I Kept the Review Step

Today, the small safeguards beside AI caught my attention more than another promise that AI would do more work. I saw a label, a visible source trail, and a review step before a change could go live. Automation was moving forward, but all three stories had fitted a brake light.

I welcome that direction. Fast tools are not the problem. The problem begins when I cannot see who made something, what evidence it used, or where a person checked it. The machine answers in one second, then I spend ten minutes searching for whether I can trust the answer. The time I saved has quietly left through the back door.

AI content now needs a signpost

European Commission guidance says that some transparency obligations under Article 50 of the EU AI Act began applying on August 2. AI providers covered by the rules must let people know when they are interacting directly with AI. They must also add machine-readable markings to AI-generated or manipulated content.

Deployers have notice duties in several situations too. The guidance mentions deepfakes, AI-generated public-interest content published without human review or editorial control, and emotion-recognition or biometric-categorisation systems. This is not one identical label for every AI result; the scope depends on the role and use case.

A label is not a truth certificate. Putting a name sticker on leftovers does not make them fresh. It does, however, tell me what I am looking at. In my own work, I want a short, visible note when AI helped and when a person reviewed the result.

Newsroom AI brought the sources to the front

OpenAI Academy held an online session for newsrooms on August 4. Its description covered using AI for research, document and data analysis, interview preparation, coverage planning, and clear briefs with sources. It also described turning one-off prompts into reusable workflows with agents, skills, and apps.

This was an educational session description, not evidence that AI automatically produces good journalism. Still, the emphasis was useful: a repeatable and reviewable process matters more than one impressive answer.

When I prepare a daily briefing, I am tempted to polish the prompt first. I should build the source column first. Confirmed facts, figures I could not verify, and my own interpretation need separate places if I want to apply the same standard tomorrow. AI may write faster, but the checklist still arrives at work before the editor.

Automated setup still paused before the merge

GitHub announced an AI-generated code-coverage setup on August 4. Code coverage shows which parts of a codebase were actually exercised by tests. The new feature can prepare a workflow that builds the code, runs tests, produces a coverage report, and uploads the result to GitHub.

The step after generation mattered most to me. GitHub does not apply the workflow immediately. It opens a pull request so a person can inspect the change before merging it, and the generated workflow uses least-privilege permissions by default. The feature is currently a public preview for GitHub Code Quality users on github.com.

One click to create a setup is convenient. One click straight to production should be slightly less convenient. When a machine drafts the change and a person reviews the difference, speed and responsibility remain on the same screen.

I left three seats open

I can turn the three stories into three simple rules for my own work:

1. Make AI involvement recognisable.
2. Keep sources and verification status beside each claim.
3. Give a person a clear place to review and approve automated changes.

As automation grows, the human role does not simply disappear; it moves. I spend less time producing every piece by hand and more time setting standards and checking exceptions. I assigned work to AI again today. I also kept the final review step. My name belongs there because responsibility for the result belongs with me.