dimension |
Why it matters |
Possible AI limitation |
| Linguistic context |
Earlier and later statements clarify what an expression refers to. |
Isolated excerpts may be interpreted without their conversational meaning. |
| Interactional context |
Questions, interruptions, and interviewer responses influence the exchange. |
The model may overlook how the interview interaction shaped the answer. |
| Cultural context |
Local expressions and social conventions influence how meaning is communicated. |
Familiar English interpretations may be imposed on culturally specific language. |
| Institutional context |
Rules, roles, and power relationships affect what participants can comfortably express. |
Statements may be interpreted without considering relevant institutional circumstances. |
| Temporal context |
Experiences and attitudes may change over time. |
Statements from different periods may be treated as simultaneous or contradictory. |
| Paralinguistic context |
Tone, pauses, laughter, and emphasis can qualify spoken meaning. |
Conventional transcripts may omit cues present in the original recording. |
These dimensions interact. A phrase that appears straightforward in a transcript may carry a different meaning when the interviewer recalls the circumstances in which it was spoken.
Can AI Detect Irony and Sarcasm in Interview Transcripts?
AI can recognize some linguistic patterns associated with irony and sarcasm, especially when the surrounding context makes the contrast between literal wording and intended meaning apparent.
For example:
"Wonderful. The internet disconnected again, right in the middle of my examination."
The contrast between "wonderful" and the described problem provides evidence that the speaker may be sarcastic.
However, sarcasm is not always signaled through obvious contradictions. Participants may use understatement, familiar expressions, or humor understood primarily within a particular community.
Research on sarcasm detection has repeatedly identified contextual dependence as a major challenge. Oprea and Magdy's work on sarcasm detection examined the importance of conversational context and speaker information, demonstrating why individual statements cannot always be interpreted adequately in isolation.
These findings do not establish a universal accuracy rate for generative AI in qualitative interviews. They do support the broader methodological concern that sarcasm cannot always be reduced to recognizable keywords or sentence patterns.
Why Ambiguity Should Not Always Be Eliminated
Researchers sometimes assume that every ambiguous statement must be assigned one correct interpretation before analysis can proceed.
Yet ambiguity may itself be analytically meaningful.
Consider a participant describing a demanding online learning arrangement:
"I suppose it works for us."
The statement could express qualified acceptance, uncertainty, resignation, or a desire to avoid criticism. Without further context, selecting one meaning may be premature.
More importantly, the participant may genuinely experience competing feelings. The arrangement might provide convenience while simultaneously creating substantial difficulties.
AI-generated analysis may attempt to resolve the ambiguity by selecting a confident interpretation. A more defensible approach may preserve the uncertainty and investigate whether it reflects a meaningful tension in participants' experiences.
How Can Culturally Specific Expressions Be Misinterpreted?
Expressions acquire meaning through language use within particular communities. A literal translation may preserve the words while changing their pragmatic force.
Consider several hypothetical Filipino interview statements:
| Expression |
Possible literal reading |
Contextual possibilities |
| "Bahala na." |
Leaving the outcome to circumstances |
Determination despite uncertainty, resignation, acceptance, or risk-taking |
| "Okay lang, sanay na." |
It is fine; I am used to it |
Genuine adaptation, normalization of hardship, reluctance to complain, or resignation |
| "Nakakahiya naman." |
It is embarrassing |
Personal embarrassment, social discomfort, concern about imposing, or perceived impropriety |
| "Sige na nga." |
All right, then |
Reluctant agreement, playful acceptance, impatience, or genuine consent |
These are illustrative possibilities, not fixed translations or universal cultural meanings. Usage varies across speakers, regions, relationships, and situations.
An AI system may select a familiar interpretation even when the participant's intended meaning differs. Researchers should therefore avoid treating culturally specific expressions as though each has one stable psychological or social meaning.
What Happens When Participants Switch Between Languages?
Code-switching can communicate distinctions that are difficult to preserve through translation.
A Filipino participant might use English when discussing institutional procedures but switch to Filipino when describing frustration or personal experience. Such a shift may be analytically relevant, although it should not automatically be interpreted as evidence of a particular emotional state.
Generative AI may translate the entire passage into fluent English and remove the linguistic shift. The resulting text may appear clearer while concealing a feature of the original account.
Researchers should retain the source-language transcript where feasible, document translation decisions, and consider whether the language choice contributes to the research question.
For multilingual analysis, a model's ability to produce grammatically correct translations does not establish equivalence of meaning.
Can AI Understand Indirect Disagreement?
