AI Made Accessibility Fast. It Still Needs a Human in the Loop.

Recently I asked an AI agent to clear the old connection requests sitting in my LinkedIn account. I told it, specifically, to leave the recent ones alone. It withdrew those too. LinkedIn makes you wait three weeks before you can invite the same person again, so I am now waiting three weeks to reconnect with potential partners who matter to my work.

I am a blind digital accessibility and accessible AI specialist in Sri Lanka. I use a screen reader every day, and I train people with disabilities and organisations to use AI. I am writing this because that small incident sits inside a much bigger story: AI has made accessibility fast, and speed without a human check creates new problems.

We waited a long time for accessible AI

For most of my working life, access meant waiting. A scanned PDF was a picture of text, and my screen reader could read none of it. An image without a description was a blank. A long report took me days to get through.

The wait was never a technical accident. Someone chose to publish the scan. Someone decided that an accessible version was an extra, to be produced later, if at all. Disabled people carried the cost of those decisions in lost time.

AI has changed the speed of all this:

• Scanned PDF to tagged PDF. A tagged PDF carries hidden labels for headings, lists and tables, so a screen reader can move through it. AI can now turn a scan into a tagged document in minutes.
• Image to image description. I can ask for a description of a photo or chart and get one immediately.
• Days of reading to seconds. AI can summarise a large document so I can decide which parts deserve a full read.

A social story for a seat belt

A Microsoft feature story, "A co-teacher for every classroom", describes Ashley Hernandez, a special education teacher in New York City. One of her students refused to wear a seat belt on the school bus, and nothing persuaded him.

Hernandez used AI to create a social story, which is a short visual guide that explains a situation and what to do in it. She added pictures chosen for him, sat beside him and read it aloud. He began fastening his seat belt on his own.

Now imagine how long it used to take to produce an Easy Read document. Easy Read is information written in short, clear sentences with supporting pictures, used by many people with intellectual disability. Having trained in Easy Read in Australia, I know the work one document takes. Rewriting a single leaflet or policy could take a specialist days, and most information was never converted at all.

The same speed that gave one student his social story can now give people with intellectual disability information they were always entitled to, when they need it. It still needs testing with the people it is written for.

One interface, and agents that act for me

Two more changes matter to me as a screen reader user.

The first is connectors. A connector links an AI assistant to my email, calendar and files, so I can ask one question in one place. I no longer have to learn the layout of five different applications, each with its own accessibility gaps.

The second is automation. AI agents now do more than answer. They carry out tasks on my behalf, on my own computer. For someone who has spent years working through inaccessible screens, that is real power.

What my LinkedIn requests taught me

This was not simply an unlucky click. My instruction was clear. The agent was very efficient. That was the problem. It completed the task quickly, including the part I had told it not to do, and the action could not be undone.

I was frustrated, because three weeks is a long time in a new partnership. But the lesson is one I now teach: a clear instruction is not a control. "Human in the loop" means a person reviews and approves before an AI system acts. It is not a sign of distrust in the technology. It is how responsible people use it.

My practice now:

1. Ask the agent to show me exactly what it plans to change before it changes anything.
2. Treat any action that cannot be undone as a step that needs my approval.
3. Start with a small batch, check it, then continue.
4. Check the result myself afterwards.

Fast access is still shared work

My experience is not universal. Others use different tools and face different barriers. Most AI tools cost money and work best in English. In Sri Lanka, that decides who gets the speed and who keeps waiting.

AI also does not release anyone from their responsibility. An AI description of a chart can be wrong. A summary can leave out the paragraph that matters. And a publisher who releases a scanned PDF has still made a choice to exclude.

So here is what I ask:

• Organisations: publish accessible documents at the source. Do not leave AI to repair them.
• AI developers: build confirmation steps before irreversible actions, and make those steps work with screen readers.
• Disabled people using AI: take the speed. Keep the final decision.

We waited a long time for access. Now that it is fast, we stay in the loop.