The engines learned to see
Structure and markup refreshed for current answer engines; the original analysis is preserved.
This month ChatGPT gained eyes. It can now take an image you give it — a photo, a screenshot, a chart, a diagram — and read what is inside, including the text printed on it. For a brand, that means a whole layer of your work just stopped being invisible to machines. The response is unglamorous and worth doing anyway: put real text in your pictures, and keep your entity consistent across every mode the engine can now perceive.
the short answer
AI engines have started to see: this month’s vision capability in ChatGPT can read the words and numbers inside an image — a chart, a diagram, a screenshot — that machines used to treat as an opaque rectangle. So a layer of your work that was invisible to them is beginning to count. Two moves follow: put real, legible text where your visuals carry meaning instead of leaving it as decoration, and keep your entity consistent across modes so what the engine sees agrees with what it reads. It is early and imperfect — but the preparation is cheap and helps humans too.
key takeaways
- This month ChatGPT gained vision — it can now take an image and read what is inside it, including the text in a screenshot, the labels on a diagram, the numbers on a chart.
- For years machines treated your visuals as opaque rectangles. That is starting to change: a layer of your work that used to be invisible to them is beginning to count.
- The practical move is unglamorous — put real, legible text where your images carry meaning, instead of letting key information live only as decoration a machine cannot read.
- Vision also widens entity consistency: the engine now reconciles what it sees (your logo, your screenshots, your diagrams) with what it reads. Make the modes agree.
- It is early and imperfect, so do not bet a strategy on it — but the preparation is cheap, helps humans too, and pays off the moment the capability matures.
what the engine can read
Same chart, two outcomes. On the left the numbers live only as shapes, so a machine sees a picture and nothing it can quote. On the right the figure and the axis are written as real text, so the same machine reads "42 per cent" and can carry it into an answer. Put the meaning where it can be read.
Why a new sense changes the brief
Every time these systems gain a sense, the surface you are optimizing for gets larger. When they could only read text, the brief was about text: clear writing, clean markup, corroborated claims. The visuals were for humans, and it was reasonable to treat them as decoration a machine would never parse. That assumption was correct right up until this month, and now it is quietly wrong. A model that can read the words in a screenshot or the labels on a chart is a model for which your visuals are no longer a blind spot — they are content, with all the upside and exposure that implies. The upside is that good visual information can now be discovered and cited like text. The exposure is that misleading or empty visuals can now be misread, and that inconsistencies between what you show and what you say are now legible as inconsistencies.
This is not a call to panic or to redesign everything by Friday. The capability is new and it is rough; it misreads, it drops characters, it sometimes narrates things that are not in the image at all. Nobody should stake a quarter’s plan on it reading every figure correctly today. But the trajectory is unambiguous, and the preparation is the kind of work that pays for itself even if vision stayed exactly as imperfect as it is this month — because every step that makes a visual legible to a machine also makes it clearer to a person.
The shift, in three parts
The opaque layer became legible; the entity became something the machine can also see; so make the meaning readable in every mode. Open each part for where it changes the work.
01 The opaque layer just started to become legible
Until now there was a clean line between what a machine could read and what it could only store. Text was read; images were filed. Your infographic, your product screenshot, the chart where your best data actually lived — to a machine these were opaque rectangles, files with dimensions and nothing it could act on. This month that line moved. ChatGPT’s new vision can look at an image and read the words inside it: the figures in a chart, the labels on a diagram, the text in a screenshot of a document. The capability is built on a version of GPT-4 that understands images, and while it is new and far from flawless, it does something the previous generation simply could not — it discerns meaning in a picture. A layer of your work that machines used to skip is starting to be one they can parse, which makes it, for the first time, a layer worth optimizing rather than ignoring. The shift is small in the moment and large in implication: not that every picture is suddenly read, but that the category of "things a machine cannot parse" just got meaningfully smaller.