Participants do not always disagree explicitly, particularly when discussing individuals or institutions with authority over them.
Suppose an interviewer asks whether a university's online examination policy is fair. A student responds:
"Well, that's the policy, so we just follow it."
The statement reports compliance. It does not necessarily express approval.
AI might classify the response as acceptance of institutional policy, but the participant could be communicating resignation, reluctance to criticize, or a distinction between compliance and agreement.
The researcher should examine the interactional context and avoid attributing approval merely because the participant did not explicitly object.
Likewise, silence or hesitation should not automatically be coded as disagreement. Such cues may have multiple explanations.
Why Translation Can Distort Qualitative Coding
When interviews are translated before coding, analytical decisions may be based on the translated version rather than the participant's original expression.
Translation is not a neutral substitution of words. Choices concerning idioms, register, emotional intensity, and culturally specific concepts can influence how an account is interpreted.
For example, translating "nakakahiya" exclusively as "ashamed" may suggest a stronger or more specific psychological state than the original context warrants.
Temple and Young's methodological discussion of qualitative research and translation emphasizes the significance of language choices in the production and interpretation of qualitative evidence.
Researchers should document whether coding occurred in the original language, translated material, or both. When important interpretations depend on translation, review by someone with appropriate linguistic and contextual competence may be necessary.
Can Providing More Context Solve the Problem?
Providing the full interview, relevant field notes, participant background where ethically appropriate, and information about the research setting can help reduce interpretive ambiguity.
However, more context does not guarantee correctness.
A transcript may omit tone or gestures. Field notes may contain the researcher's own interpretations. Background information may encourage stereotypes if the model uses group membership to infer motives.
There is also a risk that long inputs are summarized, truncated, or incompletely considered by the system.
Researchers should therefore distinguish contextual information that genuinely supports an interpretation from assumptions that merely make a generated explanation sound plausible.
How Do Contextual Errors Affect Coding and Theme Development?
A mistaken interpretation can influence subsequent stages of analysis.
Suppose AI interprets "Okay lang, sanay na kami" as evidence of resilience. It may assign the code "Positive Adaptation," group similar statements under "Student Resilience," and eventually generate a theme suggesting that students successfully overcome technological barriers.
If the expression actually communicates resignation or normalization of hardship, the analytical direction may be misleading.
The error can therefore move from an individual code into a broader thematic claim.
Researchers should scrutinize initial qualitative codes and candidate themes developed from interviews when they depend on ambiguous or culturally situated language.
Can AI Identify Several Plausible Interpretations?
Yes. Asking AI to produce alternative readings may be more useful than asking it to determine the single correct meaning.
For an ambiguous expression, the model might identify possible interpretations, indicate which contextual cues support each, and specify what additional evidence would help distinguish among them.
This approach can support analytical openness. However, the model may still omit a culturally important reading or introduce speculative possibilities that are not relevant to the participant's account.
The researcher should treat alternative interpretations as questions to examine rather than equally credible explanations.
Does Cultural Competence Guarantee Interpretive Accuracy?
No. Even researchers who share a participant's language or cultural background may interpret an expression differently.
Shared cultural knowledge can support interpretation, but participants' experiences also vary by region, generation, social position, and personal history.
Researchers should avoid assuming that either AI or a human analyst possesses complete access to another person's intended meaning.
Where uncertainty remains, it may be appropriate to preserve ambiguity, compare relevant accounts, or acknowledge interpretive limitations rather than impose a definitive explanation.
This reflects the broader distinction between describing what participants said and interpreting what their statements may mean.
04 · A Practical Example
When AI Mistakes Resignation for Resilience
Hypothetical Example
Interpreting a Filipino Student's Experience of Online Learning
A researcher interviews university students about their experiences of repeated internet interruptions during online examinations. One participant provides the following hypothetical exchange:
Interviewer: "How do you feel when the internet disconnects during an examination?"
Participant: "Okay lang, sanay na kami. Wala rin naman kaming magagawa."
A possible English rendering is: "It's okay, we're used to it. There's not much we can do anyway."
What an AI System Might Initially Conclude
Potentially Misleading Interpretation
"The participant demonstrates resilience and positive adaptation to technological challenges. Their response indicates acceptance of connectivity problems and an ability to cope effectively."
The interpretation emphasizes adaptation while overlooking the second statement, which may suggest limited perceived control.
The participant's words do not establish that the experience is positive or that the coping response is effective.
