Once articles are produced automatically, images are the manual job left. Image generators and templates shrink one to two hours of work per article to minutes.
Image errors, however, are less visible than text errors: bars that do not match their values, a picture resembling a known character, empty alt text, a 2 MB file that slows the page. They are usually noticed only when someone points them out.
This article covers what a person must still check when you automate featured images, diagrams and OG images (the image shown when a page is shared on social media). We covered pre-publish checks in general in "Automated blog publishing needs a quality gate". Here we focus on images.
Main sources for this article: Google Search Central's image SEO best practices (updated 2 March 2026) and guidance on generative AI content (updated 1 October 2026), the European Commission's FAQ on Article 50 of the AI Act and its Code of Practice (published 10 June 2026), the Japanese Agency for Cultural Affairs' "General Understanding on AI and Copyright" (15 March 2024), HTTP Archive's Web Almanac 2024 and the W3C WAI images tutorial.
| What this article answers | Key points |
|---|---|
| Why image review matters now | About 90% want to know if an image is AI-made. EU labelling duty from August 2026 (section 1) |
| Which images to automate, and how | Featured images with AI, OG images with templates, diagrams from data (sections 2–3) |
| The eight checks | Facts, rights, AI Act, provenance, alt text, speed, SEO, brand (sections 4–11) |
| Cost and effort | Generation USD 0.02–0.24 per image; setup 4–80 hours (section 12) |
| Where to start | Eight steps and a 24-item checklist (sections 16–17) |
1. What is changing: image automation is routine, and the rules have caught up
About 90% of people want to know whether an image is AI-made
In a survey published by Getty Images on 30 April 2024 (25 countries, about 30,000 adults), almost 90% of consumers said they want to know whether an image was created with AI. 76% agreed they can no longer tell whether an image is real. Hiding how an image was made can itself cost trust.
Japan, the US and the EU have clarified copyright and labelling
In Japan, a subcommittee of the Council for Cultural Affairs adopted the "General Understanding on AI and Copyright" on 15 March 2024, and the Agency for Cultural Affairs published a "Checklist and Guidance" on 31 July 2024. The US Copyright Office published its report on the copyrightability of AI outputs on 29 January 2025.
In the EU, Article 50 of the AI Act (Regulation (EU) 2024/1689) has applied since 2 August 2026. On 10 June 2026, the European Commission published the Code of Practice on marking and labelling AI-generated content.
The dotted point at the end is a grace period for providers of generative AI systems already on the market. There is no grace period for the website operators who publish the images. Content created before 2 August 2026 does not need to be labelled retroactively.
The image at the top of an article decides its loading speed
According to HTTP Archive's Web Almanac 2024, the LCP element (Largest Contentful Paint, the time until the largest element is shown) was an image on 73% of mobile pages and 83.3% of desktop pages. A heavy, automatically generated image that slips through without review slows the page directly.
2. Three kinds of image you can automate: featured images, diagrams and OG images
Image types differ in purpose and in the damage a mistake does, which sets how closely to review each.
| Type | Purpose | Impact of a mistake | Review depth |
|---|---|---|---|
| Featured image | First impression, listing cards | Rights infringement, off-brand look | Medium (rights and tone) |
| Diagram or chart | Conveys figures or a process | Wrong information spreads as is | High (check the numbers) |
| OG image | How the page looks when shared | A wrong title spreads | Low to medium (template) |
| Product or project photo | Shows the real thing | Misleading if it differs | Highest (never generate) |
The table shows that diagrams and product photos are factual components, not decoration. Generating product or project photos moves you towards the misleading claims that Japan's Act against Unjustifiable Premiums and Misleading Representations and the EU's Unfair Commercial Practices Directive target, so keep them out of automation.
