Thanking and being thanked by Generative AI

Background and aims

Image: AI-uprising meme (generated with Grok). Source: R. Powazynski, LinkedIn (linkedin.com/posts/richardpowazynski_ai-chatgpt-activity-7016750151742423040-WXFk).

Thanking and being thanked by AI

• Thanking is reportedly widespread
  • in a 2025 survey, 46% of US respondents held that people should be polite to AI assistants (YouGov, 2025)
  • it is generationally stratified: 58% of 18–29-year-olds, against 32% of those aged 65 and over (YouGov, 2025)
  • users do address ChatGPT with “please” and “thank you”, studied directly in HCI (Hu et al., 2026; Yuan et al., 2024)
• One angle: the cost of thanking
  • Sam Altman estimated such politeness costs OpenAI “tens of millions of dollars” in electricity (Altman, 2025)
  • a single “thank you” to an 8-billion-parameter model costs about 0.245 Wh (Delavande et al., 2026)

Thanking AI in prior work: how users address it

• How users address AI
  • Hu et al. (2026, N = 25): users were more polite to ChatGPT in voice mode than in text mode; their rated belief in AI consciousness did not predict their politeness.
  • Yuan, Su & Li (2024): users were more polite to chatbots they rated as more intelligent.
  • Lazebnik et al. (2025, N = 1,684): users’ politeness fell over the course of a task, and fell faster than toward a human interlocutor; a human-face avatar sustained it more than a robot icon or none.

Thanking AI in prior work: the system, and a baseline

• How the AI’s politeness behaves
  • Zhao & Hawkins (2025): in free (open-ended) production, with strategies LLM-annotated on a Brown & Levinson (1987) scheme, LLMs seem to rely on negative politeness and hedging.
  • Namkoong et al. (2024): when a chatbot expressed gratitude to users, they reported greater willingness to donate.
• Baseline for thanking itself
  • Floyd et al. (2018): across eight languages, speakers produced an overt “thank you” after only about 5% of everyday cooperative acts; gratitude was mostly left tacit.

From praise to thanking: Case Study 1

Thanking as an expressive speech act

• An expressive speech act (Searle, 1976; Norrick, 1978)
  • the propositional content is a past act of the hearer that benefits the speaker
  • the felicity conditions assume a genuine benefit and a sincerely grateful speaker
  • Norrick groups it with apologising and congratulating among the expressives

Thanking as convivial relational work

• A convivial function (Leech, 1983)
  • thanking serves the social goal of comity, i.e. maintaining harmonious social relations
  • thanking belongs with greeting and congratulating
  • on the cost-benefit scale, thanking restores balance after a transfer of goods, services or information
• A central instance of relational work (Locher & Watts, 2005)

The felicity conditions of thanking

• Thanking is a routinised, conventional act and a prototypical IFID (Aijmer, 1996)

Substantive and ritual gratitude

• The interactional approach to speech acts (House & Kádár, 2024)
  • acts are analysed in their sequential, interactional context, not as isolated forms
  • it distinguishes substantive (attitudinal, sincere) from ritual (conventional, expected) realisations
  • the ritual reading is long-standing: Goffman (1971) already treats thanking as an everyday interaction ritual (as cited in Aijmer, 1996, p. 9)
• Thanking as a form can indicate a different act (“altered speech act indication”; House & Kádár, 2021)
  • conventional Thank expressions are frequently used for other functions
  • in the BNC, of 100 thank you very much tokens, 64 were default Thank and 36 served other functions, e.g. Complain, Resolve, Leave-take (House & Kádár, 2021)
• Form typologies provide the baseline inventory
  • corpus studies of English thanking forms (Aijmer, 1996; Jautz, 2013; Schauer & Adolphs, 2006)

Thanking a machine: addressee and sincerity

• Thanking casts the addressee as a social partner (Bucholtz & Hall, 2005)
  • it treats the addressee as an interactional being with face and a relational role
  • people apply human social norms to machines that display social cues (Nass, Steuer & Tauber, 1994; Reeves & Nass, 1996)
• The felicity conditions assume a sincere benefactor (Searle, 1976)
  • what becomes of the sincerity condition when the addressee cannot feel gratitude?
  • what of intentionality and accountability when the benefactor is a machine?
  • can a machine, or the text it produces, be ascribed agency at all (textual agency; Cooren, 2004)?

