“Generative AI engineer” is the least informative title in this part of the market. It says somebody makes a machine produce something new, and nothing about whether that something is a paragraph, a product photograph, a voiceover or a fifteen-second clip — four crafts with four different sets of scars.

Almost every disappointing hire here traces back to that ambiguity: the employer meant one modality, the candidate had shipped another, and both nodded through an interview held in the abstract vocabulary the label invites.

Say which modality before you write anything else

Decide, in one sentence, what your system emits. Marketing copy. Product imagery. Synthetic speech. Short video. Then put it in the advert’s first line and in the screening call’s first two minutes. Most adverts we index for this role describe a general capability, then interview against a specific need.

It matters because strength in one modality is weak evidence of strength in another. Someone who spent two years making a text feature reliable is expert in instruction-following, tone and refusal behaviour. They may never have considered aspect ratios, colour consistency, or the fact that one bad frame ruins thirty good ones. The reverse holds just as firmly.

Skills that survive a change of modality, and one that does not

Some things carry across, and it is worth crediting a candidate for the right experience.

  • The generate, review, regenerate loop. Anybody who has shipped this work built a workflow where a human accepts or rejects, and that workflow is most of the product.
  • Working with non-determinism. Knowing the same input gives a different output, and designing support, testing and rollback around that.
  • Provenance discipline. Recording which prompt, model version and reference material produced which artefact, so a complaint can be traced.

What does not transfer is the part you probably care about most: knowing what good looks like. An engineer learns a new library in a fortnight; learning to see why a generated image looks cheap takes considerably longer, and some never do.

The list of things your system must never produce

This is where I would spend the most interview time; it is the risk that reaches your board.

Write the never-list before you hire. Competitor names. Real people. Untrue claims about your product. Medical, legal or financial advice. Prices. Anything resembling a guarantee. Imagery implying an endorsement you do not have. It is short, specific to your business, and producing it is your job, not the engineer’s.

The engineer’s job is to make it enforced rather than hoped for. Ask how, and listen for mechanism rather than intent. Good answers involve layers: constraints before generation, a classifier or rule pass over the output afterwards, a hard blocklist for names and trademarks, a human gate on anything published unreviewed, and a measured escape rate. A candidate whose entire answer is “we put it in the system prompt” has not operated one of these at scale.

Then ask the awkward complement: what did their system produce that somebody had to apologise for? Everyone who has shipped this work has a story. Ask about over-strictness too — eager filtering that blocks legitimate output is the failure nobody reports, and it quietly makes the tool useless to the team it was built for.

Search for the work, not for the label

Across a crawl of roughly 1.1 million UK careers pages, eight employers name this role in a job title. One hundred and fifty-four list “Generative AI” as a skill. That ratio is the most useful number on this page.

The discipline exists almost entirely as a capability attached to other jobs. The people you want work under titles like LLM engineer, machine learning engineer, creative technologist, or senior developer on the team that shipped the feature. Search the title and you get a handful of results, a fair share of them people who adopted the label because it was fashionable. Source on artefacts instead: a shipped feature you can use, a repository with an asset or prompt pipeline, a portfolio, a talk about a generation system that went wrong. Where the work is text, the LLM engineer pool has more depth and a better established label.

The same logic governs your advert: the generic title brings a large pile weighted towards enthusiasm, while the specific product and modality brings fewer and better applications.

Quality when quality is a matter of taste

For most engineering roles, correctness settles arguments. Here it often cannot, and candidates who have only worked where an answer is right or wrong struggle.

  • Who decided it was good enough, and how did you know before shipping? Listen for a named reviewer, a rubric, paired comparisons — anything but personal satisfaction.
  • What was your first-pass acceptance rate? A real number, even a poor one, means somebody was counting. “Most of it was fine” does not.
  • How did you keep a set consistent? One good output is a demonstration. Two hundred sharing a tone, a palette or a voice is the product, and that is the hard part.
  • What did you do when a stakeholder said it looked generated? Strong answers encode house style as reference material or examples rather than arguing about aesthetics.

Treat portfolios with suspicion — you are shown the survivors. Ask what the rejected twenty looked like. For measuring output systematically, the model evaluation engineer guide covers ground this role only touches.

Budget, and the missing benchmark

Indicatively £70,000 to £110,000 permanent, reaching £130,000 for somebody who has owned a public-facing generative product end to end. Day rates indicatively £500 to £700, higher where brand exposure is significant.

With so few titled adverts there is no reliable benchmark to point at in a negotiation, which cuts both ways. In practice you pay an experienced engineer’s rate plus a premium for demonstrable taste and shipped evidence. Be reluctant to pay that premium for prompt fluency alone, which is now widely distributed and rarely the constraint.

When you need two people rather than one

If your product spans text and imagery, or imagery and video, one hire will be strong in one and adequate in the rest. Rather than hunting for somebody excellent at all of it, split the problem: one engineer owns the platform — generation, storage, review workflow, safety controls, cost — and modality depth comes in alongside, on contract if the need is seasonal.

Where the visual side is the substance rather than a feature, the diffusion model engineer guide describes a different and scarcer hire, and conflating the two is how teams end up advertising for six months.