Where the time goes per article
The breakdown below is DevNJ's estimate, not measured statistics.
| Task | Manual (min) | Automated (min) | What people still do |
|---|---|---|---|
| Finding and editing a featured image | 30–60 | 2–4 | Choose a candidate, check rights and tone |
| Drawing a diagram | 60–120 | 10–20 | Check figures against source data |
| Making an OG image | 10–20 | 1 | Check title line breaks |
| Writing alt text | 5–10 | 2–3 | Edit the draft |
| Compressing and exporting | 5–10 | 0 | Nothing (machine checks stop it) |
Searching and exporting shrink; checking figures and rights does not. Without reserving that time, review becomes a formality.
3. Three ways to produce images: generators, templates and drawing from data
There are three ways to produce images automatically, and each suits a different kind of image.
| Method | How it works | Suits | Weaknesses |
|---|---|---|---|
| Image generator | Creates a picture from a prompt | Featured images, illustrations | Inaccurate numbers and text; rights checks needed |
| Template | Pours title and colour into a fixed frame | OG images, social posts | Can look repetitive |
| Drawing from data | Renders SVG or charts from a table or JSON | Diagrams, charts, maps | Takes time to build |
At DevNJ, featured images come from a generator with a fixed house style, and diagrams are SVG written directly in the Markdown, so the text and diagram have little room to disagree.
Image generators: split the prompt into parts
Split the prompt into a shared house style and an article-specific subject. The style part covers colour, light, white space and texture, plus prohibitions: "no text, numbers, logos or human faces". The subject part describes only the objects and scene specific to the article, in one or two sentences.
This keeps listings consistent and avoids the riskiest elements. Store each prompt with its image.
Templates: test how OG images break first
OG images can be made by screenshotting an HTML frame or filling an SVG template. Without generative AI, rights and AI Act issues hardly arise, but they break on long titles and awkward line breaks. Test the longest and shortest titles in every language first.
Drawing from data: keep the source data in one place
Keep chart numbers in source data (a table or JSON) and draw the SVG from it; if the text uses the same data, the two cannot drift. Writing every figure into the diagram's desc (read by screen readers) turns checking into a string comparison.
Machine-checkable items (size, dimensions, format, presence of alt text) stop at step 3; judgement calls (correct figures, excessive similarity) go to a person at step 4. No route may bypass step 4.
4. Check 1: do the figures match the text?
The most frequent and least noticed mistake is a diagram that says something different from the text.
Common types of mismatch
Diagram errors usually fall into one of these five types.
| Type of mismatch | Example | How to catch it |
|---|---|---|
| Length does not match value | A 30% bar longer than a 60% bar | Draw from data; compare by eye |
| Truncated axis | Axis not starting at zero exaggerates gaps | Start bar charts at zero |
| Transcription error | Text 83.9%, diagram 89.3% | Compare desc and text as strings |
| Outdated figure | Text uses 2026, diagram 2025 | Put the check date in the caption |
| Garbled text or symbols | Unreadable letters drawn by AI | Never let AI draw text |
The most effective rule: never let an image generator draw text or numbers in a diagram. As Google's generative AI guidance notes, these models predict likely output rather than retrieve facts.
Turn checking into a string comparison
Comparing diagram figures with text figures as strings beats eyeballing. With SVG, a script can pull numbers from desc and text elements; numbers baked into a JPEG or PNG cannot be compared. The W3C images tutorial also advises putting chart information elsewhere on the page.
Always put the source and check date in the caption
Captions need a figure number, subject, source and check date, for whoever updates the figures later. Stop diagrams without a source link or date at step 3.
5. Check 2: copyright, trademarks and likeness
Rights are the most frequent question. This section is a general overview; consult a lawyer for individual decisions.
Japan: generation and use are judged on similarity and reliance
Japan's "General Understanding" judges infringement when publishing AI images as in any other case, by similarity to and reliance on an existing work. Even if the user did not know the work, reliance can normally be presumed if it was in the training data, so "I didn't know" is a weak excuse.