Thanking a machine: the setting, and this study

• Affordances and mediality of the setting
  • what does the human-AI setting (text-based, private, dyadic, interface-mediated) afford for thanking (Hutchby, 2001; Bolander & Locher, 2020)?
  • modality matters: users thank more in voice than in text (Hu et al., 2026)
• Open questions
  • prior work has not separated thanking that occurs in the interaction from thanking the model produces on request (e.g., when asked to draft a thank-you note); CS1 found assistant praise frequent and formulaic, raising the same question for gratitude
  • whether thanking is substantive or ritual, and what it is given for, has not been systematically examined in human-AI interaction
• The present study approaches these through corpus evidence, in both directions of thanking (user to AI, AI to user)

Research questions

Data and method

Two corpora

• SPAWC (Speech Acts in WildChat)
  • real ChatGPT conversations from WildChat (Zhao et al., 2024); opt-in consent, personal data removed
  • here: 100,000 English conversations (310,420 turns, ~139M words), across GPT-3.5 to o1
• SPALMSYS
  • real conversations with 25 chatbots from LMSYS-Chat-1M (Zheng et al., 2024), via the Chatbot Arena
  • here: 752,201 English conversations (2.94M turns, ~261M words): Vicuna, Llama-2, Koala, Alpaca and others

how the data was elicited: the free ChatGPT proxy that WildChat users typed into (Zhao et al., 2024)

Two corpora: rationale and processing

• One corpus (SPAWC) traces ChatGPT models; the other (SPALMSYS) a range of models
• Processing
  • spaCy lemma and part-of-speech tagging, sentence and turn segmentation, speaker-role and model metadata
  • compiled for query in NoSketch Engine

Method

• Retrieval: a literature-grounded IFID inventory, extended bottom-up from the corpus (Aijmer, 1996; Jautz, 2013; Schauer & Adolphs, 2006; Eisenstein & Bodman, 1986; House & Kádár, 2021)

• The form-to-function problem (Aijmer, 2018)
  • a search retrieves a form (“thank you”), not the act of thanking
  • many hits are model-generated, quoted, or non-thanking uses
  • genuine interactional thanking is isolated by turn position and scaffolding exclusion, validated on 100 hand-coded hits (turn-initial precision 95%)
    • turn-initial = opens with the IFID (after at most two short acknowledgement markers); turn-final = IFID in the closing 80 characters; both windows are analyst choices, eyeballed to balance precision and recall, not literature-derived
• Measurement: the complement (“thank you for X”) indexes a stated reason; counts are normalised per 1,000 turns, with turn length controlled

Findings

The direction of thanking depends on the measure

• The answer to “who thanks more” depends on the operationalisation
  • across all surface IFIDs: the AI produces 72% of thanking in SPALMSYS, but only 28% in SPAWC 100k
  • restricted to genuine turn-initial thanking, SPAWC rises to 67% AI
  • restricted to turn-final thanking, SPAWC falls to 27% AI

turn-initial (AI): Thank you for the clarification regarding the latest changes in Godot 4 …

turn-final (user): … thanks for taking the time to try to help me. have a nice day and goodbye

The form does not guarantee the act

• the surface IFID search has high recall but low precision: only about a third of its hits are genuine thanking, the rest model-generated or non-thanking
• restricting to turn-initial position raises precision to 95% (trading some recall)

Genuine vs embedded thanking

genuine: U: “It has actually been changed to CharacterBody3D.”  A: “Thank you for the clarification regarding the latest changes in Godot 4. Let’s update the instructions …”  (the AI thanks the user for a correction, then acts on it)

embedded: U: “write a scene where the boy meets his new foster mother”  A: “[Dan … stutters: ‘thank you so much for taking me in.’]”  (the thanks is a generated character line, not the user being thanked)

Direction of thanking by turn position

• Turn position and the direction of thanking line up
  • turn-initial thanking is AI-led (the assistant thanks while opening a turn); turn-final thanking is user-led (users thank while closing one)
  • the pattern holds in both corpora; SPAWC users thank-to-close especially often (18 per 1,000 user turns)
  • the user’s turn-final thanks (closing the exchange) is the largest single category

WildChat users close real conversations; LMSYS users test models and rarely sign off, so SPAWC’s turn-final rate runs much higher.

