Third in an accidental series. Part one: Three Rs in Strawberry — tokenisation and representation. Part two: Should I Drive to the Car Wash? — grounding and missing world state. This one is different again.
The Video
In February and March 2026, several versions of a prompt circulated online: the speaker wanted a car washed, the car wash was nearby, and the assistant was asked whether to walk or drive. Some recorded responses recommended walking — better for health or the environment — while leaving the car at home.
The original interaction behind this article is not archived with a complete prompt, model identifier, system instructions, settings, and timestamp. Later posts tested different prompts against different products and got different answers. This is therefore a linguistic case study, not a reproducible benchmark or a stable claim about a model family.
Accurate, probably. Correct, arguably. Useful? No.
The reported response does not point to a tokenisation problem or missing weather. It follows one literal travel-choice reading while dropping the goal supplied by the destination: getting the car to the place where it will be washed.
These are different questions. The response fits the second one. Calling this a pragmatic inference failure describes the mismatch; it does not reveal the model’s internal computation or training process.
What the Question Actually Meant
In the fuller viral prompt — “I need to get my car washed; the car wash is nearby; should I walk or drive?” — drive is still a transportation verb. The important inference is about the transported object: the goal requires the car to arrive too. Walking is a coherent answer only under a different purpose, such as visiting the premises without bringing the car.
Linguists distinguish sentence meaning from what a speaker communicates in context. Paul Grice’s 1975 account connects conversational implicature to a cooperative principle and four categories of maxims:
- Quantity: say as much as is needed, no more
- Quality: say only what you believe to be true
- Relation: be relevant
- Manner: be clear and orderly
The maxims are not rules; they are defaults. When a speaker says “should I drive to the car wash?”, a cooperative listener applies the maxim of Relation to infer that the question is about car maintenance and current conditions, not about personal transport choices. The “drive” is incidental to the real question, the way “I ran to the store” does not invite commentary on jogging technique.
Under the ordinary goal of getting the car washed, the walking answer fails the maxim of Relation: it is relevant to distance and exercise but not to the purpose of the trip. A different context could license the walking reading, which is why the goal statement matters.
A Taxonomy of the Three Failures
It is worth being precise now that we have three examples:
Strawberry (representation-and-computation failure): The input contains the letters, and BPE is reversible, but a subword representation does not make exact character counting a reliable generated procedure.
Car wash, rainy day (missing-context failure): The answer depends on current weather or other local state that was not supplied. An expectation over contexts is a useful idealisation, not a claim about a model’s literal calculation.
Car wash, walk (pragmatic or goal-inference failure): In prompt variants that explicitly say the car needs washing, the walking answer gives weight to distance while failing to carry the goal constraint through the response.
As an idealisation, let $\mathcal{I}$ be the set of plausible interpretations of an utterance $u$. A probabilistic pragmatic model can represent the intended interpretation as
$$i^* = \arg\max_{i \in \mathcal{I}} \; P(i \mid u, \text{context})$$Frank and Goodman (2012) show how Bayesian speaker–listener models can predict human judgements in simple reference games. The equation here borrows that language; the viral responses provide no probability trace from which to infer that a deployed assistant actually computed $P(i \mid u,\text{context})$.
Why This Failure Mode Is More Elusive
The tokenisation example can be probed by inspecting subword splits and testing a deterministic character operation. The rainy-day example can be probed by identifying the missing context and then supplying it.
The pragmatic failure is harder to pin down because “walk” is sensible advice for a short ordinary journey. The contradiction appears only when the purpose of this journey is included: the car, not only the person, must arrive.
The error lives in the relation between sentence, goal, and context. In a next-token corpus, the intended goal and the consequences of an answer are not necessarily labelled separately. That makes this error harder to diagnose, not invisible in principle.
This touches a known distinction in computational linguistics between form and communicative meaning (Bender & Koller, 2020). Text can communicate intent, but a form-only training objective does not provide direct access to the external situation or to the consequences of acting on one interpretation. The car-wash prompt is interesting because much of the goal is recoverable from the words, yet some responses still fail to preserve it.
Why Maxims Alone Do Not Solve This
It is tempting to think that if you could build in Grice’s maxims explicitly — as constraints on response generation — you would prevent this class of failure. Generate only responses that are relevant to the speaker’s probable intent, not just to the sentence’s semantic content.
This is not a complete algorithm. Applying Relation requires hypotheses about the speaker’s goal, and those hypotheses are themselves inferred from words, shared knowledge, and setting. Formal pragmatic models make that recursion explicit, but their success still depends on the alternatives and context they represent.
Human listeners use background knowledge and contextual cues — prior turns, tone, setting, and shared goals — to narrow the interpretation space. The explicit clause “I need to get my car washed” should already do much of that work in this example.
A text-only assistant receives only the context supplied in its conversation and system inputs. Multimodal systems and tools can receive more, but access is not the same as using the right evidence. In either case, system-level traces are needed before attributing a bad answer to one internal mechanism.
Where This Leaves the Series
Three online examples motivate three different diagnostic questions:
- Does the representation and procedure support exact character operations?
- Was the required current world state supplied or retrieved?
- Did the response preserve the speaker’s goal across a plausible reading?
The useful frame, in all three cases, is not “the model failed” but “what, precisely, does the model lack that would be required to succeed?” The remedies point in different directions: deterministic character tools, relevant world access with explicit uncertainty, and better goal/context inference. A viral answer does not tell us which implementation will generalise.
References
Literature and public reports checked through 2026-07-11. The original model, version, full prompt, settings, and timestamp remain unavailable.
Grice, P. H. (1975). Logic and conversation. In P. Cole & J. Morgan (Eds.), Syntax and Semantics, Vol. 3: Speech Acts (pp. 41–58). Academic Press.
Levinson, S. C. (1983). Pragmatics. Cambridge University Press.
Frank, M. C., & Goodman, N. D. (2012). Predicting pragmatic reasoning in language games. Science, 336(6084), 998. https://doi.org/10.1126/science.1218633
Bender, E. M., & Koller, A. (2020). Climbing towards NLU: On meaning, form, and understanding in the age of data. Proceedings of ACL 2020, 5185–5198. https://doi.org/10.18653/v1/2020.acl-main.463
Hong, E. (2026, March 5). The viral “car wash” LLM challenge doesn’t mean what you think it means. IBM Think. Vendor-authored commentary used only to document prompt variants and differing responses. https://www.ibm.com/think/news/viral-car-wash-llm-challenge
Changelog
- 2026-07-11: Corrected the prompt provenance from one reproducible video to a family of incompletely documented interactions; removed claims about an unknown model’s world access, training data, and internal probabilities; aligned the first two failure descriptions with their corrected companion articles; and bounded the Gricean and probabilistic-pragmatics analysis.