You may not be able to protect AI-generated images as your own
The reverse question: can you protect an AI image as your own work?
| Region | Position | What it means in practice |
|---|---|---|
| Japan | A work if a human made a creative contribution | Keep records of instructions, selection and edits |
| United States | Prompts alone are insufficient (2025 report) | Human expression such as edits and arrangement is protected |
| EU | A work must be the author's own intellectual creation | Hard to protect without human creative input |
Japan's guidance weighs the detail of instructions, attempts, selection and later edits, but attempts or selection alone are not enough. The US Copyright Office found prompts alone give insufficient control over expression, and the EU Court of Justice in Infopaq (C-5/08) requires "the author's own intellectual creation".
In short, treat AI-generated featured images as images that may be hard to protect if copied. Have people design anything you want to own exclusively, such as logos and characters.
Trademarks and likeness: no logos, characters or real people
On trademarks, the risk is other companies' logos or similar shapes creeping into an image, especially with prompts like "in the style of a famous brand". On likeness, generators sometimes produce faces that resemble real people; in Japan portrait and publicity rights (recognised in case law) apply, in the EU the GDPR, and the deepfake disclosure in the next section.
| Element that creeps in | Possible problem | Prompt or check |
|---|---|---|
| Other brands' logos, colours, shapes | Trademark, confusion | Prompt "no logos or text"; check visually |
| Existing characters or art styles | Copyright (similarity, reliance) | Never name works or artists in prompts |
| Faces resembling real people | Likeness, deepfake | Prompt "no faces"; show backs or hands |
| Photorealistic real buildings or places | Misleading, deepfake | Fix an illustrated style; no real place names |
| Shapes copying others' products | Design rights, unfair competition | Use your own product photos |
Check the tool's terms of use
Check the service's terms for commercial use, output ownership and IP indemnity, which varies by plan. Record the terms version and date with the prompt.
6. Check 3: EU AI Act transparency duties and labelling AI content
For sites with European readers, Article 50 of the AI Act has been a given since 2 August 2026. Businesses outside the EU, including Japanese ones, can be covered where the output is used in the EU (Article 2).
Article 50(2): providers of generative AI mark outputs
Paragraph 2 requires providers of generative AI systems to mark outputs in a machine-readable, detectable way, such as watermarks or metadata; assistive standard editing is excluded. A website operator is normally a deployer, but offering your own image generation feature to users may make you a provider.
Article 50(4): deployers disclose deepfakes
Paragraph 4 requires deployers who publish deepfakes to disclose that the content was AI-generated or manipulated; a business website is professional use. Per the Commission's FAQ, a deepfake (Article 3(60)) has a high level of similarity to real people, objects, places or events that exist or could exist, and would falsely appear authentic, judged by audience expectations and context.
The middle band is a rule of thumb derived from the definition, not a line in the regulation. Stylised illustrations rarely meet the criteria, but one showing a recognisable real politician does.
How to disclose, and the penalties
Disclosure must be clear and made at the latest at first exposure (paragraph 5), with a visible label; deployers cannot rely on the provider's machine-readable marking. Evidently artistic, satirical or fictional works need only an unobtrusive disclosure. Infringements can lead to fines of up to EUR 15 million or 3% of worldwide annual turnover (Article 99).
| Example image | Deepfake? | Labelling |
|---|---|---|
| Abstract illustrated featured image | Usually not | Not required; explain on a policy page |
| SVG diagram or chart | No (not AI-generated) | Not needed |
| Photorealistic "customer" | Possibly | Label near the image, or better, don't use it |
| Photoreal recreation of a real shop | Possibly | Label; replace with a real photo |
| Brightness fix or background removal | Not if standard editing | Not needed |
| Real person composited into a new scene | Possibly | Label; avoid misleading uses |
The Code of Practice, and Japan
The voluntary Code of Practice (10 June 2026) makes compliance easier to show, and a common EU labelling icon exists. Japan has no general labelling duty, but images that flatter a product can breach its law on misleading representations. A single rule for both regions, "don't make images that look real; if you do, label them", saves deciding case by case.
