When AI really helps you write better

The promise and the pitfall

Generative AI has transformed in two years how millions of people write. Promise: produce more, faster. Pitfall: produce more, but the same as everyone else, in a smooth, forgettable average.

For AI to actually help, you have to step out of the « write me an article about X » model and step into a logic of editorial collaboration. Here's how we think about it.

AI doesn't write alone, it prepares

In our approach, AI is never the final writer. It is an assistant that:

Synthesises the sources you've chosen
Proposes an article structure based on that synthesis
Suggests angles you might not have seen
Checks factual coherence between elements

The human writer then takes this material, prunes it, corrects it, gives it a voice. That's the difference between a draft and a publishable article.

A worked example: from brief to published piece

Imagine a trade publication covering agriculture. On Monday morning, the editor sees a topic gaining ground across several countries: fertiliser prices moving ahead of the spring season. She picks five sources for the story: two official price bulletins, a farmers' association statement, a market analysis and a regional news report.

From those sources, the assistant produces a first draft: a proposed structure, a summary of each source, two possible angles and a list of claims, each tied to the document it comes from.

Then the real work starts. The writer:

Checks every figure against its source and finds that one price was attributed to the wrong month. It is corrected.
Removes a paragraph that repeated the market analysis without adding anything.
Calls a local cooperative and adds a short quote from that conversation, clearly attributed.
Rewrites the opening so it says something the publication actually believes: that smaller farms will feel the change first.

The editor then reviews the piece against a checklist before approving it. The draft saved the writer the most mechanical part of the job. The value of the article comes from the call, the correction and the point of view.

Editorial voice cannot be delegated

A text generated without context sounds like a generated text: well-formed sentences, polished transitions, but no rough edges. No surprise. No point of view.

That's what makes raw AI so recognisable, and that's also why we don't try to remove the human. We try to give them more time to do what AI cannot: decide, take a stance, surprise.

Sources: the non-negotiable

An article that does not cite its sources is an article you cannot verify. AI that does not cite its sources is AI you cannot trust.

Our approach requires every factual claim to be anchored in an identifiable source, and that this source appears in the final article. This protects your credibility and that of your media.

As a side effect, it forces the AI to stay close to reality rather than producing plausible but invented claims — which remains its biggest flaw.

Verification stays human

Even with rigorous sourcing, AI can be wrong on details. Confuse two dates, attribute a quote to the wrong person, oversimplify a complex argument. No current system fully eliminates these errors.

The fix: integrate verification into the workflow as an explicit step, not a vague intention. If you don't know who reviewed the article before publication, it wasn't reviewed.

A verification checklist for AI-assisted drafts

Before approving any draft, whoever reviews it should be able to tick every line:

Every factual claim is linked to a source that actually says it.
Numbers match the source, including units, currency, period and rounding.
Quotes are verbatim, attributed to the right person, and used in their original context.
Names, titles and organisations are spelled and described correctly.
Relative dates ("last week", "recently") have been replaced with clear dates.
The article claims nothing its sources don't support.
The headline and summary match what the body actually says.

Instructions that get better drafts

The quality of an AI draft depends heavily on what you ask for. A few habits help:

State the angle and the audience. "An explainer for small-farm owners on what changes this spring" produces a very different draft from "an article about fertiliser".
Choose the sources yourself. A draft built on sources you trust is much easier to verify than one built on whatever turned up.
Ask for structure before prose. Review the outline first; fixing a structure is quicker than fixing finished paragraphs.
Ask it to flag uncertainty. A draft that marks the claims it isn't sure about saves you from hunting for them.
Say what to avoid. Banned phrases, sensitive topics, claims your legal team won't accept.

Keeping your voice

Write the opening and the conclusion yourself. That is where a point of view shows most.
Keep a short style guide with examples of sentences that sound like you, and a few that don't.
Cut generic transitions such as "moreover", "in conclusion" or "it is worth noting".
Read the piece aloud. Sentences no one would say usually shouldn't be printed either.

Form follows substance

Once the substance is solid, AI can also help on form: rephrase a heavy sentence, propose an alternative title, check tone. These tasks are well suited to an assistant and free up time for what requires judgment.

The common mistake is to start with form — « make this text more dynamic » — before substance is held. A beautiful sentence about a wrong idea is still a wrong idea.

What we want to avoid

Articles that all sound alike because they come out of the same mould.
Fake quotes or invented statistics.
The temptation of volume at the expense of value.
The author's voice erased under a layer of algorithmic neutrality.

What we want to make possible

Write twice as much, but better.
Cover topics you wouldn't have had time to address.
Keep editorial standards intact while reducing friction.

Well-used AI is not a shortcut, it's a lever. The difference shows over time.

Key takeaways

Use AI to prepare: synthesise sources, propose structure, suggest angles.
Every factual claim needs an identifiable source that appears in the article.
Make verification a named step with a checklist and an owner.
Fix the substance first, then use AI to help with form.
Keep the opening, the conclusion and the point of view in human hands.