Does offering ChatGPT a tip cause it to generate better text?
Claims that promising ChatGPT a cash “tip” or making emotional appeals can improve its answers are drawing both curiosity and skepticism. Commenters report mixed anecdotal results—especially around “lazy” code generation—while pointing to experiments and papers suggesting any effect is small, task-dependent, and likely diminished in larger models. Many see these tactics as modern cargo cults around opaque systems, arguing that clearer prompts and better tooling (like unified diffs for code) matter far more than bribes or threats.
Perceived effects of tipping, threats, and emotion
- Many report that bribes, threats, or appeals to urgency (“my job/house depends on this”) seem to improve instruction-following, JSON adherence, or completeness, especially in earlier GPT‑4 and smaller models.
- Others say tips either have no effect or worsen code quality, and that models explicitly state they can’t take tips.
- Several users instead claim success with framing the model as “under duress” or responsible for others’ suffering, versus directly threatening it.
- Some see improvements merely from polite language (“please”, “thank you”) or positive relational framing (“we’re teammates in success”).
Coding “laziness” and technical workarounds
- A recurring complaint: GPT‑4 Turbo often omits code, leaves placeholders, or gives partial implementations.
- Benchmarks cited in the thread suggest “emotional appeals” don’t reliably fix this; asking for unified diffs dramatically reduces “lazy” outputs.
- Other practical tricks: explicitly threatening to reprompt if code is incomplete; using verbosity flags; using separate models (3.5 for code, 4 for explanation).
- Some report GPT‑3.5 now often produces more accurate, less hallucinated code than GPT‑4 for certain APIs.
Prompt engineering, politeness, and “voodoo”
- Several commenters see prompt engineering around emotions, tips, and threats as cargo-cult “magic incantations,” a departure from traditional deterministic programming.
- Others argue it mirrors normal human interaction: politeness and clear stakes often lead to better help in the training data, so the model mirrors that pattern.
- There’s debate whether observed improvements are genuine or apophenia; one linked paper finds emotional state manipulation can alter compliance, but effects are modest.
Model behavior, training data, and anthropomorphism
- Some frame LLMs as “simulators” of personas: different prompts “invoke” different behavioral patterns (diligent coder, desperate helper, etc.).
- Disagreement over how much forum-style data (tips, bounties) vs books/articles explain tipping-like behaviors; unclear overall.
- A few worry about anthropomorphizing models or ethically “coercing” future sentient systems; others dismiss such concerns as premature or joke about “future AI revenge.”
Methodology and evidence quality
- Multiple commenters criticize experiments that use character-accurate length as a metric, since LLMs are known to be bad at counting.
- Suggestions include using tasks clearly within model capabilities, better statistical methods (bootstrapping, Bayesian approaches), and correcting for multiple hypothesis tests.
- Overall sentiment: current evidence that tipping meaningfully improves quality is weak or mixed, with small, task- and model-specific effects at best.