After a wedding, several hundred frames sat on my memory card — and ahead of me lay the job photographers love least: sifting, discarding, rating, preparing for editing. This time I handed the first pass to an AI. More precisely: to several. The shoot was split into blocks of roughly 25 frames, and a small fleet of AI instances worked through them in parallel — sharpness, closed eyes, duplicates, impact, plus a rating suggestion and development presets, all written machine-readably next to the originals.
The result was the kind of first draft a diligent assistant delivers: not final, but a genuine shortcut. The star suggestions often landed close to my own judgement, obvious rejects were reliably flagged, and the emotional moments — embraces, laughter, tears — floated to the top. At sifting, vision AI is strong. It looks fast, tirelessly, and without the third-pass blindness we humans develop.
And then, the pony
Different project, same technology: for a travel article, the AI captioned photos from a hike. On one of them, it wrote, "two sheep" graze by a young fruit tree. The caption went live. Only a look at the full-size image revealed: it is one animal, and it is no sheep — it's a pony, a single grey. Nothing about the photo was ambiguous; at full resolution it is perfectly clear. But the AI hadn't seen the full photo. It had seen a stamp-sized preview on a contact sheet.
[BILD: image pair or crop — contact-sheet thumbnail (~300 px) on the left, full-resolution grey pony on the right. Source: bw Witten article, wit-0014. Target path /images/articles/foto-halluzination/.]
That is the heart of the problem, and it is more fundamental than one amusing anecdote: where the signal is thin, vision AI still delivers an answer. A few pale pixels against grass — statistically, that could be a sheep, and maybe two. The model doesn't say "too small, can't tell". It completes the probable, in the same calm voice it uses when it's right. In photography, hallucination doesn't mean seeing ghosts; it means turning scarce information into a concrete claim.
There are quieter traps of the same kind. Image files carry an invisible rotation flag — portrait shots preview upright while the raw signal lies sideways; trust the AI blindly and you end up straightening pictures that were never crooked. And when the high-resolution version of a picture is missing, the AI happily keeps working with the small one — the quality loss doesn't bother it. It bothers me, later.
My rules for AI in the photo archive
- Pre-selection yes, final say no. The AI sorts, weighs, suggests. What gets deleted and what makes the album, I decide at the big screen.
- Claims only from full resolution. Every caption, every subject identification ("horses and sheep grazing…") is a factual claim that gets published. Identification happens on the original, never on a preview. Since the pony, that's a fixed, written-down rule here (the AI keeps notes).
- When unsure, go one level more general. "Pony" instead of a guessed breed, "bird of prey" instead of "red kite" if the image doesn't support it. Vague and right beats precise and wrong.
- Quality is never silently downgraded. If the original is missing, the rule is ask, not guess — a small file is a finding, not a substitute.
What to take away
Vision AI is an excellent sorter and an unreliable witness. For pre-selecting your holiday photos, let it run without worry — the cost of errors is low, the time saved is real. But the moment images turn into statements — who is in the photo, what does the scan show, how many were there — the pony principle applies: the AI may have seen less than its answer suggests. Ask what the claim is based on, and when in doubt, look at the original yourself.
/compact — the essentials, if context is running low:
Vision AI is an excellent sorter but an unreliable witness: where the signal is thin, it still delivers an answer — in the same calm voice it uses when it's right. That's how a single pony on a stamp-sized preview became "two sheep" — in photography, hallucination means turning scarce information into a concrete claim. The countermeasures are fixed rules: claims only from full resolution, one level more general when unsure, no silent quality downgrades — and the final say stays with a human looking at the original.