What Other Interpretations Are Possible?
| Possible reading |
Supporting consideration |
What remains uncertain |
| Adaptation |
The participant reports becoming accustomed to interruptions. |
Whether adaptation is experienced positively. |
| Resignation |
The participant states that little can be done. |
Whether the statement reflects broader feelings of powerlessness. |
| Normalization of difficulty |
Repeated problems appear to have become familiar. |
Whether the participant considers the situation acceptable. |
| Indirect criticism |
The expression may communicate dissatisfaction without direct complaint. |
Whether criticism was intended in this interaction. |
These readings are not necessarily mutually exclusive. A participant may adapt to a difficult situation while also feeling resigned to it.
A More Defensible Provisional Interpretation
Revised Interpretation
"The participant describes becoming accustomed to recurring internet interruptions while also expressing limited perceived ability to change the situation. The account may suggest normalization of technological difficulties, although the participant's emotional evaluation and intended meaning require further contextual examination."
This interpretation remains close to the evidence and preserves uncertainty about meanings not explicitly established.
How the Researcher Should Verify the Meaning
1. Preserve the original-language statement.
Retain the Filipino wording rather than relying exclusively on an English translation.
2. Examine the surrounding conversation.
Review preceding questions, subsequent explanations, and any relevant interactional context.
3. Review available audio or field notes.
Where ethically permitted and methodologically relevant, examine tone, pauses, and other recorded contextual cues.
4. Consider alternative readings.
Evaluate whether the statement reflects adaptation, resignation, criticism, or overlapping experiences.
5. Compare with the wider dataset.
Examine whether other participants describe similar expressions or experiences and how their accounts clarify the emerging pattern.
6. Document the interpretation.
Record the evidence, translation decisions, unresolved ambiguity, and any AI suggestions considered.
The objective is not to identify a culturally correct label from a list. It is to develop an interpretation that is sensitive to the participant's language and circumstances while acknowledging the limits of the available evidence.
06 · What This Means for You
How to Use AI Without Losing Cultural and Conversational Meaning
Use AI-generated interpretations to identify possibilities, not to settle meanings that remain uncertain.
When working with multilingual or culturally situated qualitative data, the researcher should preserve the original language where feasible and examine whether translation changes the analytical significance of an expression.
Choose the Appropriate Verification Strategy
If an expression has several plausible meanings
Preserve the ambiguity and examine contextual evidence before selecting an interpretation.
If the interpretation depends on tone or irony
Review the original recording or relevant field notes where available and permitted.
If the interview contains code-switching or local expressions
Retain the source-language text and verify translations with appropriate linguistic and contextual expertise.
If AI attributes an unstated emotion or motive
Require supporting evidence or revise the claim to distinguish inference from participant testimony.
If contextual meaning remains unresolved
Acknowledge uncertainty rather than presenting a speculative interpretation as established fact.
Ask AI to Preserve Ambiguity Rather Than Eliminate It
A practical prompt might be:
"Examine the supplied anonymized interview excerpt in its original language and conversational context. Identify what the participant explicitly communicated and propose plausible alternative interpretations of any irony, ambiguity, code-switching, or culturally specific expressions. Explain which contextual cues support each reading and what information remains missing. Do not assume a single intended meaning, invent emotional states, or treat cultural stereotypes as evidence."
This may encourage more careful analysis, but the model's suggested interpretations still require researcher verification.
Record Translation and Interpretive Decisions
When meaning depends on a particular translation or contextual judgment, document the original expression, the translation used, alternatives considered, and the reasons for the final interpretation.
Such decisions may be recorded through qualitative analytical memos. The record should reflect the researcher's actual reasoning rather than an AI-generated account of decisions that never occurred.
Do Not Assume AI Has the Researcher's Situated Understanding
Researchers may possess contextual knowledge arising from fieldwork, professional experience, language familiarity, or relationships with the research setting.
Such knowledge is not infallible, but it may be essential for evaluating interpretations that cannot be established from the transcript alone.
AI can suggest contextual possibilities. It should not be treated as possessing the same situated relationship to the participants or as independently resolving the researcher's interpretive responsibilities.
These considerations also inform questions about positionality in qualitative research and the limitations of delegating interpretive work to AI.
Protect Confidentiality When Supplying Context
Contextual details can improve interpretation but may also increase re-identification risks. Information about a participant's workplace, community, family circumstances, or distinctive experiences may be sensitive even when names have been removed.
Before submitting such material to an AI service, verify that the processing is permitted by participant consent, ethics approval, institutional policy, and applicable data protection requirements.
When AI materially contributes to contextual interpretation, describe its role according to applicable requirements for reporting AI-assisted qualitative analysis.