02 Your entity just became something the machine can also see
For schema and entity work this is the part that matters. An entity has been, in practice, a thing assembled from text — the name you use, the description you publish, the independent sources that corroborate both. Vision adds a second channel into that assembly. The same system can now look at your logo and read the name on it, look at the screenshots in your documentation and see which product they show, look at the images others have published of you and compare them to what it already believes. It is reconciling what it sees with what it reads. When those agree — the wordmark matches the text, the interface in the screenshot matches the one you describe, the diagram labels match your terminology — the entity sharpens, because two independent channels now point at the same thing. When they disagree, or when your visuals carry no legible signal at all, you have handed that reconciliation to luck. Entity consistency used to end at the edge of the text; it now runs through everything the machine can see, which means the same care you spend reconciling your name and description across sources now extends to your logo, your interface, and the pictures the world takes of you.
03 So make the meaning readable, in every mode
The work this calls for is almost boringly practical, which is the good news. Write real alt text that says what an image shows, not a keyword stub. Put the numbers behind a chart into actual text — a caption, a data table, a labelled axis a model can read — so the meaning does not live only as pixels a machine has to guess at. Label your diagrams in words rather than leaving them as wordless art. Keep your name, your wordmark, and your product’s look consistent across every image, so the seeing channel and the reading channel agree about who you are. None of this is exotic, and all of it helps the humans on your pages today, which is why it is safe to do now even though the vision capability is still rough. This is the same discipline the AC Group has applied to entities for 27 years — make the meaning unambiguous in whatever form the machine can take it in — extended to a form the machine has only just learned to take in.
Two pages, one new reader
Take two product pages making the same claim about results. The first presents its evidence as a beautiful rendered chart — the numbers are in there, but only as pixels, with no caption, no data table, and an alt attribute that reads "chart". To a human it looks authoritative; to the new reader it is a handsome blank, a picture with no quotable content. Everything that would make the claim citable is locked in a form the machine cannot open.
The second page makes the same chart but writes the meaning out alongside it: a one-line caption with the headline figure, a small data table beneath the graphic, axis labels in real text, and alt text that actually describes what the chart shows. It looks just as good to a person and reads cleanly to a machine, which can now lift the figure into an answer and attribute it. Same data, same design effort — but only the second page is legible to the reader that just arrived. When the engine reconciles the screenshots, the logo, and the labels on that second page against everything it already knows about the brand, the modes agree and the entity gets sharper. That difference, invisible a month ago, is the whole reason to make your visuals readable now rather than after everyone else has.
What to do with this
Start with the visuals that carry your most important claims. For each one, ask a blunt question: if a machine could only read this image and nothing around it, what would it learn? If the honest answer is "nothing", the meaning is trapped in pixels, and the fix is to write it out — a caption with the key figure, a data table under the chart, labels in real text, alt text that describes rather than keyword-stuffs. You are not decorating less; you are making sure the decoration is backed by something readable.
Then widen it to the entity. Make your name and wordmark consistent across every image a machine might encounter, so the seeing channel and the reading channel never contradict each other about who you are. Check that the product in your screenshots matches the product in your copy, that the diagrams use your real terminology, that the images others have of you would corroborate rather than confuse. None of this is glamorous and none of it is wasted: it serves the humans on your pages today and positions you for the readable visual layer that is plainly coming. Doing it early, while the capability is still rough and most competitors are ignoring it, is exactly the kind of unglamorous head start the AC Group has built entities on for ' + years + ' years — make the meaning unmistakable in every form the machine can read, including the one it learned this month. We have watched the readable surface widen before — from titles to body text to structured data — and the pattern holds each time: the teams that made the new layer legible early were reliably the ones the engines understood best once it mattered.
What "seeing" does not mean yet
It is worth being precise about the limits, because the gap between "can read images" and "reads your images reliably" is wide this month. The capability misreads text it should get right, drops characters and whole words, and occasionally narrates details that are not in the picture at all. It is uneven across image types — cleaner on a crisp screenshot than on a dense, stylized infographic — and it is available in one product, not across every engine a person might use. So "the engines learned to see" is a direction, not a finished fact. If you read it as a promise that a machine will now flawlessly extract every number from every chart you publish, you will be disappointed, and you may over-invest in a capability that is still finding its footing.
Read correctly, though, the limits do not weaken the case; they shape it. Precisely because the reading is imperfect, the burden is on you to make the meaning easy to read — a model that struggles with text baked into a busy graphic will do far better with a plain caption and a real data table sitting right beside it. You are not trying to dazzle the new reader; you are trying to be unmissable to it even on a bad day. The roughness of the capability is an argument for legibility, not against preparing for it.
