AI is great, but it needs humans at the right time

At a time when enterprises (believe they) are expected to produce more and more content all the time and in many languages, they are desperately looking for help. And what better to turn to than AI when it comes to generating and translating content. Right?

Many therefore are adopting a new “efficient” workflow:

1. AI generates blogs, communications and other content.
2. AI translates this AI-generated content into multiple languages.
3. Human linguist is brought in to correct any mis-translations, and iron out inconsistencies and cultural faux-pas (so-called post-editing)

In principle, not such a bad idea. It certainly speeds things up, and it saves translation cost. It gets that content out there fast, to the audience who is eagerly awaiting more.

Yet, in reality, this creates a feedback loop of mediocrity that’s making everyone miserable.

Clients are unhappy with the end result, because their expectations are not met. Their in-country reviewers or marketing managers tell them that the material is not in line with brand guidelines, the tone of voice doesn’t meet the target group expectations, etc. And the post-editors are unhappy because they are aware that the time and budget they have been allocated is not sufficient to do a good job, and there is no opportunity to do a re-write where that is what is actually needed. They are frustrated also because the efforts they have put into creating client-specific termbases, and absorbing stylistic guidelines and client preferences, are not respected by the AI that generated the source and the target.

To obtain a satisfactory outcome, post-editing stops being ‘efficient’ and becomes unpaid ghost-writing – performed under time pressure.

We need to be aware that unless carefully guided and curated, AI-generated source plus AI-generated translation will exacerbate any of the problems that are so often mentioned in discussions about AI. These include blandness, inconsistencies, unnatural phrases and uneven style.

Raw AI-generated English tends to look surprisingly good at first glance. Sentences are grammatically correct, no spelling errors, the tone sounds vaguely “marketing”. And there are plenty of buzzwords to create the impression of a professionally written text by someone who knows their subject-matter.

But scratch beneath the surface, and you’ll find a linguistic nightmare:

• Overloaded sentences that undermine readability (By unifying fragmented data sources, tearing down the silos, automating repetitive tasks, accelerating decision-making across teams, and empowering the gathering of data that will enable deeply informed decision-making companies can now liberate their finance teams to do what they do best”; “As more and more organizations face ever increasing demands on content, shrinking budgets, rapidly evolving customer expectations and the never-ending need to prove ever-faster ROI, this innovative platform empowers you to extract information from gathered data, collaborate with colleagues across the globe, and create and deliver reports faster than you would have imagined.”)
• Bland, unexciting, formulaic texts that put the audience to sleep rather than excite them.
• Repetitive imperatives (you know what I mean: Enhance your performance, Improve your understanding, Level up your productivity, Boost your capacity, Unleash your creativity, Empower your finance team, …)
• Vague and overused jargon that is abstract and does not engage the reader
• Overuse of comparatives (smarter, faster, more efficient, more personalized
• Overuse of adjectives (best-of-class, world-class, seamless, scalable, industry-leading, next-generation)
• Desperate attempts to think of new metaphors and phrases (“your workflows glide across bottlenecks”; “AI becomes your lighthouse in the data fog”; “mobile outreach”, “grow your impact”, “supercharge your productivity”)
• Inconsistent tone and register (half-transcendental manifesto, half-product manual)

English readers may glide over these quirks. The English language has an uncanny ability to gloss over poorly developed ideas. It is more flexible and forgiving (almost frivolous) and can get away with ambiguity. Unfortunately, that’s not true of many other languages.

Translation amplifies every flaw in the source

That’s true whether a human is involved, or AI. Give AI a good, clear, neatly expressed and readable text, and it will produce a pretty good translation most of the time. Certainly if you have spent some time training it and giving it what it needs. But when the source is vague, uses idiosyncratic syntax, or is inconsistent or terminologically unstable, MT magnifies every flaw.

The result? German sentences that make native speakers wince. French copy that sounds like it was written by a confused robot. Spanish content that contains words but communicates nothing. In other words: raw AI English + unguided AI translation = a linguistic snowball rolling downhill.

Essentially, AI translation converts the AI-created original into stiff, unnatural target language, often in a convoluted way.

The Light post-editing myth

Light PE assumes several things that AI-to-AI workflows do not generally achieve: The source text must be coherent and make sense. Terminology is defined and consistent. Errors are surface-level. Meaning is clear enough not to require painful “interpretation” and guesswork.

When AI-generated English is fed into MT, one or more of the above a lacking. As a consequence, the post-editors face:

1. Semantic detective work

Half their time is spent deciphering what the original AI-generated English was actually trying to say.

2. Terminology digging

Without proper glossaries, every key term becomes a research project.

3. Tone reconstruction surgery

Fixing the machine’s accidental tone takes longer than writing from scratch.

4. Full paragraph transplants

Light PE? Hardly. This becomes full transcreation masquerading as “just a quick polish.”

5. Double trouble

Instead of saving time, the process adds an extra layer of work: repairing two texts instead of one.

Instead of fixing one text, they’re essentially fixing two: the confused source and its confused translation.

Without Prompting and Glossaries, AI Just Freestyles

No AI system magically knows:

• the client’s terminology,
• the preferred phrases or the no-no’s,
• tone of voice,
• product naming rules,
• idiosyncrasies like particular spelling, punctuation, etc.,
• handling of gender-neutral language,
• reviewer preferences.

You need glossaries, style guides and prompting as guard rails. Without them, AI imitates “generic, bland marketing language that lacks authenticity” — the exact opposite of what enterprise clients actually want.

AI without guidance cannot do a good job, not as a copywriter, nor as a translator.

---

Bottom line: Cheap inputs create expensive outputs

1. The AI-to-AI content pipeline isn’t reducing workload. It’s shifting it onto the very humans it was supposed to replace, while making their jobs infinitely harder.
2. If companies want real efficiency, they need to invest at the front of the pipeline. Proper prompting, clean terminology databases, and intelligent oversight by people who understand both the client’s needs and their audience’s expectations.
3. AI is incredibly powerful. But like every tool, it only works well when a human tells it what to do.
4. The choice is simple: Invest time and effort and money upfront, or pay much more to fix the mess later.