If you must use ChatGPT, prompt it like you wrote it

The default register of every major LLM is roughly the same. Business-formal-neutral. Hedged. Structurally complete sentences. Transitions like 'in conclusion' that no native speaker uses in informal writing. The model produces this register because it's a safe average of its training data, and the average is nobody's actual voice. When you send an email in the model's default register without editing, the recipient reads it as 'this person used AI', not as 'this person wrote me an email'. The difference looks small. It changes how the message is received.

The simple prompt fixes

The cheapest improvements come from giving the model specific stylistic instructions. 'Make it sound informal.' 'Reduce the polish.' 'Cut by 50%.' 'Sound like a human, not a corporate communication.' Each of these moves the output measurably away from the AI default. None of them reproduce your specific voice, but they close most of the distance between 'detectable AI' and 'plausibly written by a human in a hurry'.

The single highest-impact instruction is usually 'cut by 50%'. AI-drafted emails default to elaborate. Brevity is the easiest fix and the one most users skip. Three sentences usually replace seven without losing information. The recipient prefers three. The sender takes ten extra seconds to issue the cut instruction. Trade is overwhelmingly favourable.

The example trick

The next level up is providing examples. Three or four of your own past emails, pasted into the prompt as 'this is how I write', genuinely shifts the model toward your voice. Not perfectly. Closely enough that the recipient stops noticing the AI register. The investment is twenty minutes once. The benefit accrues every email you send for the rest of the year. Most people don't bother because the investment doesn't feel worth it until you've felt the cost of receiving enough off-voice emails yourself.

The structural reason this works is in-context learning. The model anchors strongly to recent examples in its prompt. With three high-quality samples of your voice, it produces output much closer to that voice than the default. With no samples, it has nothing to anchor on except its training average. The discipline is just remembering to include the samples consistently.

The read-aloud test

The simplest quality control I've found is reading drafted emails aloud before sending. If the second sentence makes you wince, the recipient will too. The model's failures are most audible when spoken - the over-formal transitions, the unnecessary hedges, the slight tonal awkwardness. Most AI emails fail this test on the second sentence. Editing for the read-aloud version produces dramatically better-received messages.

This takes about thirty seconds per email and catches most of the egregious failures. Skipping the test is how the bad versions get sent. The skip happens because the email feels finished - the AI produced a coherent block of text, the structural shape is right, the grammar is correct. All of these are true and irrelevant. The voice is wrong, and you can only hear the wrongness by reading the words out loud.

The model produces a register that is nobody's actual voice.

The cultural variable

The default LLM register is American business-formal. In other professional cultures it lands differently. French business correspondence has different rhythms; Japanese has different politeness markers; Indian English has its own conventions around formality. AI tools default to American norms because that's what they were trained on. For non-American professional contexts, the 'sound like yourself' work also includes 'sound like the cultural register expected of you', and the prompts have to address both.

This is rarely written into prompt-engineering guides, which are themselves usually written in American English. For anyone working across cultures, the meta-skill is noticing when the model's output is in the wrong cultural register entirely, and explicitly correcting for it. 'Use British English, formal but warm, no Americanisms' is a useful baseline instruction for many UK contexts. Equivalent baselines exist for other cultures and they save a lot of editing.

The simpler alternative

There is, of course, a simpler version. Write the email yourself. Use AI as an editor for grammar, length, or tone - not as the author. The total time is about the same once you account for the cleanup of AI's drift away from your voice, and the output is reliably yours. For short emails (under a hundred words), this is almost always the right approach. For longer pieces of writing, AI as a drafting partner can save genuine time, but only if you commit to the editing afterwards.

The broader observation is that AI-assisted writing rewards discipline disproportionately. The people producing AI-written emails that you can't tell are AI-written are doing more work, not less, than the people producing detectable AI emails. The productivity gain from AI is real and overstated. The quality gain requires investment most people aren't making, which is why most AI emails arrive irritating and the discipline of producing better ones remains underdeveloped. The simple prompt instructions in the original post - sound informal, reduce polish, cut length - are the minimum viable practice. They're not enough on their own. They're a start.