Large language models (LLMs) such as ChatGPT have moved rapidly from research labs into everyday life, especially in multilingual, digitally connected societies like India. From students seeking study help to professionals drafting job applications, millions of users now rely on conversational agents for information search, content production, and informal social exchange. Because prompts—the short strings of text users type to solicit model responses—sit at the intersection of human language and machine affordances, they constitute a uniquely informative corpus for both linguists and engineers. Prompts are at once ordinary utterances and technocommunicative artifacts: they reveal users’ communicative goals, cultural norms, linguistic repertoires, and tacit expectations about machine competence.
This paper reports a mixed-methods analysis of 238 unedited user prompts from Indian users interacting with ChatGPT. Our goals are threefold: (1) to document the linguistic features that characterize Indian user prompts, with particular attention to code-switching and transliteration; (2) to map interactional patterns—how prompts function as speech acts, how users manage context and follow-ups, and the strategies they adopt to guide model behavior; and (3) to draw implications for design and deployment of language-aware conversational AI. By combining quantitative NLP processing (language identification, syntactic profiling, clustering) with qualitative discourse coding, we aim to bridge micro-level linguistic description with macro-level interactional patterns that matter for both sociolinguistic theory and practical model improvement.
The corpus comprises 238 unedited prompts submitted by Indian users to a mainstream conversational LLM over a defined collection window. Prompts were de-identified and stripped of any direct personal identifiers prior to analysis. Because prompts are user-generated, the study follows a conservative ethical stance: only prompt text was analyzed; no system logs, IP addresses, or demographic metadata were used. Where examples are provided below, they are paraphrased and anonymized to avoid re-identification. The study treats the prompts as public-facing communicative acts and analyzes them under standard research ethics consistent with Institutional Review Board guidance for text corpus work.
Although the dataset is modest in size, it offers high signal because entries are complete, user-authored turns with minimal editorial filtering. For analysis we first applied automatic preprocessing:
Normalization: Unicode normalization and consistent handling of Indic scripts and Roman-script transliterations. Many Hindi and other Indic language tokens appear in Latin script (so-called transliterated text, e.g., “namaste”, “acha”).
Language detection: A multilingual pipeline combining probabilistic language detectors (e.g., fastText language ID) and rule-based heuristics for short utterances was used. Code-switched strings were flagged when detectors reported mixed probabilities or when transliteration patterns were present.
Tokenization and sentence segmentation: Tools were selected to be robust to code-mixing; for Devanagari tokens we used Indic-aware tokenizers, and for mixed Latin/Devanagari sequences tokenization was tuned to avoid spurious segmentation.
Morphosyntactic annotation: Where possible, we applied multilingual models (Stanza / UD tools) to obtain POS tags and dependency parses. Results were interpreted cautiously for mixed inputs because off-the-shelf parsers are optimized for monolingual text.
Preprocessing prioritized preserving the original surface form (punctuation, capitalization, and spacing) as these features themselves are meaningful signals for user intent and politeness. For example, users often rely on punctuation (question marks, ellipses) or imperative forms to bias the model’s response style.
We used a mixed-methods pipeline combining quantitative NLP analysis with hand-coded qualitative work.
Quantitative components
Descriptive statistics: Prompt length (in tokens and characters), distribution across language labels (English, Hindi in Devanagari, romanized Hindi or “Hinglish”, other Indian languages), and frequency of punctuation/orthographic markers were summarized to provide corpus structure.
Syntactic and lexical profiling: POS distributions and dependency relation counts highlighted prominence of question constructions, imperatives, and nominal requests. Type–token ratio (TTR) and measures of lexical density were computed to compare short conversational prompts versus longer, task-oriented prompts.
Clustering and topic induction: Embedding-based clustering (using multilingual transformer embeddings such as mBERT/XLM-R representations) allowed grouping prompts into functional clusters (e.g., information requests, task delegation, creative writing, code/technical queries). Topic modeling (LDA/BERTopic) aided extraction of thematic groupings.
Qualitative components
Coding scheme development: Drawing on speech act theory (Searle) and discourse analysis traditions, we constructed a hierarchical codebook with categories for communicative function (assertive, directive, interrogative, commissive, expressive), politeness strategies (directive vs mitigated), and context dependence (standalone vs context-dependent).
Thematic analysis: Using Braun and Clarke’s approach, we iteratively identified themes emerging across prompts (education/career, exam preparation, translation, emotional support, technical help). We intentionally coded for socio-cultural markers (references to family, honorifics, exams) that signal expectations and social positioning.