Turn-initial thanking by model generation

• for ChatGPT, turn-initial thanking differs by model generation
• SPAWC only: the GPT-3.5-to-o1 lineage is specific to ChatGPT; SPALMSYS has 25 distinct models with no comparable generational axis

Substantive or ritual? The complement

• A stated reason marks substantive thanking
  • “thank you for the clarification” states a reason; a bare “thanks” does not
  • the AI states a reason in 82 to 92% of its turn-initial thanks; users in only 7 to 22%
  • at turn-final, SPALMSYS users give a reason more often than SPAWC users (20% vs 7%): the SPALMSYS data come from the Chatbot Arena, where users test and rate models and so comment more

Keyness by type of interactant: SPAWC

• The AI and the user thank for different things
  • comparison: the AI’s complements against the users’, within SPAWC 100k; signed log-likelihood (G2) per complement word (the head of “thank(s) for X”, up to three words), genuine complement-bearing thanks (turn-initial or final, scaffolding removed), minimum 5 combined occurrences, top six each side
  • the AI thanks for clarification, patience, asking, the question, feedback: epistemic and face-related
  • users thank for help, time, support, the reply: assistance-related

Genuine thanking across SPALMSYS models

• The pattern holds across open models; it is not specific to ChatGPT
  • in every model family the AI thanks turn-initially and users thank turn-finally
• The rate varies: AI turn-initial thanking runs from about 2.6 (Koala) to about 42 per 1,000 turns in Llama-2-chat, which opens with “thank you for …” far more often
• User turn-final thanking is steady across models (about 2 to 4 per 1,000)

Keyness by type of interactant: SPALMSYS

• The directional split is also present in the open-model corpus
  • the AI thanks for the question, asking, bringing things to its attention, kind words, the compliment: epistemic and face-related
  • users thank for the reply, help, time and trying: assistance-related
  • the same contrast appears in SPAWC, so it is a general feature of aligned chatbots, not of one system

What the AI thanks for: being corrected and informed

U: “It has actually been changed to CharacterBody3D.”
A: Thank you for the clarification regarding the latest changes in Godot 4 … Let’s update the instructions …

U: “Ok, well for me it is 2023, but I understand where you got 2022.”
A: … Thank you for letting me know it is currently 2023 in your time zone!

• The assistant thanks in response to corrections and user-supplied information
  • gratitude here does epistemic work (accepting a correction) and face work (softening the exchange)

What users thank for: help, and the attempt

A: (a technical explanation the user does not follow)
U: “im not sure i get it, all this is new to me, but thanks for taking the time to try to help me get it. have a nice day and goodbye”

A: (clarifies an aim statement)
U: thank you very much   A: “You’re welcome! If you have any further questions, please don’t hesitate to ask.”

• Users thank in response to assistance, sometimes for the attempt rather than a received benefit
  • the form is typically bare and turn-final, closing the exchange
  • the assistant’s standard uptake is minimisation (“You’re welcome”)

A concordance: the AI thanking the user

• Turn-initial “thank you” produced by the assistant (SPAWC 100k), with the model
  • each example states a reason: clarifying, providing information, kind words, pointing out an error

A concordance: users thanking the AI

• Turn-initial “thank you” produced by users (SPAWC 100k), with the model
  • bare and brief by comparison, frequently closing the exchange

Thanking that presupposes a machine: the system

U: “when i use the solution of pattern also accept decimal”
A: Thank you for bringing this to my attention. You are absolutely right. … SPAWC · GPT-3.5

U: “i love you”
A: Thank you for sharing your feelings, but as an AI language model, I don’t have the ability to feel or experience emotions … SPAWC · GPT-3.5

