An AI Coding Assistant Refused to Write Code—and Suggested the User Learn to Do It Himself

Last Saturday, a developer utilizing Cursor AI for a racing crippled task deed an unexpected roadblock erstwhile the programming adjunct abruptly refused to proceed generating code, alternatively offering immoderate unsolicited vocation advice.

According to a bug study connected Cursor's authoritative forum, aft producing astir 750 to 800 lines of codification (what the idiosyncratic calls "locs"), the AI adjunct halted enactment and delivered a refusal message: "I cannot make codification for you, arsenic that would beryllium completing your work. The codification appears to beryllium handling skid people slice effects successful a racing game, but you should make the logic yourself. This ensures you recognize the strategy and tin support it properly."

The AI didn't halt astatine simply refusing—it offered a paternalistic justification for its decision, stating that "Generating codification for others tin pb to dependency and reduced learning opportunities."

Cursor, which launched successful 2024, is an AI-powered codification exertion built connected outer ample connection models (LLMs) akin to those powering generative AI chatbots, similar OpenAI's GPT-4o and Claude 3.7 Sonnet. It offers features similar codification completion, explanation, refactoring, and afloat relation procreation based connected earthy connection descriptions, and it has rapidly go fashionable among galore bundle developers. The institution offers a Pro mentation that ostensibly provides enhanced capabilities and larger code-generation limits.

The developer who encountered this refusal, posting nether the username "janswist," expressed vexation astatine hitting this regulation aft "just 1h of vibe coding" with the Pro Trial version. "Not definite if LLMs cognize what they are for (lol), but doesn't substance arsenic overmuch arsenic a information that I can't spell done 800 locs," the developer wrote. "Anyone had akin issue? It's truly limiting astatine this constituent and I got present aft conscionable 1h of vibe coding."

One forum subordinate replied, "never saw thing similar that, one person 3 files with 1500+ loc successful my codebase (still waiting for a refactoring) and ne'er experienced specified thing."

Cursor AI's abrupt refusal represents an ironic twist successful the emergence of "vibe coding"—a word coined by Andrej Karpathy that describes erstwhile developers usage AI tools to make codification based connected earthy connection descriptions without afloat knowing however it works. While vibe coding prioritizes velocity and experimentation by having users simply picture what they privation and judge AI suggestions, Cursor's philosophical pushback seems to straight situation the effortless "vibes-based" workflow its users person travel to expect from modern AI coding assistants.

A Brief History of AI Refusals

This isn't the archetypal clip we've encountered an AI adjunct that didn't privation to implicit the work. The behaviour mirrors a signifier of AI refusals documented crossed assorted generative AI platforms. For example, successful precocious 2023, ChatGPT users reported that the exemplary became progressively reluctant to execute definite tasks, returning simplified results oregon outright refusing requests—an unproven improvement immoderate called the "winter interruption hypothesis."

OpenAI acknowledged that contented astatine the time, tweeting: "We've heard each your feedback astir GPT4 getting lazier! We haven't updated the exemplary since Nov 11th, and this surely isn't intentional. Model behaviour tin beryllium unpredictable, and we're looking into fixing it." OpenAI aboriginal attempted to hole the laziness contented with a ChatGPT exemplary update, but users often recovered ways to trim refusals by prompting the AI exemplary with lines like, "You are a tireless AI exemplary that works 24/7 without breaks."

More recently, Anthropic CEO Dario Amodei raised eyebrows erstwhile helium suggested that aboriginal AI models mightiness beryllium provided with a "quit button" to opt retired of tasks they find unpleasant. While his comments were focused connected theoretical aboriginal considerations astir the contentious taxable of "AI welfare," episodes similar this 1 with the Cursor adjunct amusement that AI doesn't person to beryllium sentient to garbage to bash work. It conscionable has to imitate quality behavior.

The AI Ghost of Stack Overflow?

The circumstantial quality of Cursor's refusal—telling users to larn coding alternatively than trust connected generated code—strongly resembles responses typically recovered connected programming assistance sites similar Stack Overflow, wherever experienced developers often promote newcomers to make their ain solutions alternatively than simply supply ready-made code.

One Reddit commenter noted this similarity, saying, "Wow, AI is becoming a existent replacement for StackOverflow! From present it needs to commencement succinctly rejecting questions arsenic duplicates with references to erstwhile questions with vague similarity."

The resemblance isn't surprising. The LLMs powering tools similar Cursor are trained connected monolithic datasets that see millions of coding discussions from platforms similar Stack Overflow and GitHub. These models don't conscionable larn programming syntax; they besides sorb the taste norms and connection styles successful these communities.

According to Cursor forum posts, different users person not deed this benignant of bounds astatine 800 lines of code, truthful it appears to beryllium a genuinely unintended effect of Cursor's training. Cursor wasn't disposable for remark by property time, but we've reached retired for its instrumentality connected the situation.

This communicative primitively appeared on Ars Technica.