Industries / Finance and business services / Advertising and marketing / Graphic and web design
Illustration
Illustration is drawing and making the pictures that go into a client's brand, publications and website, by hand, in drawing software, from stock libraries or with AI image generators, and settling who owns the result.
7
Milestones achieved, from 1963 to 2025.
3
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1963
Drawing on a screen
Ivan Sutherland's Sketchpad, written for his doctoral thesis at MIT and run on its TX-2 computer, let a user draw, erase, magnify and shrink shapes directly on the display with a light pen, a pen-shaped sensor held against the screen. The Computer History Museum calls it "generally recognized as the first computer-aided drafting (CAD) program."
Credit
Ivan Sutherland, MIT
1987
Editable curves on the desktop
Adobe released Illustrator 1.0, its first desktop drawing program, in March 1987, by the Computer History Museum's record of the company. It stores shapes as vectors, mathematical outlines that scale without blurring, which artists build with a pen tool that places Bézier curves, curves set by a few anchor points and handles. Adobe says it let artists make work that is "precise, perfect, and adjustable," and the type designer Mark Simonson credits it with "introducing Bézier curves and the pen tool to type designers and lettering people."
Credit
Adobe
1990
Photoshop 1.0 ships
Photoshop edits raster images, pictures stored as grids of colored dots called pixels, and became the partner to Illustrator's curves. Adobe licensed it from the brothers Thomas and John Knoll in 1988 and bought the rights in January 1989, and version 1.0 "started shipping early in 1990," the Computer History Museum says; over the next ten years more than 3 million copies were sold.
Credit
Thomas Knoll, its sole engineer for version 1, and John Knoll; Adobe
2006
Stock images for cents
iStockphoto, begun by Bruce Livingstone as a photo-swapping site, sold credits that bought photos for $0.25 each, while traditional picture agencies, in Livingstone's words, cared about "keeping their margins high -- 70 percent plus." In 2006, after the site reached $12 million in revenue, Getty Images bought it for $50 million, VentureBeat reported.
LessonCheap licenses filled the market with pictures; photographers later told Livingstone that competition was "getting too fierce" and that they "weren't getting the same revenue anymore.".
Credit
Bruce Livingstone; Getty Images
2022
Pictures from a sentence
Within five months three services let anyone type a description and get a finished-looking picture: OpenAI's DALL-E in April 2022, Midjourney in July and Stable Diffusion, whose trained model Stability AI released publicly on 22 August under a license that allows commercial use. Outside economists then measured the effect. After DALL-E and Midjourney, image-related freelancers on the platform Upwork saw a 3.7% drop in monthly jobs and a 9.4% drop in monthly earnings, by the count of Xiang Hui, Oren Reshef and Luofeng Zhou, and on one large freelancing platform posts for image-making jobs fell 17%, by the count of Ozge Demirci, Jonas Hannane and Xinrong Zhu (Management Science, 2025).
Credit
OpenAI; Midjourney; Stability AI
2023
AI images lose US copyright
On 21 February 2023 the US Copyright Office ruled on Zarya of the Dawn, a comic by Kris Kashtanova: it kept protection for her text and for her selection and arrangement of the pages, but found that the images made with Midjourney "are not the product of human authorship," comparing the prompts to the general directions of "a client who hires an artist." Its January 2025 report added that protection does not come from "the mere provision of prompts." In March 2025 the federal appeals court in Washington, D.C., ruled in Thaler v. Perlmutter that the Copyright Act requires a human author, while noting that this "does not prohibit copyrighting work that was made by or with the assistance of artificial intelligence," and the Supreme Court declined to hear the case on 2 March 2026.
LessonIn the US a client can register the human-made parts of an illustrated work, but not pictures generated from prompts.