Case study selection: From each cluster, representative prompts were selected for detailed microanalysis. These exemplars illustrate how code-switching, transliteration, and rhetorical packaging function to shape model responses.
Reliability and validity
Inter-coder reliability was assessed on a held-out subset of prompts using Cohen’s kappa for categorical codes; categories with low agreement were refined. We also triangulated automatic labels with manual checks to mitigate errors from automatic language detection on short, mixed inputs. Given the modest corpus size, emphasis was placed on interpretive depth rather than exhaustive statistical generalization.
The corpus lacks user demographic metadata (age, education, region), which constrains sociolinguistic inferences about community-level practices. Short prompts challenge automated language identification and syntactic parsing; thus quantitative results must be read alongside qualitative insights. Finally, because prompts are uncontextualized single turns without full conversation histories, our analysis of multi-turn strategies relies on prompts that explicitly include prior context (e.g., “As I said before…”).
Language in these 238 prompts is not merely a vehicle for requests; it’s a register in which speakers negotiate identity, expertise, and expectations with a machine interlocutor. Three crosscutting linguistic phenomena stand out: code-switching and transliteration, syntactic and rhetorical preferences, and variations in length and complexity.
Prevalence and surface forms. A salient pattern is widespread code-mixing between English and Hindi, often realized as romanized Hindi tokens embedded within English syntactic frames (e.g., “Can you help me write a formal mail in Hindi with polite tone?” where “mail” is English and additional Hindi transliterations or particles may appear). We observed several code-switching types:
Intersentential switching: complete sentence alternation across languages (less common).
Intrasentential switching: alternating language within the same sentence or clause — the predominant pattern.
Tag switching: short discourse tags or particles from Hindi (e.g., “yaar”, “ji”, “achha”) appended to English requests to convey familiarity or politeness.
Functions of switching. Code-switching serves pragmatic functions beyond lexical choice: it indexes register and social stance. Hindi small talk markers and honorifics can soften directives (e.g., “Please explain, yaar”), while English is often preferred for technical vocabulary (programming, exam names). Transliteration is highly frequent: users prefer Latin script for Hindi words, leveraging the convenience of Roman keyboards and the expectation that the model will understand transliterated tokens. This practice creates noise for language detectors but is communicatively efficient for users.
Implications. For model developers, mixed-script inputs present two challenges: first, language identification must handle transliteration robustly; second, semantic embeddings need to align transliterated words with their script-native counterparts so that semantic similarity is preserved across script variations.
Dominance of questions and imperatives. Analyses reveal a heavy skew toward interrogative and directive forms. Users frequently adopt concise interrogatives (“How to prepare for X?”) or imperatives framed as requests (“Write a 300-word cover letter for a sales role”). This aligns with the action-oriented nature of prompt use: users are task-driven and expect concrete output.
Politeness and mitigation. Even in directives, mitigation strategies are common: “Can you please…”, “Could you help me by…”, and prefatory framing (“I need to send this to my manager, so please make it formal”). Mitigation serves to signal social norms (deference in formal contexts) and to cue the model toward a respectful register. When users switch to Hindi particles (e.g., “ji”), the politeness function becomes culture-specific and uses local honorifics.
Rhetorical compression and repetition. Because many prompts are typed quickly on mobile devices, rhetorical simplification is frequent: ellipsis, truncated clauses, and repeated emphasis (e.g., “urgent urgent please”) are used to foreground urgency. Repetition acts as an intensifier, and models often treat repeated tokens as higher priors for user priority—something engineers should note when weighting prompt tokens.
Specialized framing for tasks. Task prompts frequently include scaffolding tokens—“Act as”, “Explain like I’m five”, “Give bullet points”—which function as higher-level instruction tokens that steer model style and granularity. Such metacommands are an emergent part of lay prompt engineering.
Short vs long prompts. We find a functional bifurcation along prompt length. Short prompts (1–10 tokens) are typically conversational checks or quick information requests: “What is GST?” or “Translate: XYZ”. Longer prompts (30+ tokens) tend to specify task constraints (tone, word limit, audience) and are associated with higher-precision responses. Long prompts often contain multi-sentence context, explicit evaluation criteria (e.g., “use formal tone, include three points”), and sample content to be transformed.
Complexity measures. Syntactic complexity correlates with task sophistication: prompts that request rhetorical or evaluative outputs (e.g., cover letters, essays) use more subordinated clauses and conditional frames. Lexical density is higher in technical or educational prompts, where domain terms appear frequently. Conversely, social or emotive prompts feature more expressive vocabulary and interjections.
