Describing what you cannot see

Point a camera at something and ask what it is. For blind and low-vision users this has genuinely changed what can be done alone — reading post, checking whether the milk has turned, finding the right door, knowing what is in a photograph a friend sent.

It is also the use in this module where being confidently wrong does the most damage, because there is no glance available to catch it.

Ask for what you need, not for a description

“Describe this image” gets you a paragraph of scene-setting, most of which is not what you wanted.

Ask the actual question:

What does the sign say, word for word? Is there anyone in this room? What is the expiry date on this packet? Which of these two boxes is the blue one? Is the cooker ring on?

Specific questions get shorter, more accurate answers, and they let you tell when the tool has not understood — a non-answer to a specific question is obvious in a way that a vague paragraph is not.

Ask for a general description when you actually want one — a photo of family, a view — where there is no single fact you are after.

Ask it to say when it cannot tell

The most valuable instruction in this lesson:

If you cannot tell, say so. Do not guess.

Without it you get a plausible answer to everything. With it you get “the image is too blurred to read the third line”, which is genuinely useful — you can retake the photo.

Follow it up when precision matters:

How confident are you about the number? Would a clearer photo help?

The failure modes

Know these specifically, because you cannot spot them by looking.

Misread digits. Numbers are where this most often goes wrong — expiry dates, doses, reference numbers, prices, house numbers. A 3 read as an 8 arrives with exactly the same confidence as a correct reading.

Invented plausible detail. Asked what is on a table, a model may add things that are usually on tables. The description reads naturally and contains something that is not there.

Text that is nearly right. Long text gets transcribed with small substitutions — a word swapped for a similar one, a line skipped. Fine for gist, not fine for instructions.

Stale scene. The photo is a moment. “The door is closed” was true when the shutter fired.

Confident answers about people. Age, mood, relationship, whether someone looks unwell — treat all of these as guesses, because that is what they are.

Where it must not be the only source

Draw this line before you need it:

Medication. Never identify a tablet or a dose from a photo alone. Use a pharmacist, a labelling system, or a dedicated medication-identification service.

Anything you are signing. Get a human or an official accessible copy. Organisations are usually obliged to provide one, and asking costs a phone call.

Money. Note denominations, card details, meter readings — verify by another route.

Safety. Whether an appliance is off, whether a substance is what the bottle says, whether a road is clear.

This is not because the tool is bad. It is because “I could not check” is different from “I checked”, and these are the cases where that difference is expensive.

Use the purpose-built ones too

General assistants are convenient. Dedicated tools exist for this exact job — some with the option to connect to a sighted human volunteer or a trained agent when the machine cannot tell.

That combination is the practical answer for anything important: try the fast automated route, escalate to a person when it matters. There is no prize for refusing help.

Try it now (7 minutes)

Photograph something with small print — a packet, a bill, a label.

Ask a specific question about a number on it, adding say so if you cannot tell. Then check the answer against the real thing, or ask someone. That comparison tells you how much to trust this for your own use, which is worth more than any general claim.

Check your understanding

1. Instead of 'describe this image', you should usually ask: For a longer, more detailed description The specific question you actually need answered For a description in bullet points

2. The most valuable instruction to add when using a camera this way: Be concise If you cannot tell, say so — do not guess Describe colours precisely

3. Which must never rest on an image description alone? Identifying a tablet or its dose Finding out what a poster is advertising Checking whether a photo is in focus

Recap

Ask the specific question rather than for a description, and always add “say so if you cannot tell”. Know that digits, invented detail and near-right transcription are the failure modes, and that none announce themselves. For medication, signatures, money and safety, use a second route — and use the purpose-built tools, including the ones that can reach a human.