Last week, I came across various headlines about a study conducted by the New York Times, which found that readers preferred short stories generated by AI to those written by humans. Previous studies had found the same in relation to poetry and adverts.
Before you decide to delegate your writing to generative AI, though, you should consider that texts may serve different purposes. Some, such as speeches, are meant to connect emotionally with the reader; others, such as meeting notes, are meant to record information.
As Paul Roetzer argued, in episode 203 of The Artificial Intelligence Show podcast1, whether or not you use AI to support or take over your writing tasks, very much depends on what the text is meant to achieve. Text that requires emotional nuance, reflects your personal identity, voice or specific experience, or is of strategic importance to you, should be written by the human, with no or minimal use of AI. Conversely, text that follows standardised formats, is meant to be produced quickly and at high volume, and the value comes from the information not the style or the intent, might gain from being written with or even by AI. This reasoning mirrors research findings that people downgrade creative writing when it was produced by AI, but not necessarily when it is advertising text.
Thus, when considering the use of generative AI in writing tasks, we should first ask ourselves:
- In this piece of writing, does the value come from my personal identity / voice / judgement?
=> Focus on authenticity
2. Does this text need to be produced quickly, consistently and at scale?
=> Focus on efficiency
Roetzer proposed a scale with 5 different levels of AI involvement, depending on whether readers would value authenticity or efficiency the most. It may be summarised as:
- Text that requires high authenticity and there is low efficiency pressure, puts us in levels 0 or 1, with human led writing.
- Text that requires a balance of authenticity and efficiency, puts us in level 2, where human and AI collaboration is justified.
- Text that does not demand authenticity and requires high efficiency, puts us in levels 3 or 4, where it may make sense to lean on AI for writing.The table below summarises the key characteristics of xxx’s scale, but
The table below summarises the key characteristics of the scale, but I encourage you to read the original post, here, for more detail, including examples.
| Level | Writing | Priority | Characteristics |
| 0 | All human | Authenticity | • Authentic, original and nuanced • Reflects personal perspective • Emotionally resonant • Copyrightable |
| 1 | Mostly human | Voice matters, but AI can support selectively | • Originality and perspective are important • High degree of control needed |
| 2 | Half & Half | Both quality and efficiency matter | • Fast production of first drafts following specific formulae • Human refines, validates and adds nuance |
| 3 | Mostly machine | Consistency and scale | • Standardised, repetitive, informational • High-volume outputs • Limited originality or voice |
| 4 | All machine | Efficiency | • Automated and data-driven content • High volume, with low differentiation • Objective and transactional content • No need for personal voice |
To make this framework practical, you can think of it as a simple decision tool. Before deciding how to produce a piece of writing, consider whether its value comes primarily from authenticity or efficiency. The balance between these two dimensions determines the appropriate level of AI involvement, which can be summarised as:
- Use human writing to express what is distinctive.
- Use AI to scale what is standard.
Where are you finding that AI helps in your writing?
- On a side note, this is one of my favourite podcasts to get up to date on AI developments that are relevant for businesses, particularly marketing, consulting and market research. I really like their emphasis on using the productivity gains (if any) to upgrade what the organisation does, rather than do the same with fewer people. ↩︎
Ana Canhoto 