7. Check 4: provenance data (C2PA and IPTC)
C2PA and IPTC metadata record how an image was made in a machine-readable way. Both are often missing even though you think you added them, so include them in your checks.
C2PA (Content Credentials)
C2PA is an open standard for recording, with a signature, the origin and edit history of content (Content Credentials); its steering committee includes Adobe, Google, Meta, Microsoft, OpenAI and Sony. Google announced in September 2024 that "About this image" in Search would use it, but nothing is shown for images without C2PA data.
IPTC DigitalSourceType
IPTC photo metadata has a DigitalSourceType field describing how an image was made, and Google supports four values, including trainedAlgorithmicMedia (generated by a trained model) and compositeWithTrainedAlgorithmicMedia (a composite with generated elements). Google's generative AI guidance says AI-generated images sold on e-commerce sites must carry this value, and recommends it for other sites too.
| Mechanism | What it holds | How easily it is lost | How to check |
|---|---|---|---|
| C2PA | Creation and edit history, signature | Can be lost on recompression or conversion | Content Credentials Verify |
| IPTC DigitalSourceType | Generated, captured or composite | Can be lost on resizing or conversion | Read it with exiftool etc. |
| IPTC copyright and credit | Creator, credit, rights URL | Same as above | Same as above |
| Structured data (ImageObject) | Creator, licence, credit | Rarely lost (lives in the page) | Rich Results Test |
| Visible label | Wording such as "AI-generated" | Never lost | Visual check |
Prevent "added but gone"
Metadata is often stripped during resizing or conversion. Read the metadata of a final file served from your CDN, and add a visible label where it matters.
8. Check 5: alt text and accessibility
Having AI draft alt text is useful, but never publish the draft unchecked.
Missing alt text is still one of the most common errors
WebAIM's 2026 test of one million home pages found images without alt text on 53.1% of pages, second only to low contrast. Google's generative AI guidance lists image alt text among items for human review, since AI alt text can describe things not in the image.
How to write alt text for each type of image
Following the W3C WAI (Web Accessibility Initiative) alt decision tree, how to write alt text depends on the image's role.
| Image type | How to write alt text | Example |
|---|---|---|
| Decorative featured image | Empty (alt="") or a short subject | alt="" |
| Meaningful illustration | The meaning, briefly | "Inspection sheet and magnifier" |
| Chart or diagram | The key point; figures in text or table | "Share by format; JPEG leads" |
| Image containing text | The text itself | The wording of a notice |
| Image used as a link or button | Destination or action | "Go to contact page" |
Repeating a card's title in the featured image's alt text means it is read twice. If the same information sits next to the image, leave alt empty.
Three WCAG criteria for diagrams
Diagrams touch the following WCAG 2.2 success criteria.
- 1.1.1 Non-text Content: provide a text alternative; show complex diagrams in text or a table
- 1.4.5 Images of Text: use real text rather than text baked into images
- 1.4.11 Non-text Contrast: give bars and lines at least 3:1 contrast against adjacent colours
Inline SVG with role="img", aria-label and desc works with screen readers. In Europe, the European Accessibility Act has applied since 28 June 2025 (see "Redesigning your website for the European Accessibility Act").
9. Check 6: file size, format and page speed (LCP)
File size is easy to automate and easy to forget; generators often output multi-megabyte PNGs.
The LCP threshold and how often it is an image
web.dev recommends an LCP of 2.5 seconds or less for at least 75% of visits, and anything over 4.0 seconds counts as poor. In Web Almanac 2024, 16% of mobile pages lazy-loaded their LCP image. Because this delays the start of loading, web.dev says never to lazy-load the LCP image.
Formats: JPEG dominates, AVIF is rare
The same study shows that fewer than one in ten pages use WebP or AVIF, formats that make it easier to cut file size, for their LCP image.
Image generators often output PNG, so build conversion to JPEG, WebP or AVIF into the export step.