Transliteration effects on complexity. Transliterated Hindi increases token variability and confounds straightforward lexical counting. For example, the same Hindi word may appear as “sambhalna”, “sambhalnaa”, or in Devanagari script; clustering methods that account for orthographic variation are therefore essential.
Domain signaling. Certain lexical clusters act as reliable domain signals: “resume/CV”, “cover letter/interview”, “JEE/NEET/board” (exam-related), “code/error/compile” (programming). These clusters help the model disambiguate user intent even when syntactic cues are minimal.
Named entities and cultural references. Users frequently include culturally specific references (festival names, local institutions, colloquial expressions) which models trained primarily on global English corpora may partially understand but may miss nuanced local connotations.
Pragmatic markers. Fillers such as “please”, “thanks”, and emotive punctuation (smileys) function as softeners and social cues; they shape the tone that users expect in return.
Prompts are not isolated utterances but elements of interactional sequences: they position the user relative to the model (expert vs novice), manage expectations, and signal the desired form of the reply. Below we unpack core interactional patterns and the socio-pragmatic work users perform when engaging ChatGPT.
Based on clustering and manual coding, prompts broadly fall into three functional categories, each with subtypes:
A. Information requests (Inquiry)
Factual lookup: “What is the deadline for filing taxes in India?”
Explanatory: “Explain the difference between supervised and unsupervised learning.”
Comparative/evaluative: “Which MBA program is better for analytics?”
Information requests often require concise, authoritative outputs. Users tend to favor succinct interrogatives with minimal contextual framing when seeking facts; however, for complex explanations they often ask for examples or analogies.
B. Task delegation (Production)
Text generation: “Write a 200-word professional bio.”
Transformations: “Translate this paragraph into Hindi, keep formal tone.”
Problem solving / coding: “Debug this Python snippet / suggest improvements.”
Task delegation prompts often embed explicit constraints: tone, length, audience, or formatting. These metaprompts function as instruction templates and show that many users implicitly understand the model as a tool for articulated production.
C. Social and affective interaction (Social talk)
Emotional support: “I feel stressed about exams; any tips?”
Idle chat / humor: “Tell me a joke about office life.”
Identity negotiation: “How to respond to a rude manager politely?”
Though less frequent than the first two categories, social prompts reveal how users project human roles onto the model—advisor, friend, coach. They also highlight expectation management: users may expect empathy or validation from the model even when it cannot truly experience feelings.
Standalone prompts. A large share of prompts are self-contained single turns where all necessary information is present. These favor immediate, one-shot outputs and are optimized for speed.
Context-dependent prompts. Some prompts explicitly reference prior content (“As above”, “From my previous message”), indicating multi-turn sequences. When users include prior context within a single prompt (copy-pasting earlier model output), they demonstrate an awareness of the model’s limited short-term memory and strategically compress context into the prompt itself.
Follow-ups and iterative refinement. Users often follow a generated reply with iterative directives (“Make it shorter”, “Use simpler words”). This pattern reveals a collaborative workflow: initial automated draft + human refinement through targeted prompts. Iterative prompting is a form of emergent co-authoring where the user and model negotiate content through a sequence of micro-edits.
Cooperative strategies. Many users frame requests to encourage cooperation: politeness markers, explicit role assignment (“You are my career counselor”), and constraint tokens. Such framing is effective at steering model behavior toward desired registers and output shapes.
Adversarial / test prompts. A minority of prompts are designed to test or trap the model (e.g., asking it to reveal internal mechanics, to contradict itself, or to produce disallowed content). These illustrate user curiosity about model boundaries and occasionally provoke defensive or refusal patterns in the model response.
Negotiation of authority. Users alternate between deferring to the model for facts and contesting its outputs. For example, after receiving a response users sometimes challenge factual claims (“Are you sure about that source?”), seeking source attribution or correction. This points to an emerging epistemic relationship: the model is treated as an authoritative yet contestable source.
Role of honorifics and politeness markers. Indian users frequently append honorifics (e.g., “ji”) and local discourse markers to adjust formality. Those linguistic choices subtly index social distance and desired politeness.
Expectations of local knowledge. Users expect the model to understand local institutions, exam formats, and cultural practices. When the model fails, users often rephrase or supply local context, revealing a practice of contextualizing the model to local realities.
Language as social index. Choice of language (English vs Hindi vs local language) often correlates with perceived audience and social aspiration. English is preferentially used for formal, career, or academic tasks, while Hindi or local languages are used for intimate, emotional, or community-oriented queries. This distribution reflects sociolinguistic prestige patterns and suggests different interactional expectations across registers.
