Choosing an AI Draft: Read the Finalist Without Its Rival

A designer asks an AI assistant for two versions of a client email. One is short and direct; the other adds a warmer opening and a more graceful closing. With both on screen, the second seems much better. But when it arrives alone in the client’s inbox, the extra warmth may matter less than the sentence explaining which file needs approval. The comparison helped the designer choose. It did not reproduce the client’s experience.

That is a reason to give the finalist a different kind of review. Compare drafts when you need to see what changes. Then put the alternatives away and read the chosen version where it will be used. The question shifts from ‘Which is better?’ to ‘Does this do the job?’ Those questions can produce different answers without either review being wasted.

Why comparison changes the task

Christopher Hsee and Jiao Zhang explored this problem in research published in May 2004 on distinction bias. They argued that people choosing or predicting experiences with alternatives together can give weight to differences that matter less when someone experiences only the selected option. Their studies examined joint and separate evaluation. The research did not test AI-generated emails; applying the distinction to draft review is a practical proposal, not a finding about a particular assistant.

The point is not that differences between drafts are unimportant. A missing deadline, a wrong price or a changed instruction may matter enormously. The useful question is whether a difference has earned its weight because it changes the recipient’s experience or simply because it is easy to notice beside the alternative. Two endings may look quite different on your editing screen while both leave the reader equally clear about what to do next.

Give the draft a task before giving it a rival

For the designer’s email, the task could be: the client knows which attachment to review, what kind of approval is needed and when a reply is due. Use that as the initial review brief. A more polished opening is welcome if it helps the relationship, but it cannot compensate for a request that remains vague. If the two drafts both fail the brief, choose neither until the missing information is supplied.

An AI prompt can make the comparison more useful by asking for alternatives with meaningful differences. Instead of ‘Give me five better versions’, ask for one version that puts the action first and another that introduces the context first. Explain who will read it and what they need to decide. Then compare the trade-off you actually wanted to investigate. This is an editorial method to try; it does not guarantee that the generated versions will be accurate or suitable.

Read the finalist in its actual surroundings

Put the selected email into a preview with its subject line and attachment names. Hide the competing draft. Can you tell, from that single message, which version of the design is under review? Does ‘approve this’ mean approve the layout, the wording or the whole job for print? A rival can make a draft look clear by being less clear. The recipient will not have that rival to explain the distinction.

Our article on checking project shorthand in AI outputs offers a related check. If both drafts use a familiar label such as ‘quiet version’, the comparison might never draw your attention to it. Someone receiving the final message may still need to know what the label covers. Reviewing one draft alone brings the missing context back into view.

For a web page, use the published layout or a close preview rather than the chat window. For a handout, read the instructions with the actual example beside them. If the item leads to an action, follow that action. A button that sounded clear beside three alternatives can still send readers somewhere unexpected. Record what fails in the chosen version, then make the smallest change that fixes it.

Keep any factual corrections from the first review when you move to the second. Hiding the alternatives should not mean forgetting why you rejected them. If one version named the wrong delivery date, put the verified date in the brief and check it in the finalist. This separates two useful records: the evidence that a fact is right, and the judgement that one phrasing feels better beside another.

Use a short second review

• Keep a one sentence brief beside the chosen draft. Check the reader’s task first, including any facts or conditions the draft must preserve.
• Read only the finalist in context. Ask what someone who missed the editing discussion would need to know.
• If a difference between alternatives drove your choice, name the result you expect it to improve. Check that result rather than continuing to rank the wording.

A colleague can help by seeing the finalist without the options or your preferred answer. Ask them to tell you what they think they are being asked to do. If they identify the wrong attachment or scope of approval, you have something concrete to revise. If they understand the request but prefer a different greeting, decide whether another round of editing serves the purpose of the message.

Keep the side-by-side view for questions it answers well: what changed, which facts disappeared and which structure better fits the brief. Close it for the final check. The draft you send has to work without the one you rejected.

Source

• Hsee and Zhang (2004), Distinction bias: Misprediction and mischoice due to joint evaluation, Journal of Personality and Social Psychology, 86(5), 680–695.