Size guidelines by use
These are DevNJ's guidelines. Adjust them against your own Core Web Vitals.
| Use | Dimensions (px) | Format | Size limit (guide) | Loading |
|---|---|---|---|---|
| Featured image (top of article) | 1200×630 | JPEG, WebP, AVIF | 180 KB | Not lazy; fetchpriority="high" |
| Listing card | 600×315 etc. | WebP, AVIF | 60 KB | Lazy |
| OG image | 1200×630 | JPEG, PNG | 300 KB | N/A (fetched by platforms) |
| Article diagram | viewBox 520 wide | Inline SVG | 10–20 KB | Inline with HTML |
| Article photo | Up to 1200 wide + srcset | WebP, AVIF | 200 KB | Lazy |
Machine checks should stop images over the limits, LCP candidates with loading="lazy", and images without width and height (a cause of CLS). We covered monitoring in "Automating Core Web Vitals monitoring".
10. Check 7: Google's image SEO guidance and Open Graph
Images represent the article in image search, Discover and social sharing, so keep to Google's basics.
Google's image SEO best practices
The key points of Google Search Central's image SEO page (updated 2 March 2026) are below.
| Item | What Google says | Watch out for in automation |
|---|---|---|
| Supported formats | BMP, GIF, JPEG, PNG, WebP, SVG, AVIF | Match the extension to the file type |
| Alt text | Descriptive and in context; no stuffing | People edit the AI draft |
| File names | Short and descriptive | Derive from the slug |
| Responsive images | srcset and sizes; always keep src | Never drop src |
| Image sitemaps | Help Google find images | Verify the CDN domain too |
Article structured data and Discover images
For the image in article structured data, Google recommends a representative image of at least 50,000 pixels (width × height), in 16:9, 4:3 and 1:1 (on maintenance, see "Automating structured data maintenance"). Discover recommends images at least 1200 px wide, over 300,000 pixels and 16:9, with max-image-preview:large, and asks you to avoid logos, text-heavy images and exaggerated imagery.
OG images: include og:image:alt
The Open Graph protocol says a page with og:image should also specify og:image:alt; generate it from the article's alt text. OG images with a baked-in title keep circulating after a title change, so regenerate them whenever the title changes.
11. Check 8: brand consistency
The last check is visual consistency, which a human eye used to protect.
Five ways a brand drifts
With 30 or 50 images side by side, drift becomes obvious.
| Type of drift | What happens | Fix |
|---|---|---|
| Colour drift | Each image has a different tone; the list looks patchy | Fix brand colours in the style prompt |
| Style drift | Photoreal, 3D and watercolour get mixed | Make the style prompt a shared component |
| Repeated composition | Every article shows a desk and a plant | Add a motif specific to each article |
| Generic AI look | The typical generator look feels cheap | Specify white space and texture |
| Stray text | Meaningless letters appear in the background | Prompt "no text"; check visually |
Review on the listing page
Review on the listing page once a month; if a new image stands out, fix the prompt and regenerate. DevNJ fixes an off-white ground, terracotta accents, white space and paper texture, and varies only the subject.
12. Cost and effort ranges
Cost depends more on building the workflow and on human review than on generation. Yen figures use EUR 1 = about JPY 176 (European Central Bank reference rate, JPY 175.97 on 22 October 2026).
Generation: a few cents per image
Google's Gemini API pricing (checked 29 October 2026) lists the following per-image prices for image models, at standard 1K resolution (about 1024 px square) unless stated.
| Model | Per image (USD) | Batch (USD) |
|---|---|---|
| Gemini 3.1 Flash Lite Image | 0.0336 | 0.0168 |
| Gemini 3.1 Flash Image | 0.067 | 0.034 |
| Gemini 3 Pro Image | 0.134 | 0.067 |
| Gemini 3 Pro Image (4K) | 0.24 | 0.12 |
Three candidates for each of 30 monthly articles is 90 images, about USD 6 at USD 0.067. Three minutes of review each is 1.5 hours, or EUR 120 at EUR 80 an hour. Most of the cost is review time.
