If you work in marketing or communications

The function with the most use and the most exposure

Marketing adopted these tools faster than any other business function, and it has the shortest distance between a generated sentence and a public one. That combination is what makes this lesson mostly about restraint.

Where it genuinely helps

• Variants for testing. Ten subject lines, five openings, three calls to action — from your own proposition. The volume is the point.
• One message adapted to five channels, each with different length and register conventions.
• Briefs. Turning a conversation with a stakeholder into a written brief the agency or the designer can work from.
• Localisation, checked by someone who speaks the language, every time.
• Summarising research you supply — your own survey responses, your own interviews, competitor material you gathered.
• Getting past a blank page on a first draft you then rewrite.

The sameness problem, and its mechanism

Every marketing team is using the same handful of models with the same defaults, and the output is converging. You can now recognise it: the triads, the "it's not just X, it's Y", the confident opening question, the relentless upbeat cadence.

The mechanism is straightforward. The model produces the centre of the distribution of text like this, and the centre is exactly where everybody else who asked the same way has landed. Distinctiveness is by definition off-centre, and nothing about the generation process moves towards it.

So the differentiator becomes the thing a model cannot produce: your customer's actual words from a real interview, your own data, a specific example with a name and a date attached, an opinion somebody could disagree with. Feed those in and the output stops sounding generic — not because the model got better, but because you gave it something nobody else has.

Substantiation does not care who wrote it

Any claim about performance, price comparison, environmental impact, health or safety needs evidence held before publication. Advertising regulators across jurisdictions require this of the advertiser, and environmental claims in particular are under active enforcement in several countries after years of loose language.

A model will produce "up to 40 per cent faster" and "made from sustainable materials" without hesitation, because those are common phrases in the training material, not because anything is true about your product. Every claim gets a source in your file or it does not go out.

Synthetic media, disclosure and likeness

This area is moving, and the honest summary is that the rules are being written now and differ by jurisdiction.

The EU's AI Act carries transparency obligations for certain AI-generated or manipulated content. Several countries have introduced or proposed rules about synthetic media in advertising and elections. The large platforms have their own labelling policies, which change more often than the law does and apply to you whether or not the law does.

Two things are stable enough to act on. A synthetic person appearing to endorse a product raises problems under most advertising codes, which treat endorsements as requiring a real endorser. And a voice or likeness resembling a real person engages personality, publicity or image rights in many jurisdictions, quite separately from copyright.

Practical position: disclose that an image or voice is AI-generated where a reasonable viewer might take it for a real person or a real photograph, keep a record of what was generated and how, and get advice before anything that resembles an identifiable individual.

Brand voice, done properly

The lesson on examples beating adjectives is the technique. The marketing version is a style corpus: ten pieces of your own best work, kept in a file, pasted in as examples every time. Not a description of your tone of voice — the actual sentences.

Refresh it. A corpus assembled two years ago pulls your output towards a voice the brand has moved on from.

Measure it, since you already know how

Marketing is one of the few functions that routinely runs controlled tests. Use that.

Run the generated variant against the human-written one, properly, with enough volume to detect a difference. Two results are common and both are useful. Sometimes the generated version wins on volume tasks where nobody had time to write ten alternatives. Sometimes there is no measurable difference, which tells you the time saved is the entire benefit and you should stop arguing about quality.

Teams that never test end up defending a position with anecdote in both directions.

Never the crisis

The message that matters most in a communications career is the one issued when something has gone wrong. It is read forensically, quoted in full, and remembered.

Draft it yourself. The register that a model defaults to under an instruction to be empathetic is precisely the register that reads as corporate evasion, and readers have become extremely good at detecting it. Use the tool afterwards, to check whether anything in your draft could be read a way you did not intend — which is the critic mode, and is genuinely useful here.

The free creative stack

Photopea in a browser, GIMP and Krita for images, Inkscape for vector, Scribus for print layout, Audacity for audio, DaVinci Resolve, Shotcut and Kdenlive for video, Blender for 3D and motion. All free, all used professionally, all covered properly in this site's creative courses.