A checklist before any AI-generated news video goes out

I handle communications for a small organisation that publishes to a local audience, and I started generating video summaries of our written pieces about a year ago. I have also killed four of them before publication, which is the part worth writing about, because everything I have read on this subject is either a product pitch or a general warning that AI is dangerous, and neither helped me build a process.

What follows is the process I actually use. It is not elaborate. It exists because I got things wrong early and would rather someone else skipped that.

What actually goes wrong

Only one of these is the failure people expect.

Invented detail. This is the famous one and in my experience the least common, because I generate from a piece I have already written rather than from a prompt. The tool works from my document, so it does not invent facts. What it does do is generalise. A sentence saying the council approved the plan for the eastern section became narration saying the council approved the plan, which is a different and false statement. Summarisation drops qualifying clauses first, and in reporting the qualifier is frequently the accuracy.

Collapsed sequence. Written pieces establish chronology through structure. Narration is linear and has to commit to an order, and twice I have had a generated script imply that one thing happened because of another when the original text only said both happened. Causation appears without anyone adding it.

Borrowed certainty. This is the one I did not anticipate. My written pieces hedge properly, with attributions and words like reportedly and according to. A confident synthetic voice reading a hedged sentence somehow lands as less hedged. I do not fully understand the mechanism, but I have watched people take a stronger impression from the video than from the article, and that is on me rather than on them.

The four checks I do by hand

Every generated script, without exception, before anything renders.

One: every attribution survived. I search the script for the names and sources that appear in the original. If the article said a spokesperson stated something and the script says it happened, that is a rewrite, not a summary.

Two: every qualifier survived. I go through the original looking for where, when, if, among, reportedly, alleged, and check each one is still there. This is the check that catches the most.

Three: no new causation. I read the script asking whether it implies any because that the original did not contain.

Four: nothing is stated more strongly than the source. A generated script tends to tighten prose, and tightening removes hedges.

This takes about as long as the video runs. It is not optional and it does not get faster, and anyone who tells you the workflow is push button has not published anything they were accountable for.

Disclosure, and why I stopped debating it

For the first two months I put a line in the description. Now I put it in the video, on screen, in the first five seconds, and I say it in the narration.

The reason is a conversation with a reader who asked whether the presenter was a real person at our organisation. She had not been deceived exactly, since the information was accurate and the description said so. But she had spent forty seconds thinking she was being addressed by a colleague of mine, and when she found out she was not, her question was what else she had assumed wrongly.

That is the actual cost. Not a complaint, not a correction, just a small permanent reduction in how much someone trusts everything else you publish. The line costs nothing.

The wider standards conversation is worth following rather than reasoning about from scratch. AP's news values and principles set out how attribution and sourcing are supposed to work regardless of format, and Poynter publishes ongoing work on how newsrooms are handling synthetic media specifically.

What I use it for

The list has got shorter over the year rather than longer.

Visual summaries of pieces already published and checked. The article is the source and the video points at it. This is the bulk of what I do.

Structured information that does not change interpretation. Meeting times, service changes, opening hours, results. There is no judgement in this material and the risk of distortion is close to zero.

Explainers on things that keep coming up. How a particular process works, what a term means. These are written carefully once and reused, and they are the ones where the effort is most repaid.

Multilingual versions of the above. This is the biggest practical gain for us. I use Leadde.ai, which will translate a finished video and produce a separate version rather than overwriting it, handling the on screen text along with the narration. Our audience includes a substantial number of people who read a second language more comfortably than English, and until last year we simply did not serve them.

I have someone check every translation. Ordinary prose comes through well. Place names, official titles and anything legal do not, and an approximate translation of a court outcome is not a thing I am willing to publish.

What I never use it for

Anything unverified. If I am not certain, the answer is to not publish in any format, and the existence of a fast production route makes that discipline harder rather than easier.

Anything involving an accusation against a person. The wording in those pieces is chosen sentence by sentence, and a paraphrase is a different document with the same apparent authority.

Anything about a death, an injury or a disaster. Partly judgement about what the format conveys, partly that these are exactly the pieces where the reporting is most conditional and the summarisation risk highest.

Anything time critical. This is the one that surprises people, given the category name. My check takes fifteen minutes, and a genuinely breaking story does not have fifteen minutes. So breaking news is the thing I use this for least, and if I ever find myself skipping the check because something is urgent, that is the moment the process has failed.

Corrections after publication

Publish something wrong in text and you fix the text. Video is harder, and I did not have a plan for this the first time it happened.

What I do now: correct the source article first, regenerate the video from the corrected article, replace it at the same location, and add a correction note that is visible without playing anything. I keep the superseded version as a duplicate rather than deleting it, because someone may have acted on it and I want to be able to say exactly what it said.

The thing I got wrong initially was editing the video directly to patch the error. The article and the video then said slightly different things, and I had created a second problem while fixing the first.

Questions I get asked about this

Does it save time overall? Yes, but less than people assume. Generation is minutes. The check is roughly the runtime of the video. Against filming and editing it is a large saving, and against publishing the article alone it is an additional cost that has to be worth something.

Do the videos perform better than the articles? They reach different people rather than more people. The audience for a three minute summary overlaps less with our article readership than I expected.

What settings actually matter? The audience and objective fields, and keeping the length setting short. A generated script pitched at nobody in particular is the default outcome and the main reason first attempts feel generic.

Should captions be on? Always, from the first version. Most of this is watched with the sound off, and adding captions across a back catalogue later is much more work than turning them on at the start.

Is the technology the risk? No. The risk is that publishing becomes cheap enough that the checking step starts to feel disproportionate. That is a process problem and it has to be solved before the tooling arrives, not after.

How do you decide who appears as the presenter? We use a generated presenter rather than a likeness of anyone who works here, which sidesteps a whole category of problem. If you do want to use a real colleague's face, get their agreement in writing, be specific about what the avatar will be used to say, and agree what happens if they leave. A former employee still presenting your output is an awkward conversation nobody plans for.

Does any of this replace a reporter? It replaces a production step, not a reporting step. Everything I generate comes from something a person already researched, wrote and checked. If there were no article underneath, there would be nothing worth summarising, and the format would just be a confident voice reading an unverified claim.

What would make you stop using it entirely? If I found myself publishing something I had not fully checked because the video was already rendered and it felt wasteful not to use it. The sunk cost is small enough that this should never happen, and the day it does is the day the process is running me rather than the other way round.

Where I would start

If I were beginning now, I would take one already published explainer that gets referenced repeatedly, generate a version, and run the four checks slowly enough to notice what they catch.

Whatever the checks find on that first attempt is the thing to watch for permanently. In my case it was qualifiers, every time, and knowing that has been worth more than any setting I have adjusted since.