Chapter 33 — Key Takeaways
The screenshot-and-reread card for AI and generative imaging. Mostly tables and checklists; read it before you touch a generative tool or post an AI-assisted image.
The one idea
A camera measures light that was really there. A generative model predicts pixels that were not. Everything else in this chapter follows from that one difference. AI output is plausibility, not evidence — "it looks real" and "it is real" are now fully decoupled.
Core definitions
| Term | Definition in one line |
|---|---|
| Generative AI | A system that produces new content (here, images) by predicting what is statistically likely from patterns in training data — no scene behind it. |
| Diffusion model | Today's photorealistic generators; build an image by reversing noise step by step, steered by a text prompt, until an image emerges. |
| AI denoise | Software that removes sensor noise from a real capture by learning noise vs. detail — refines the record (gentle end of the spectrum). |
| AI upscaling | Enlarges an image and invents plausible fine detail to fill the new pixels — restoration shading into fabrication as you push it. |
| Generative fill / extend / replace | Selects an area and has the model replace it with newly generated content — invents pixels no light ever made (far end). |
| Provenance | The documented, checkable origin and edit-history of an image: where it came from and what was done to it. |
| Content credentials | Tamper-evident, signed metadata embedded in the file that records how an image was made/edited and travels with it. |
| Disclosure (AI) | Clearly telling viewers when/how AI generated or substantially altered an image, in plain language and (ideally) embedded provenance. |
The capture-to-conjured spectrum (sort every tool by how much it INVENTS)
| Tool | What it does to the record | Side of the line |
|---|---|---|
| AI masking / selection | Invents nothing — only understands your frame so adjustments land precisely | Editing |
| AI denoise | Interprets the real signal; strips sensor noise (back off before it smears real texture) | Editing |
| AI sharpen / lens correction | Corrects the real capture's optics; no new content | Editing |
| AI upscale (modest) | Invents plausible fine detail; harmless small, fabrication if pushed hard | Editing → edge |
| Generative fill / extend / replace | Invents whole new content — objects, regions, skies, faces | Generating |
| Text-to-image generation | Invents the entire image; no scene, light, or moment | Generating |
Rule of thumb: everything up to upscale refines what you captured; upscale and fill begin adding what you didn't.
The editing vs. generating decision
The test (inherited from Chapter 29, sharpened for AI): Enhancement presents the best version of what was actually there. Manipulation changes the record of what was there. For AI: did the source pixels come from the scene (editing) or did the model invent them (generating)?
The hard middle is removal. Resolve it with one question:
Was the removed element incidental to the photograph's subject or claim? - Incidental clutter / sensor dust → defensible enhancement (note it). - Anything that affects the image's claim about reality → manipulation.
And remember: the line is usually invisible in the result. A good edit and a good generation look identical. So the line must live in you as a decided rule, before your finger is on the tool — you cannot judge it from "does it look edited?"
"When is it honest?" — quick decision table
| Situation | Honest default |
|---|---|
| Noisy high-ISO real capture | AI denoise — fine; refines your real frame |
| Need a precise local edit (dodge/burn, sky darken) | AI masking — fine; invents nothing |
| Modest enlargement for a print | AI upscale in moderation — fine; disclose if pushed hard |
| Personal art, removing incidental clutter | Defensible (Ch.29 test) — note what you did |
| News / documentary / journalism | No added or removed content. Disclose any process precisely. |
| Real-estate, product, dating, "this is how it is" | No generative add/remove — it falsifies a claim. |
| Openly-labeled AI art / composite | Fine — present it as AI, label plainly, embed credentials |
| Any image of a real, recognizable person | Consent + right of publicity apply (Ch.32) regardless of tool |
Disclosure checklist (use BOTH layers)
- [ ] Layer 1 — plain-language label a human actually reads, placed where the image is seen.
- [ ] Layer 2 — content credentials / provenance embedded in the file (signed, verifiable, travels with it).
- [ ] Match detail to stakes: light note for personal art; precise account for news/documentary; legally required for synthetic people in ads in some places.
- [ ] Be specific — list what was done ("denoise, mask, grade, crop; nothing added/removed"). Specific builds trust; vague can erode it.
- [ ] Put it in the alt text too, so non-sighted viewers get the same disclosure ("AI-generated…").
- [ ] Test: would a normal viewer, in the normal way they meet this image, understand it was AI? If only by digging → not disclosed. Technically disclosed ≠ disclosed.
Copyright — the three independent questions
| # | Question | The stable shape of the answer |
|---|---|---|
| 1 | Do I own the output? | Captured photos are clearly yours; purely-prompted images may be protectable by no one (human-authorship requirement). More human creative control → stronger claim. |
| 2 | Was training on scraped images lawful? | Open, litigated, varies by country — and your work may be in the sets. |
| 3 | Can a generation infringe? | Plausibly yes — imitating a protected style/character, or a real person's likeness (right of publicity, Ch.32). "The AI did it" has protected no one. |
They are independent — answering one does not settle the others. Local law is changing; verify currently. Business/licensing mechanics live in Chapter 35.
Free hedge: keep your RAW files + capture metadata as proof of human authorship and that a real scene was captured (back up per Chapter 30's 3-2-1 rule).
Where the photographer is irreplaceable (Figure 33.7)
| The model CAN | Only a photographer CAN |
|---|---|
| Plausible generic imagery | Witness — this really happened, I was there |
| Average of all styles | Presence — being at the real, singular moment |
| Fast, cheap, infinite variants | Relationship — earned trust, real consent |
| "Looks like a photograph" | Voice — the choices only you would make |
| No stake, no memory, no duty | Responsibility — duties owed and honored |
The economics: as plausible images flood toward free, they become worthless (infinite supply). The scarce thing is the true, witnessed, voiced, accountable image — so it becomes more valuable. Compete on truth, presence, and meaning — never on plausible quantity.
Top mistakes → fixes
| Mistake | Fix |
|---|---|
| Judging a tool by where the button sits ("it's in my photo app, so it's fine") | Judge by what it does to the record, not its location |
| Assuming "it looks real" means it is real | Photorealism is the model's default output, not evidence — ask if the image must be true or only convincing |
| Burying disclosure (40th hashtag, stripped metadata) | Put a plain label where the image is seen, at the stakes it carries; embed credentials too |
| Letting a labeled "demo" generation circulate later unlabeled | A fabrication is honest only while disclosed — never let it travel as a photograph |
| Pushing AI denoise/upscale until it smears or invents | Back off at the point restoration tips into fabrication; you hold the dial |
| Thinking "the AI did it" shields you | You chose the prompt and published — liability and consent duties stay with you |
| Assuming a prompted image is automatically yours | Captured photos are yours; pure prompts may be owned by no one |
Portfolio Checkpoint (this chapter)
Write your stated stance on AI (≈150–300 words): which tools you use, where on the capture-to- conjured spectrum your line sits, what you will never do, and how you disclose. Specifics a viewer could hold you to — vague is a failed answer. And/or add one clearly-disclosed AI-assisted or comparison piece (e.g., the editing-vs-generating pair from Figure 33.4 / Case Study 33.2), with the disclosure as part of the piece.
Curation note: file it beside your Chapter 32 ethics audit — together they are the integrity of your portfolio. You'll fold the stance into the artist statement in Chapter 34.
One-line summary to carry
Honest use of AI is not "no AI." It's AI that refines the record (not invents it) plus a clear, specific account of what was done — and the irreplaceable thing you sell is the witnessed, present, accountable, voiced image a machine can never make.