QA for AI Automation: Checking Facts and Inspecting the Photos

AI can rebuild a website, work through a video archive, and produce a page that looks convincing. It can also put the wrong facts into that page and select photos that have no business being there. My website rebuild did both. I was proud of what Arya built from my YouTube videos, but reviewing the result exposed mistakes that a finished-looking page did not make obvious.

A convincing biography can still contain a false claim

One page said I had won all my matches at white belt. That was wrong: I lost many matches when I started. I corrected the wording to reflect those losses at 170 lb and the wins that came later at 185 lb.

The useful lesson here is about checking AI-written history. A model can turn a real collection of videos into an inaccurate summary. The sentence sounded like a clean account of my progress, but it changed the record. Seeing the right person, subject, and general story on a page is not enough to establish that each claim is true.

In this case, I knew the history and could catch the error. For an automation workflow, that suggests a concrete review step: pull out factual claims and compare them with the original material. Pay particular attention to absolute statements such as “all,” dates, numbers, and descriptions of achievements. If a source does not support the wording, revise the claim rather than filling the gap with a plausible story.

Photo selection requires looking at the actual images

I downloaded my Facebook data and asked my local model to flag photos that were just me. The resulting selection included pictures that were not just me. Arya flagged some of those, and I decided they were safe to use. Other choices were clearly wrong for the website: a couple of family photos and a picture of food.

Those are different decisions. An image can fail a literal “just me” filter and still be acceptable after I inspect it. A family photo can be a valid image from my archive and still be inappropriate for this page. A food photo can be correctly retrieved from the data while contributing nothing to the biography. Finding a file is only one part of selecting an asset.

The review needs the actual photo. A filename, a model's description, or a confidence score cannot tell me everything I need to decide whether it belongs on my public website. I need to see who is in the frame, what the image shows, and whether it helps explain the page. Human approval also needs to remain distinct from the model's initial classification.

Use photos as evidence without inventing their context

I added photos showing me at 170 lb and 185 lb, including one from later in my white-belt period. Looking at those different versions of myself helped me review how the page represented my history.

The image and its context do separate jobs. The photo shows what is visible; my account supplies details such as the weight and period. A workflow should keep those sources together instead of asking a model to guess a date, a result, or a personal story from appearance alone. Looking at the pixels is necessary, and checking the supporting information is necessary too.

Turn the review into a repeatable QA process

This rebuild suggests a practical checklist for reviewing similar automation work:

1. Check the claims. Compare the generated wording with the relevant video or other original source. Correct unsupported summaries.
2. Open the selected assets. Inspect the images themselves rather than accepting the selection list as proof.
3. Apply the page's purpose. Decide whether each photo is relevant and appropriate for public use, including who else appears in it.
4. Review the finished page. Read the corrected text alongside the chosen photos so the combined result tells the supported story.

That checklist is the workflow lesson from the review, rather than a claim that the original automation already performed every step. The automation did a great deal of useful work. The factual correction and the unsuitable photo choices show why QA needs to examine its output directly before treating that work as complete.

Source: original video, 1:03:05–1:07:25.