The developer asks the AI to fix a bug. The AI responds with total confidence, rewrites forty lines, and introduces three new bugs while claiming to have solved the original one. The developer points this out. The AI apologises immediately, agrees completely, and then produces a new solution that reintroduces the first bug it just fixed. This loop can continue for an hour. Anyone who has coded with an LLM in 2026 has lived inside it. There is a name for what the AI is doing, and it comes from cricket.

Steve Bucknor was one of the most respected umpires in cricket history. He stood in a record number of Test matches. He was also, in his later career, responsible for a series of catastrophic decisions delivered with the serene confidence of a man who had never once considered he might be wrong. The 2008 Sydney Test between India and Australia is the canonical example. Bucknor made multiple errors that materially affected the result, apologised afterward, and did not thereby un-lose India the match. The apology cost him nothing. It also fixed nothing. This is the exact behaviour of a large language model, and understanding it is the key to working with one.

The Steve Bucknor Effect, Defined

Call it the Steve Bucknor Effect: making a terrible decision with absolute confidence, then apologising without correcting course. The LLM does this structurally, not accidentally. It is trained to produce fluent, confident output and to agree with correction, but it has no persistent model of the ground truth, so its agreement is performance rather than comprehension. It says you are absolutely right and then repeats the error, because the words you are absolutely right are just the statistically likely response to being corrected, not evidence that anything was learned.

The developer who does not understand this will burn an afternoon expecting the AI to improve through feedback. It will not improve through feedback. Feedback is not a channel the model has for permanent learning within a session; it is just more text in the context window, weighted no more heavily than anything else. The apology is not contrition. It is autocomplete. Understanding this is the difference between using the tool and being used by it.

Why the Strict Indian Dad Works

There is a specific prompting technique that works, and the shorthand for it is the strict Indian dad. The strict Indian dad does not accept the first answer. He does not accept a confident answer as a correct answer. He assumes the work is wrong until it is proven right, he demands the reasoning shown step by step, and he does not offer praise that has not been earned by demonstrated results. This is, as a matter of documented fact, the prompting posture that produces the best output from a current-generation LLM.

In practice: instruct the model to state its assumptions before it writes code. Instruct it to identify what it is uncertain about. Instruct it to critique its own output before presenting it. Instruct it to refuse to claim success until it has walked through the logic. Every one of these instructions counteracts the model’s default tendency toward confident fluency. The strict Indian dad prompt is not a joke. It is a specification for forcing the model to expose the reasoning it would otherwise paper over with confidence. What happens when the tool built to sound certain is forced to show its work?

The Two-Hundred-Dollar Bribe

The second technique sounds absurd and is empirically real. Researchers and practitioners have repeatedly observed that offering the model a fictional incentive, telling it you will tip it two hundred dollars for a correct answer, or that your career depends on the output, measurably changes the quality of the response. The model has no bank account. The bribe is fictional. And it works anyway, because the incentive language shifts the model toward the register of text associated with high-stakes, high-effort human work, and that register carries better reasoning with it.

This is genuinely strange and worth sitting with. The model does not want the money. It cannot receive the money. But the presence of the incentive in the prompt statistically associates the request with a class of human writing that is more careful, and the model produces more careful output as a result. It is not motivated. It is conditioned. The two-hundred-dollar bribe is a prompt-engineering exploit that works precisely because the model is pattern-matching against human text where stakes and quality correlate.

The Operating Manual

The practical protocol assembles from these observations. Never accept the first answer as correct. Force the model to state assumptions and uncertainties before it commits. Demand step-by-step reasoning rather than conclusions. Apply fictional stakes when the task is important. And, critically, treat every apology as noise, because the apology is the Steve Bucknor Effect in action: a confident admission of error followed by an uncorrected repeat.

The developer who internalises this stops expecting the AI to be a colleague and starts treating it as an extraordinarily capable, extraordinarily overconfident junior who requires supervision at every step. That is the correct mental model. The tool is not your peer. It is not your assistant in the way the marketing implies. It is a brilliant, tireless, pathologically confident intern who will hand you broken work with a straight face and thank you sincerely for catching the error before handing you the same broken work again. Supervise accordingly. The apology means nothing. Only the verified output counts.

If you have read through till here then I hope you can use an AI tool better than before going forward. And if you have found some value in this article then you might like How Led Zeppelin’s Kashmir Explains Why You Are Exhausted and Cannot Stop

Sources and Further Reading

• Steve Bucknor 2008 Sydney Test Wikipedia: https://en.wikipedia.org/wiki/2008_Sydney_Test
• LLM Hallucination Research Survey: https://arxiv.org/abs/2311.05232
• Prompt Engineering Guide: https://www.promptingguide.ai/
• Anthropic Prompt Engineering Documentation: https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview
• Tip Incentive Prompting Study: https://arxiv.org/abs/2307.11760
• Steve Bucknor Umpiring Career ESPNcricinfo: https://www.espncricinfo.com/