Cheap insurance, human first
The reason to act now even on an early, imperfect capability is that almost none of the work is speculative. Real alt text helps screen-reader users today and has for years. Captions and data tables make a chart clearer to every human who reads it, not just to a machine. A consistent name and look across your images is plain brand hygiene. Each of these earns its place on a page whether or not an engine ever reads the image — which means you are not betting on vision maturing; you are doing good work that happens to also pay off if it does.
That is what makes this an easy call rather than a hard one. The downside case — vision stays rough for another year — still leaves you with more accessible, clearer, more consistent pages. The upside case — it matures quickly and becomes a real channel — finds you already legible while competitors who waited are scrambling to rewrite their pixels into words. When the cost of preparing is work you should be doing anyway and the payoff is a head start on a shift that is clearly underway, the only mistake is to treat "it is early" as a reason to do nothing.
Vision and entity: quick answers
What does it mean that an AI engine can "see" an image?
It means the model can take an image as input and reason about what is in it, rather than ignoring everything that is not text. The new vision capability in ChatGPT, built on a version of GPT-4 with image understanding, can look at a photo, a screenshot, a chart, or a diagram and describe it, read the words printed inside it, and answer questions about it. For years a machine treated your infographic or your screenshot as an opaque rectangle — it saw a file, not the meaning inside. That is starting to change. The words and numbers baked into your visuals are becoming legible to the same systems people now ask for answers, which means a whole layer of your work that used to be invisible to machines is beginning to count. It is early and imperfect — the model still misreads and occasionally invents — but the direction is clear enough to prepare for.
Should I put text inside my images now?
Put real, legible text where it carries meaning, and stop relying on text that exists only as decoration. If a chart’s key numbers live only as pixels in a pretty rendering, a machine reading the image has to guess at them and will often guess wrong; if the same numbers are also present as actual text — in a caption, a data table, a label the model can read — then both a human and an engine get them right. The rule is not to cram words into every picture; it is to make sure the meaning a visual carries also exists in a form that can be read rather than only admired. That includes genuine alt text that describes what the image shows, real labels on diagrams, and the underlying figures written out near the graphic. Decorative imagery with nothing to read stays exactly as useful to a machine as it always was, which is to say not very.
How does vision change entity consistency?
It widens what the engine reconciles. An entity used to be something a machine assembled from text — your name, your description, the sources that mention you. Now the same machine can also look at your logo, your screenshots, the diagrams in your documentation, and the images others post of you, and it will try to square what it sees with what it already knows. If those modes agree — the name on the logo matches the name in the text, the product in the screenshot matches the product you describe — the entity gets clearer. If they conflict, or if your visuals say nothing legible at all, you have left the reconciliation to chance. The work of keeping an entity clean no longer stops at the words; it extends to making sure that what the machine sees lines up with what the machine reads.
Is this worth acting on already, or is it too early?
It is early, and you should act on it anyway, because the cost is so low and the work pays off regardless. Vision in these tools is new and imperfect this month — it misreads text in images, loses characters, and sometimes describes things that are not there — so nobody should bet a strategy on it reading every chart perfectly today. But the things you would do to prepare are things you should be doing for humans already: real alt text, legible labels, the numbers behind a graphic written out, a consistent name and look across every image. None of that is wasted if the vision capability stays rough for a while, and all of it pays off the moment it matures. Preparing early for a readable visual layer is cheap insurance on a shift that is clearly underway.
A note on sources and timing
This is written in October 2023, as ChatGPT’s new voice and image features roll out — image understanding powered by a vision-capable version of GPT-4, alongside the arrival of a new image generator inside the chat. We have been deliberately measured about it, because the capability is genuinely new and genuinely imperfect: by its makers’ own account it misreads text in images, drops characters, and sometimes describes what is not there. We have described only what was public and working as of this writing, and nothing announced later. The durable point does not depend on the tool being flawless: machines have begun to read the text inside your images, your entity now extends to what they can see, and making that meaning legible is cheap, human-friendly work worth doing before it is urgent. That is the kind of entity discipline the AC Group has practiced for 27 years — now stretched to a sense the engines have only just acquired.