Three stages of workflow
The table is DevNJ's estimate, not measured statistics, and varies with the site and volume.
| Stage | What it includes | Setup time | Monthly (EUR) |
|---|---|---|---|
| 1. Procedure and manual checks | Style prompt, size limits, the 24-item checklist | 4–8 hours | 0–10 |
| 2. Templates and machine checks | OG template, SVG diagrams, automatic size, dimension and alt checks | 16–40 hours | 5–30 |
| 3. Generation and approval flow | Built-in generation, provenance data, approval screen | 40–80 hours | 20–100 |
At EUR 80 an hour, stage 1 costs EUR 320–640 (about JPY 56,000–113,000), stage 2 EUR 1,280–3,200 (about JPY 225,000–563,000) and stage 3 EUR 3,200–6,400 (about JPY 563,000–1.13 million). Monthly costs cover the generation API, image delivery (CDN) and storage.
For small and mid-sized sites, stage 2 delivers most of the benefit, since templated OG images and diagrams raise almost no rights or AI Act issues. Consider stage 3 above 20 articles a month.
13. Common failures
These are failures DevNJ has seen.
| Failure | What happened | Prevention |
|---|---|---|
| Letting AI draw a chart | A chart with bars that did not match the values went live | Draw charts with figures from data |
| "In the style of..." prompts | The image resembled a famous work | Never name works or artists |
| Photoreal people | A "customer" resembling a real person appeared | Do not let the generator draw faces |
| Publishing the raw file | A 3 MB PNG pushed LCP to over 5 seconds | Convert on export; check size |
| Trusting auto alt text | Alt text described things not in the image | People edit the AI draft |
| Stale OG image | Old image kept circulating after a title fix | Regenerate on title change |
The first and fourth rows are the most common, and both can be prevented at the production stage.
14. Japan and Europe compared
A comparison of the points that affect images. Consult a specialist for individual decisions.
| Item | Japan | Europe (EU and the Netherlands) |
|---|---|---|
| Labelling AI images | No general duty | Deepfakes must be disclosed (from 2 August 2026) |
| Copyright infringement test | Similarity and reliance (Agency guidance) | National law and EU directives; TDM opt-out (2019/790 Art. 4) |
| Protection of AI output | If a human made a creative contribution | Requires the author's own intellectual creation |
| Images of people | Portrait and publicity rights (case law) | Personal data under the GDPR; likeness under national law |
| Misleading images | Act against Misleading Representations | Unfair Commercial Practices Directive |
| Accessibility | Reasonable accommodation mandatory since April 2024 | European Accessibility Act (from 28 June 2025) |
Europe has the clearer labelling duty, so one labelling policy aligned with Europe is simplest; check misleading claims and likeness under both systems.
Under Article 4 of the DSM Copyright Directive, rights holders can opt out of text and data mining in machine-readable form, for example in terms or robots.txt.
15. Worked examples (three fictional companies)
Three fictional companies, made up for illustration (not real businesses).
| Company (fictional) | Situation | Suitable stage | First decision |
|---|---|---|---|
| A: Japanese parts maker, 40 staff | 4 blog posts a month in JA/EN; many diagrams | Stage 2 | SVG diagram template, figure-check procedure |
| B: Dutch Japanese-tea shop, 6 staff | 8 product posts a month in JA/EN/NL | Stage 1→2 | Real product photos; AI for featured images only |
| C: Travel guide for Europe, 25 staff | 20 area guides a month; many landscapes | Stage 3 | No photoreal generation of real places |
A first builds a template that draws every diagram from source data. B photographs its own products and uses AI only for featured images. C decides that real places get real photos and AI only draws obvious illustrations before investing.
16. Steps to get started
Eight steps. The first four need no tools.
