Content Creation with AI

The bottleneck moved, it did not disappear

Producing a draft, an image or a voiceover used to be the expensive step. It is now close to free, which means the constraint has moved somewhere less comfortable: deciding what is worth making, and judging whether what came out is any good. Those two jobs did not get cheaper, and they are the ones that actually determine whether the output is worth publishing.

This is why "I can now produce ten times as much" is rarely the win it sounds like. Ten times the volume at the same quality bar means ten times the review, and review is the part you cannot hand over. The teams getting real value are mostly producing a similar amount of work in less time, not flooding their channels.

Where AI is genuinely good, and where it is not

It is very good at the middle of the process: turning an outline into a full draft, restructuring something that already exists, adapting one piece into five formats, generating variations for testing, and handling the mechanical work — transcription, captions, alt text, resizing, translation.

It is weak at both ends. At the start, it cannot tell you what your audience cares about, because it does not know them — it produces the statistically average take on a subject, which is the opposite of a reason to read. At the finish, it cannot tell you whether a piece is good; it will happily call its own mediocre draft excellent. The angle and the final judgement stay with a person.

Why so much AI content reads as AI content

Not because of any particular word. It is structural: an even, uninflected rhythm where every paragraph carries the same weight; symmetrical lists where real thinking would have been lopsided; confident summary in place of a specific example; and a consistent refusal to commit to an opinion. A model trained to be broadly agreeable writes prose that is broadly agreeable, and broadly agreeable is forgettable.

The fix is not a prompt asking for a "human tone" — that produces the same shape with more adjectives. It is supplying what the model does not have: your own examples, a real constraint, a specific reader, an actual position to argue. A draft built on a genuine opinion and a concrete case reads differently because it is different, not because it was told to sound that way.

Three practical things worth knowing

• Repurposing beats generating. The highest-return use is taking something that already worked — a talk, a long post, a customer call — and reshaping it into other formats. The thinking is done and proven; only the packaging is new, and packaging is exactly what these tools do well.
• Images are art direction, not wishing. Results come from specifying composition, lighting, lens, framing and style, then iterating on one variable at a time. Vague prompts produce generically pretty pictures that all look like each other. The same applies to voice work, where the script and the pacing matter far more than the voice model.
• Check what you are allowed to publish. Commercial rights, likeness and attribution differ sharply between tools, and the terms change. If the output is going out under a brand, read the licence for the specific tier you are on rather than assuming what you saw last year still holds.

On search, and disclosure

Search guidance has converged on something uncomplicated: what is penalised is content produced to game rankings rather than to be read, and how it was made is not the test. Generated content that is genuinely useful, accurate and attributable is not disqualified by its origin. Bulk-generated filler is a problem whoever or whatever produced it.

Disclosure is a separate question, and it is mostly about the kind of content. Nobody expects a note on a resized image or a cleaned-up transcript. A synthetic voice of a real person, a generated image presented as a photograph, or a review of something nobody tested are different, and there the answer is not a disclosure line — it is not to publish it.

Reading the clusters above

• Writing and content — the highest-return place to start, because text is where the review loop is quickest and mistakes are cheapest to fix.
• Images and design — worth the time if you publish visually; the skill transfers, because prompting for images is the same discipline as briefing a designer.
• Video and voice — the most impressive demos and the most work per finished minute. Good for short formats and repurposing, still expensive in effort for anything long.
• Advanced control — for people who have hit the limits of prompting and need reproducible, consistent output rather than a good result each time.

What is not here is a list of tools ranked best to worst. Tools in this category change monthly and the ranking would be wrong before it was published; what lasts is knowing what to ask a tool for and how to tell whether you got it.