Applying speech act theory, prompts function mainly as directives (requests for action) and interrogatives (requests for information), with fewer assertives or expressives. However, the illocutionary force is often masked: a short “Resume?” may be both an information query and an implicit request for a résumé draft.
Understanding illocutionary intent is crucial for models because surface forms alone are insufficient: the same phrase can have distinct desired outcomes depending on user goals. For instance, “Explain X” might be intended for a novice learner (ELI5) or for an expert who wants a succinct summary. Users often supply metaprompts ("simplify", "formal") to clarify intent.
From the corpus, several recurrent user strategies emerge:
Role specification: “Act as a professional interviewer” to set voice and stance.
Output scaffolding: specifying format: bullet points, word count, or sections.
Error correction and verification: explicit requests for sources or cross-checks.
Constraint stacking: combining multiple constraints (tone, length, audience) in single prompts.
Transliteration plus glossing: providing transliterated local terms with quick glosses to ensure correct interpretation (“samosa (a snack)”).
These strategies are practical forms of prompt engineering by non-expert users and suggest that usable interfaces should make such scaffolding easier through templates or UIs that expose common constraints.
Prominent sources of miscommunication include:
Ambiguity in short prompts: under-specified requests result in generic or off-target outputs.
Script mismatch: transliterated terms may be misinterpreted, especially proper nouns.
Cultural misalignment: model outputs that ignore local conventions (e.g., inappropriate formality) may be rejected by users.
Overfitting to global norms: models trained on predominantly global English corpora may default to idioms and cultural references unfamiliar to Indian users.
Addressing these requires both better multilingual training data and user interface affordances that allow users to specify context in structured ways without heavy typing.
This section synthesizes the analyses above into interpretable findings, explores implications for model design and social impact, compares Indian prompting practices with existing literature from Western contexts, and outlines future research directions.
1. Multilingual interaction is the norm, not the exception. The corpus shows pervasive code-switching, transliteration, and mixing of register across English and Hindi. This multilingual practice reshapes prompt structure: users routinely embed local discourse markers and domain-specific transliterations in otherwise English frames. For model builders, this underscores the need for robust cross-script lexical alignment and transliteration-aware tokenization.
2. Prompt form mirrors functional demand. Short, minimal prompts are associated with quick factual lookups and conversational checks; longer, more elaborate prompts are linked with complex production tasks (resumes, essays, translations). Users have organically developed prompt scaffolds—metaprompts and role assignments—that function as lay prompt engineering techniques.
3. Cultural markers modulate interactional expectations. Honorific particles and local idioms signal politeness preferences and desired tone. Users expect models to understand not only lexical meaning but also cultural valence and appropriateness. Where the model fails to align with local tone, users either adapt the prompt (adding explicit tone constraints) or abandon the system for human assistance.
4. Iterative co-authoring is a common workflow. Users frequently refine model outputs through short corrective prompts. This iterative process—initial generation followed by human edits—positions models as drafting aids rather than final authorities.
One of the most actionable findings is the strong relationship between prompt specificity and output quality. Detailed prompts that include role, audience, tone, and length constraints reliably yield outputs closer to user expectations. Conversely, vague prompts produce outputs that require substantial human revision. This suggests two complementary design strategies:
Educate users about effective prompting: lightweight onboarding or built-in templates (e.g., “I need a formal email; audience: manager; length: 150 words”) can raise the baseline quality of outputs.
Improve model steers: models should be better at requesting clarification when prompts are under-specified rather than producing low-value generic outputs. A simple clarification question (“Do you want formal or informal tone?”) can substantially increase relevance.
Technical challenges. Transliteration and code-mixing confound standard pipelines. Language identification often fails on short, mixed strings; downstream parsing and entity recognition degrade accordingly. Remedies include:
Training tokenizers and embeddings on transliterated corpora.
Implementing transliteration normalization layers that map Latin-script Hindi tokens to Devanagari equivalents before downstream processing.
Fine-tuning multilingual models (mBERT, XLM-R) on code-switched corpora to improve robustness to intrasentential switching.
Socio-political implications. The linguistic prestige attached to English influences task allocation: high-stakes, aspirational tasks are more likely to be framed in English. This creates a risk that model performance disparities across languages may reproduce social inequities. Ensuring equitable performance across local languages is therefore not only a technical goal but a social justice imperative.