Credit
US Copyright Office; US Court of Appeals for the D.C. Circuit
2025
Anti-AI cloaks shown removable
Glaze, from Shawn Shan, Ben Y. Zhao and colleagues at the University of Chicago (2023), adds "barely perceptible perturbations," small changes to the pixels, to an artwork so that AI models cannot copy the artist's style, and NightShade is a related tool. In 2025 researchers from the University of Cambridge, the Technical University of Darmstadt and the University of Texas at San Antonio presented LightShed, which detected NightShade-protected images with a 99.98% true-positive rate and a 100% true-negative rate and then stripped the changes out.
LessonProtection hidden in the pixels can be removed by later software, so artists cannot rely on it alone.
Credit
Hanna Foerster and colleagues (LightShed); Shawn Shan, Ben Y. Zhao and colleagues (Glaze)
Watch for
Editable SVG from a prompt
Recraft says its V4 model, released in February 2026, makes "production-grade vector images" and that its "Vector variants return editable SVG," the standard file format for vector drawings, at $0.08 per vector image or $0.30 for the Pro vector version, by its price list.
Achieved whenan outside party publishes a test in which working illustrators make standard client edits to V4 vector files, with the timings and files released.
Claimed by
Recraft, February 2026
Getty's appeal on model files
On 4 November 2025 England's High Court held that Stability AI's image model, Stable Diffusion, which "does not store or reproduce any Copyright Works," is not an "infringing copy" of Getty's pictures; Getty had dropped its claim about training, accepting there was no evidence it took place in the UK, before the ruling. Getty Images and iStockphoto argue on appeal that the model, made outside the UK and brought into it, is an infringing copy under section 27 of the Copyright, Designs and Patents Act 1988, and the Court of Appeal heard the case on 11 and 12 November 2026.
Achieved whenthe Court of Appeal gives judgment.
caselaw.nationalarchives.gov.ukjudiciary.ukipkitten.blogspot.com
Claimed by
Getty Images, in arguments filed in July and October 2026
Authorship after 624 prompts
Jason Allen has asked a federal court in Colorado to rule that he is the author of a picture he made by entering prompts into Midjourney "at least 624 times" and then editing the result in Photoshop, after the Copyright Office refused to register it as a whole. In its brief of 6 February 2026 the Office says he "likely could have successfully registered the Work had he simply agreed to limit his claim by disclaiming the underlying material that he did not author."
Achieved whenthe court rules on those motions (Civil Action No. 24-cv-02665-WJM).
Claimed by
Jason Allen, in Allen v. Perlmutter, with both sides' motions for summary judgment filed by February 2026
Open
01
A tracing test for drawings
A public test of tracers, tools that turn a scanned drawing into vector paths, on about 100 openly licensed or public-domain hand drawings. Score each tracer's output for fidelity, how closely it matches the original by two standard image-similarity measures (SSIM and LPIPS), and for editability: the number of paths and anchor points, stray fragments, and the minutes working illustrators need to make each file ready for a client.
The briefCredit · unclaimed02
Credentials on artists' platforms
A tally of whether Content Credentials survive upload. These are signed records, under the C2PA standard, of who made an image and how. Upload human-made illustrations to the sites illustrators use, such as Behance, Dribbble, ArtStation, DeviantArt, Adobe Stock, Shutterstock, iStock, Instagram, Pinterest, LinkedIn, X and Etsy, and check for each site whether the record is still in the file the site serves, still in a downloaded copy, and shown to viewers.
The briefCredit · unclaimed03
Editing tests for AI vectors
A blind test in which working illustrators make standard client edits to SVG illustrations from AI generators and to hand-made control files: recolor to a three-color brand palette, move one element, delete the background and resize for a small icon. Record the minutes taken, success or failure, the counts of paths and anchor points, and a 1 to 5 rating for how easy the file was to edit.
The briefCredit · unclaimed
Open 01 · the brief
A tracing test for drawings
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of tracers, tools that turn a scanned drawing into vector paths, on about 100 openly licensed or public-domain hand drawings. Score each tracer's output for fidelity, how closely it matches the original by two standard image-similarity measures (SSIM and LPIPS), and for editability: the number of paths and anchor points, stray fragments, and the minutes working illustrators need to make each file ready for a client.