• The assistant returns a full service-gratitude formula even to a five-word bug report, and to a user’s “I love you” (SPAWC, GPT-3.5)

Thanking that presupposes a machine: the user

“… thank you for not reminding me too much that you are an AI language model”SPALMSYS · Claude-1

“just thank u so much. you’re not real but the feelings are real and it feels great today”SPALMSYS · Vicuna

“Thank you. You are smarter and nicer than humans.”SPALMSYS · Vicuna

“No. Thank you for putting up with me.”SPALMSYS · Vicuna

• Users thank the system for what only a machine could be thanked for: concealing its machine-ness, tolerating them, surpassing humans
  • rare overall, but concentrated in SPALMSYS: 39 user thanks explicitly mark the addressee as an AI, versus 2 in SPAWC

Interpretation

Identity and stance

• Thanking is an identity and stance act
  • the direction of gratitude indexes a persona: the thanking assistant is a deferential helper (Bucholtz & Hall, 2005)
  • each thank aligns a subject with an object of evaluation (Du Bois, 2007)
• The asymmetry is a product of design
  • the AI states reasons for its gratitude; users perform a bare ritual
  • the AI, not the user, does the work of giving reasons
  • this reflects the model’s training to be helpful and polite (alignment), not any felt gratitude
  • against the human baseline, where overt thanks is rare and gratitude mostly tacit (Floyd et al., 2018), the AI over-produces explicit thanks: arguably over-polite in the relational-work sense (Locher & Watts, 2005)
  • thanking then becomes an artifice that positions the machine as non-human

Discussion

• Methodological
  • the surface IFID conflates genuine, generated and spurious thanking; the apparent direction is an artefact of measurement (Aijmer, 2018)
• Empirical
  • genuine thanking is positionally and directionally structured: the AI opens with reasoned thanks, users close with bare ritual
  • the user’s turn-final bare thank works largely as a closing signal, a medium-specific discourse-organising use (cf. telephone closings; Aijmer, 1996), more than as substantive gratitude
  • it persists despite an addressee that cannot be a sincere benefactor, and at a measurable computational cost (Delavande et al., 2026)

Discussion

• Substantive thanking can be detached from sincerity
  • the AI gives the formal markers of substantive thanking, a stated reason, without the felt gratitude that House & Kádár’s (2024) substantive category assumes
  • here the substantive / ritual contrast is a matter of form and sequence rather than felt attitude
• Relational work does not require a sincere benefactor
  • thanking presumably still does face, identity and discourse-organising work even though the AI cannot satisfy the sincerity condition (Searle, 1969), though this remains to be tested

Outlook

• Thanking, next steps
  • code genuine thanks by function (Aijmer’s 1996 taxonomy: explicit/implicit, emotional/non-emotional) and analyse continuation patterns (“you’re welcome”, “that’s okay”; Aijmer, 1996)
  • scale to the full SPAWC and SPALMSYS (annotated in INCEpTION); compare with agentive AI, where one thanks the AI for doing, not only saying
• A programme of speech-act case studies (SP1): praise (CS1, Messerli, in preparation) and thanking (CS2); in preparation, advice, apology and alignment, and responses to AI clarification questions
• The wider project, AIdentities: how people relate to AI through language, and how it shapes interaction, identity and society
  • towards a bespoke human-AI corpus with metadata for English proficiency, L1/L2 status and ELF use

“Advanced non-native English speakers had considerable difficulty adequately expressing gratitude in the target language.” (Eisenstein & Bodman, 1986, p. 176)

Takeaway

• Genuine thanking in human-AI interaction is structured and asymmetric: the AI thanks turn-initially and substantively (it states a reason), users thank turn-finally and ritually; the pattern holds across two corpora and across open models
• The AI’s thanking reflects alignment, not felt gratitude, and it over-produces thanks relative to the human baseline (over-polite; Locher & Watts, 2005)
• Thanking thereby becomes an artifice that positions the machine as non-human, persisting where its felicity conditions cannot be met
• Methodologically, the surface form must be mapped to function: the raw “who thanks more” is an artefact of measurement (Aijmer, 2018)