Existing research on prompt behavior (primarily from English-dominant user bases) emphasizes the rise of few-shot prompting, role-assignment prompts, and instructive metaprompts. Indian prompts share many of these features but diverge along two axes:
Higher incidence of code-switching and transliteration, reflecting multilingual keyboard practices and sociolinguistic repertoires.
Domain focus skewed toward education and exam preparation, reflective of the social prominence of entrance exams and competitive educational pathways in India.
These differences suggest that transfer of prompt engineering practices from Western contexts will be incomplete unless models accommodate script variation and prioritize training on locally relevant domains.
Our analysis contributes to three overlapping literatures:
Sociolinguistics of digital media: prompts function as a new register shaped by affordances of conversational AI. They reveal emergent norms for human–machine communication in multilingual contexts.
HCI and conversational design: the study documents real user-level prompt engineering strategies and highlights the efficacy of role-assignment and output scaffolding, pointing to interface improvements.
Speech-act theory in human–AI interaction: prompts instantiate a spectrum of illocutionary forces where directive speech acts dominate but are frequently mitigated by social markers; recognizing this helps models select appropriate pragmatic strategies.
Limitations. The study’s sample size, while richly informative, is limited, and the absence of demographic metadata prevents strong sociological claims. Automated analyses are constrained by off-the-shelf parsers that struggle with code-mixing and transliteration. Finally, because prompts are single turns, some multi-turn interactional phenomena are necessarily under-observed.
Future directions. We propose several next steps:
Scale up: gather larger, stratified corpora across regions, languages, and user cohorts to test generalizability.
Longitudinal studies: follow users over sequences to understand evolving strategies and trust dynamics.
Intervention experiments: test interface features (prompt templates, clarification prompts) to measure measurable improvements in output quality and user satisfaction.
Model advances: invest in transliteration mapping layers and code-switching fine-tuning to close performance gaps.
Model deployment in multilingual societies raises governance questions. Data collection must protect user privacy and avoid harvesting sensitive content. Further, model evaluation should include language equity metrics so that systems do not privilege English-speaking use cases. At an institutional level, educational and governmental stakeholders should be engaged to develop culturally informed guidelines for deploying conversational AI in high-stakes contexts (exams, legal advice, medical triage).
The 238 unedited prompts analyzed here reveal an ecology of human–AI interaction that is richly multilingual, pragmatically inventive, and oriented toward practical task completion. Users have already invented pragmatic prompt engineering techniques—role assignment, scaffolding, iterative refinement—that can be leveraged to improve model design. For conversational AI to be equitable and useful in India and similar multilingual environments, research and development must prioritize code-switching robustness, transliteration alignment, and interface affordances that make context explicit without burdening users.
Appadurai, A. (1996). Modernity at Large: Cultural Dimensions of Globalization. University of Minnesota Press.
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency.
Biber, D. (1988). Variation Across Speech and Writing. Cambridge University Press.
Bourdieu, P. (1986). The forms of capital. In J. Richardson (Ed.), Handbook of Theory and Research for the Sociology of Education (pp. 241–258). Greenwood.
Braun, V., & Clarke, V. (2006). Using thematic analysis in psychology. Qualitative Research in Psychology, 3(2), 77–101.
Brown, T., Mann, B., Ryder, N., Subbiah, M., et al. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems, 33.
Couldry, N., & Hepp, A. (2017). The Mediated Construction of Reality. Polity.
Couldry, N., & Mejias, U. A. (2019). The Costs of Connection: How Data Is Colonizing Human Life and Appropriating It for Capitalism. Stanford University Press.
Fairclough, N. (1995). Critical Discourse Analysis: The Critical Study of Language. Longman.
Floridi, L. (2014). The Fourth Revolution: How the Infosphere is Reshaping Human Reality. Oxford University Press.
Floridi, L., & Cowls, J. (2019). A unified framework of five principles for AI in society. Harvard Data Science Review.
Gumperz, J. J. (1982). Discourse Strategies. Cambridge University Press.
Hovy, D., & Spruit, S. (2016). The social impact of natural language processing. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (ACL).
Noble, S. U. (2018). Algorithms of Oppression: How Search Engines Reinforce Racism. NYU Press.
Reeves, B., & Nass, C. (1996). The Media Equation: How People Treat Computers, Television, and New Media like Real People and Places. Cambridge University Press.
Searle, J. R. (1969). Speech acts: An essay in the philosophy of language. Cambridge University Press.
Wodak, R. (2015). The Politics of Fear: What Right-Wing Populist Discourses Mean. SAGE.