Why it is still open
Published benchmarks score looks, not cleanup work. Adobe's VectorArk paper of May 2026 says such methods are "typically evaluated on synthetic benchmarks." StarVector's benchmark scores fidelity. VTracer, an open-source tracer, says its output is "much more compact" than Illustrator's Image Trace but publishes no data. Tools for tracing scans have existed since Adobe Streamline in 1989, by the Computer History Museum's record.
Built with what exists
VTracer (MIT license), Potrace, Illustrator's Image Trace, Vectorizer.ai and the open StarVector-1B and StarVector-8B models on Hugging Face; Python scripts that render each SVG and count its path commands.
First step
Gather 50 public-domain line drawings, run VTracer and Potrace on their default settings, and count the anchor points in each file.
Done looks like
A public repository with the drawings and a license for each, outputs from at least 5 tools, a script that reproduces every score, and one table giving, per tool, the median SSIM and LPIPS, the median number of anchor points and the median fix time from at least 3 illustrators doing at least 20 files each.
Open 02 · the brief
Credentials on artists' platforms
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tally of whether Content Credentials survive upload. These are signed records, under the C2PA standard, of who made an image and how. Upload human-made illustrations to the sites illustrators use, such as Behance, Dribbble, ArtStation, DeviantArt, Adobe Stock, Shutterstock, iStock, Instagram, Pinterest, LinkedIn, X and Etsy, and check for each site whether the record is still in the file the site serves, still in a downloaded copy, and shown to viewers.
Why it is still open
Earlier tests covered other things. IPTC, the news media's standards body, last tested older photo metadata, not C2PA, in spring 2019. Indicator checked AI labels on 516 AI-made posts across 5 social platforms in October 2025 and found only 169 correctly labeled, and its March 2026 update still covered only AI-made content on social platforms. None of these tested human-made pictures or stock and portfolio sites.
Built with what exists
c2patool, an open-source command-line tool from the Content Authenticity project that adds and reads C2PA records; contributor accounts on each site; a spreadsheet; optionally a web agent such as Browser Use 1.0 to repeat the uploads and downloads, within each site's terms (see the Note "Deleting 80% to start over").
First step
Sign one self-made illustration with c2patool, upload it to Behance and Instagram, download it again and read the record back.
Done looks like
A published table of at least 12 sites by 3 checks by 2 file types (JPEG and PNG), run twice a month apart, with the raw files and c2patool reports.
Open 03 · the brief
Editing tests for AI vectors
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A blind test in which working illustrators make standard client edits to SVG illustrations from AI generators and to hand-made control files: recolor to a three-color brand palette, move one element, delete the background and resize for a small icon. Record the minutes taken, success or failure, the counts of paths and anchor points, and a 1 to 5 rating for how easy the file was to edit.
Why it is still open
Current SVG benchmarks do not measure human editing effort. VectorGym (2026) judges output with an AI model, and its 17 human raters were "all technical engineers or AI and design experts," with no editability measure. LayerTracer counts paths and asked 46 "design enthusiasts" which result looked better. SVGenius uses 18 automatic measures.
Built with what exists
Recraft V4 Vector through its API; general AI models that write SVG code, such as those VectorGym tested (Gemini 3 Pro, GPT-5.1 and Claude Sonnet 4.5); Adobe Illustrator or the free Inkscape for the edits; a timer and a shared score sheet.
First step
Write 20 briefs based on common jobs (a spot illustration, an icon, a hero image for a website), generate them with Recraft V4 Vector and one general model, and time one illustrator on the recolor edit.
Done looks like
At least 5 generators by 20 briefs by 3 illustrators, so 300 or more timed sessions, with the files, timings and a table of median minutes per edit and failure rate per generator published.