· updated
When Constrained Generation Beats Open Generation
Lessons from shipping AI in a multilingual newsroom and running an eval-gated content pipeline: constrain the failure surface, not the model's talent.
The first instinct with a capable model is to let it write. The systems I’ve shipped that actually survived contact with production all did the opposite: they shrank what the model was allowed to produce, on purpose, and got faster and more trusted because of it. Constrained generation isn’t a compromise you make because models are weak. It’s a design position you take because failures have owners.
If you’re scoping a generation feature, the decision this essay helps with is how much freedom to give the model, and where the human sits relative to its output. I learned the answer twice, in two very different settings: rebuilding an editorial CMS with AI workflows for a news network publishing across eight languages, and later running my own eval-gated content operation solo. The three questions at the end are the framework I use now.
What we constrained in the newsroom, and why
A news brand can’t publish “mostly right.” Every story carries the masthead, and in a newsroom the unit of accountability is an editor, not a model. So when we added AI to the editorial CMS, nothing generated free-form into the publish path. Every capability was bounded on three axes:
Bounded input. Quick drafts weren’t open composition. They were grounded in existing source content from partner properties. The model’s job was transformation with a reference, not invention. Same for cross-language translation: a story that already passed editorial judgment in one language became the ground for drafts in others.
Bounded output. The most valuable AI features weren’t articles at all. They were fields: suggested titles, slugs, tags, SEO metadata, social creatives. A title suggestion can be wrong at almost no cost: the editor sees five options, picks one or types their own. The blast radius of a bad generation was designed down to a single, reviewable, replaceable unit.
Bounded decision. The publish decision never moved. The editor of each language approved every AI-assisted draft before it went out. That wasn’t a temporary training-wheels phase; it was the permanent architecture, and it’s why editors adopted the tools desk by desk instead of fighting them. Nothing threatening lived in the workflow, because the workflow couldn’t publish without them.
The result of all that constraint wasn’t slower output. Editorial output tripled. Constraint is what made the speed safe enough to use.
The same lesson, relearned solo
Years later, running my own automotive content property with no editors and no safety net, I rediscovered every one of those boundaries the hard way. Open generation produced confident nonsense (plausible prices and specs that were simply wrong) and template-smell pages that search engines rightly discount.
What fixed it was the same three-axis constraint, rebuilt in pipeline form. Drafting is grounded in a research file collected from verifiable sources, and live prices come from data pipelines rather than the model (bounded input). Deterministic QA gates enforce structure and phrasing and block publication on failure (bounded output). Publishing stays a human action, mine, with post-publication verification behind it (bounded decision).
When open generation is the right call
Constraint has a cost, and sometimes it’s the wrong trade. Open generation wins when three things are true at once: the output is a draft by definition (a human will rework it anyway), a wrong output costs seconds rather than trust, and creative range is the actual product. For brainstorming directions, exploring angles, or first-pass prose a writer will own, constraint buys you nothing but friction.
The failure mode is inheritance: teams prototype in open mode because it demos beautifully, then ship the prototype’s architecture into a context where errors have real owners. The demo never shows you the blast radius.
A decision framework for PMs
Three questions decide how much constraint a generation feature needs:
- What’s the blast radius of one bad output? A wrong title suggestion dies in a dropdown. A wrong published fact dies in public, wearing your brand.
- Can correctness be checked by code, by grounding, or only by a human? Push everything checkable into gates and grounding. Whatever only a human can catch defines where your human belongs.
- What are your reviewer economics? Constraint is what makes review cheap. An editor can approve five field suggestions in seconds; they cannot fact-check a free-written article in seconds. If review is too expensive, you didn’t constrain the output enough.
The pattern underneath all of it: don’t constrain the model’s talent; constrain the surface where its failures can land. Give it rich input and full creative range inside a boundary, and make the boundary the product decision you defend. If you’re the PM, that boundary is yours to draw and to justify, because when an output goes wrong in public, someone will ask why the system was allowed to produce it.
The newsroom outcomes above are from my time as Senior PM at Asianet News (2022–2025); the pipeline mechanics are documented in the motomotar case study.