TL;DR:
An assistant can describe what is visible in a picture: objects, layout, text, apparent mood. It cannot know what the picture does not show.
Who, where and when are usually inferences from clues, not facts.
Good answers separate what is seen from what is guessed, and say how sure they are.
You are studying history and a slideshow shows a black and white photo with no caption: a crowd outside a building, men in hats, a banner above the door. You ask the assistant, "What is this?" A good answer will describe it and then admit what it cannot say. A bad answer will name a city and a year with total confidence, and might be completely wrong.
The short answer
An assistant describing an image is working from the pixels, the same way you work from what your eyes receive, with no extra hidden knowledge about that particular picture. It can say what is in the frame. It can offer reasoned guesses about the setting. It cannot reliably say facts that are not visible, such as the date, the photographer or the names of people, unless text in the picture says so.
What happens step by step
Capture. The picture, or the part of your screen that holds it, is turned into an image the system can process.
Detect. Patterns of pixels are recognized as shapes and things: people, hats, a doorway, letters on a banner.
Read any text. Letters that appear in the image are recognized separately, which is why a clear sign is more useful than a face.
Relate and describe. The pieces are put into a description: what is where, what seems to be happening.
Infer, carefully. Clues such as clothing style, building shape or lettering style point toward a time and a place, with a margin of error.
Walk-through: a historical photo in a slideshow
Take our invented photo and split the claims into three groups, which is the honest way to answer.
Claim | Kind | How sure |
|---|---|---|
A crowd of people stands outside a building | Seen | High, it is plainly in the frame |
Most of the men are wearing hats | Seen | High, though those at the back are hard to count |
A banner hangs over the door; the readable letters spell part of a word | Seen (text) | Medium, depends on resolution |
The photo is probably from the first half of the twentieth century | Inferred | Low to medium; clothing and image quality are clues, not proof |
It was taken on a particular day in a particular city | Cannot be known | Not from this picture alone |
Who the people are, or why they gathered | Cannot be known | Needs a caption or a source |
Look at how the table works. The top rows are things you can point at. The middle row is a reasonable guess that comes with a warning. The bottom rows are the questions a photo cannot answer by itself. A good assistant writes its answer in the same order and tells you which part is which: "The photo shows... It appears to date from... I can't tell where or who from the image alone."
To tighten a guess, bring in outside evidence. A caption, the slide notes, the file name or the page the image came from can turn "probably" into "documented". Archives and museums that hold historic photographs publish the details, and those details beat any guess from pixels.
What can be seen, and what can't be known
Can be seen: objects, counts of obvious things, colors, layout, readable text, expressions and posture at a basic level.
Can be inferred, with care: era, setting, activity, mood.
Cannot be known from the image: dates and places without text, names, intentions, events before or after the moment, anything outside the frame.
Edited or generated images deserve extra caution. A convincing picture is not proof that the moment happened, and a description of it is only a description of what the picture shows.
Uncertainty, and how to ask for it
You can steer an assistant toward honesty with the question itself.
Ask "What can you see?" before "What is this?"
Ask it to label each claim as seen, inferred or unknown.
Ask what other information would settle the open questions.
Confidence is not accuracy: A fluent answer sounds sure whether it is right or not. The wording of the answer is a weaker signal than the evidence behind it.
What it can't do
It cannot confirm identity from a face, and a responsible tool should not try to name private people.
It can miss small details, count large crowds wrongly or misread faded text.
It cannot tell you what happened right after the shutter clicked.
Privacy and control
A picture on your screen may contain more than you intended: a name in a tab, a message notification, someone else's face. Before you share your screen, close what is not part of the question. A tool should look only when you ask: DuckyHelper sees your screen while the pill is open and for that answer, then stops. Its sees your screen page covers how that works in plain terms.
Try it yourself
Find any old photo with a caption hidden and ask the assistant to describe it. Then sort its sentences into seen, inferred and unknown, as in the table. Reveal the caption and see which guesses held up. You will build a feel for how much weight a description can carry.
For tables rather than pictures, see how assistants read tables and spreadsheets on screen.
Frequently asked questions
Can an assistant tell me who is in a photo?
Treat any name as a guess unless a caption or visible text says it. Naming people from their faces is unreliable and often inappropriate.
Why did it describe something that was not there?
Descriptions can include invented details when the picture is unclear. Ask it to point to where in the image it sees each thing.
Is a description enough for a school project?
Use it as a starting point. Cite the source of the photo, not the assistant's guess.
Get started