- Sort your images. Classify the images in your last 20 articles as featured, diagram, OG or product photo
- Choose a method per type. Featured images with AI, OG with templates, diagrams from data, products as real photos
- Fix one style prompt. Colour, white space, texture, and "no text, logos or faces"
- Set size and dimension limits. Start from the table in section 9
- Add machine checks. Size, dimensions, format, alt text, lazy loading, caption source and date
- Assign reviewers and time. Who checks figures, and who checks rights and the AI Act
- Keep records. Prompt, service and terms version, reviewer and date for each image
- Review the listing monthly. Look for brand drift and check Core Web Vitals
From step 5 onwards, automate the items that actually caught problems when you ran them by hand. For the overall approach, see also "Start site operations automation with an inventory".
17. Checklist (8 checks × 3 items = 24 items)
Send items a machine can judge to automatic checks and items that need judgement to human review.
1. Figures match the text
- Diagrams with figures are drawn from source data (not by an image generator)
- The figures in diagrams match those in the text
- Captions include figure number, source and check date
2. Copyright, trademarks and likeness
- Prompts contain no names of works, artists or brands
- No elements resemble other companies' logos or existing characters
- No faces resemble real people
3. EU AI Act labelling
- No photoreal images that look like real people, places or events
- If used, "AI-generated" is shown where the image is first seen
- Edits to real photos stay within standard editing
4. Provenance data
- Generated images carry IPTC DigitalSourceType
- Metadata is still present in the published file
- E-commerce product images carry TrainedAlgorithmicMedia
5. Alt text and accessibility
- Every image has an alt attribute (empty if decorative)
- A person has read and edited AI-drafted alt text
- Diagram content also appears in the text or a table
6. File size, format and speed
- Featured images are 1200×630 and 180 KB or less
- The image at the top of the article is not lazy-loaded
- PNGs are converted to JPEG, WebP or AVIF
7. Google image SEO and Open Graph
- The image in structured data represents the article
- og:image and og:image:alt are output
- The OG image is regenerated when the title changes
8. Brand consistency
- The shared style prompt was used
- No meaningless text appears in the background
- The image does not stand out on the listing page
18. Summary
- Automating images starts with deciding what a person looks at before publishing, before deciding how to generate
- Generation costs about USD 0.02–0.24 per image; most of the cost is review time
- Draw diagrams with figures from source data, with SVG for example, so they match the text
- Most rights problems can be prevented at the production stage by keeping works, logos and faces out of prompts
- For European audiences, Article 50 of the AI Act requires labelling of images that look real; not making such images is the easiest policy to manage
- Check that provenance data is still present in the published file, not just that you added it
- Machine-check alt text, size and OG output, and have people edit AI drafts before publishing
Sources
Google policies and documentation (Google Search Central, Google)
- Google image SEO best practices — Google Search Central: supported formats
- Guidance on using generative AI content on your website — Google Search Central: pre-publish review including alt text
- Image metadata in Google Images — Google Search Central: IPTC fields
- Article structured data — Google Search Central: 50,000-pixel images
- Google Discover — Google Search Central: images at least 1200 px wide
- How we're increasing transparency for gen AI content with the C2PA — Google: C2PA in "About this image"
- Gemini API pricing — Google AI for Developers: per-image prices of image models
Page speed and accessibility (web.dev, HTTP Archive, W3C, WebAIM)
- Optimize Largest Contentful Paint — web.dev: the 2.5-second threshold
- Performance — Web Almanac 2024 (HTTP Archive): image LCP on 73% / 83.3% of pages
- An alt Decision Tree — W3C WAI: alt text by image type
- Complex Images — W3C WAI: putting chart information on the page
- Understanding SC 1.1.1 Non-text Content — W3C: non-text content
- Understanding SC 1.4.5 Images of Text — W3C: images of text
- Understanding SC 1.4.11 Non-text Contrast — W3C: 3:1 contrast for graphics
- The WebAIM Million — WebAIM: missing alt text on 53.1% of pages
- The Open Graph protocol: og:image and og:image:alt
EU AI Act and European rules (EUR-Lex, European Commission)
- Article 50: Transparency obligations — AI Act Service Desk (European Commission): Article 50(2), (4), (5) and (7)
- Transparency obligations under Article 50 of the AI Act — European Commission: the three deepfake criteria
- Code of Practice on Transparency of AI-generated Content — European Commission: the marking and labelling code
- Commission publishes Code of Practice on marking and labelling AI-generated content — European Commission: publication on 10 June 2026
- Regulation (EU) 2024/1689 (AI Act) — EUR-Lex: Articles 2, 3(60), 50 and 99
- Directive (EU) 2019/790 (DSM Copyright Directive) — EUR-Lex: reservation of text and data mining
- Infopaq International (C-5/08) — EUR-Lex: the author's own intellectual creation
Provenance (C2PA, IPTC)
- C2PA — Coalition for Content Provenance and Authenticity: overview of Content Credentials
- Content Credentials Verify: checking an image's provenance
- Digital Source Type vocabulary — IPTC: definitions of values for generated images
- IPTC Photo Metadata User Guide — IPTC: metadata for AI-generated images
Copyright and labelling in Japan and the US
- AI and copyright — Agency for Cultural Affairs, Japan: the General Understanding on AI and Copyright
- Copyright and Artificial Intelligence — US Copyright Office: Part 2, Copyrightability
- Fair labelling — Consumer Affairs Agency, Japan: misleading representations
Surveys and statistics
- Nearly 90% of Consumers Want Transparency on AI Images — Getty Images: survey of about 30,000 people in 25 countries
- Euro foreign exchange reference rates: Japanese yen — European Central Bank: reference rate used for yen conversions
All sources were checked on 29 October 2026. Legal interpretations and prices change, so check the official sites before making decisions and consult a specialist for individual cases. The effort and cost figures are provisional DevNJ estimates, not measured statistics, and vary with the size of the site and team.
FAQ
- Is it a problem to generate blog featured images with AI?
- Using AI is not a problem in itself; Google says it looks at the quality of content, not how it was made. Before publishing, though, a person should check that the image does not resemble existing works, characters, logos or real people, that it fits the article, and that the file is not too heavy. Google also asks for a human review of AI-written alt text.
- Do I have to label AI-generated images as AI-generated?
- In the EU, Article 50 of the AI Act has applied since 2 August 2026. When you publish, in a professional context, an image that resembles real people, places or events and could be mistaken for genuine (a deepfake), you must disclose visibly that it was generated or manipulated by AI. Obvious illustrations and diagrams usually fall outside this, but the test depends on resemblance and on what viewers would expect. Japan has no general labelling duty at present.
- Can I let an image generator draw the charts and diagrams in my articles?
- For diagrams that contain figures, we recommend not doing so. Image generators do not guarantee that numbers are right: bars that do not match their values, wrong scales and garbled text are common. Draw diagrams from the table or data with SVG or a chart library, and take the numbers in the text from the same source data so the two cannot drift apart.
- Who owns the copyright in AI-generated images?
- Japan's Agency for Cultural Affairs guidance (March 2024) says AI output can be a copyright work only where a human made a creative contribution. The US Copyright Office concluded in January 2025 that prompts alone are not enough, and EU law requires an author's own intellectual creation. Assume you may not be able to stop others from reusing such images.
- What file size and format should a featured image have?
- DevNJ uses 1200 x 630 pixel JPEGs of 180 KB or less for featured images. Google Discover recommends images at least 1200 px wide, and the same size works as an Open Graph image. The large image at the top of an article is usually the LCP element, so do not lazy-load it, and keep it light, using WebP or AVIF where you can.
- How much does it cost to automate image production?
- Generation itself is cheap: Google's image generation API costs roughly USD 0.02 to 0.24 per image (October 2026). Most of the cost is building the workflow and human review. By DevNJ's estimate, a written procedure with manual checks takes 4 to 8 hours to set up, templates and machine checks 16 to 40 hours, and generation with provenance data and an approval screen 40 to 80 hours.
Planning a site that has to work in both markets?
Get in touch
