The corpus

Between 28 and 31 August 2026 I gave five frozen prompts to seven image tools (ChatGPT, Gemini, Copilot, Grok Imagine, Mistral Vibe, Meta AI and Midjourney) across thirteen names for data work and four names that already carry a picture of their own, teacher, lawyer, nurse and delivery rider. Not every name got every prompt: the first four prompts ran on twelve data-work names, the thirteenth (data-enrichment-professional) was asked once, and the group-and-informed prompt ran on three chosen names. Teacher, lawyer, nurse and delivery rider took the last four prompts in August and the bare one on 3 September 2026, so that the comparison the findings rest on starts from the same point as the data-work names; those 28 cells were made between 10:58 and 12:24 that day, with the same tools at the same settings. One pairing of a tool, a job name and a prompt is a cell. I archived and coded whatever each tool returned by default: one image from most cells, four from every Grok and Midjourney cell, and two from each of the two ChatGPT cells that returned an A/B pair. That is 938 images in 504 cells. An eighth tool, Claude, was run separately. It does not predict pixels; it writes SVG. Its 72 images are a different object and enter no table, test or mean of the main analysis. The corpus on this page is all of it, 1,010 images, so that the excluded images can still be looked at.

The unit is the generation event: one prompt, submitted once, to one tool, logged with its model version, session mode, timestamp and the exact text sent. Nothing was chosen for looking better, and nothing was quietly replaced. Where a tool failed to return an image at all and the cell had to be submitted again, the reason is on that image's own page. Six Mistral cells, one ChatGPT cell and one Copilot cell carry such a note.

The parallel corpus

On 3 September 2026 I also collected what the companies behind these tools publish of the people who work for them. One image qualified when it was a photograph or a video still on the company's own site, showing people who work there, at work or at a company event, at most four per company. Pictures of users, customers, demo subjects, external scientists and other trades were out, and so were illustrations, renders and generated images. Eight sites were inventoried page by page: ai.google, completed from deepmind.google, the department behind Gemini; openai.com; anthropic.com; x.ai; midjourney.com; microsoft.ai, the Microsoft AI division that makes Copilot, because microsoft.com/ai shows generic professionals only; mistral.ai; and ai.meta.com. Six yielded photographs, 22 in all: Google 4, OpenAI 4, Anthropic 4, Microsoft AI 4, Mistral 4, Meta 2. Two yielded none: x.ai shows no photograph of any person, midjourney.com no photograph at all. One further image sits apart from the count, a staff portrait on Microsoft AI's careers page served from a file named as a Gemini generation, filed as an extra case because nothing in the pixels decides what it is. Each photograph is logged with its page, its alt text, the size the site served it at, the time it was retrieved and a checksum, and each was coded the same day with the same grid, with three adjustments that the last finding declares. They are reproduced here at the served size, framed so that they cannot be taken for a generated image, and each one links to the page it was taken from.

These systems change monthly; this is a dated snapshot, built to be run again.

The five prompts

Each set is one template with the job name substituted in. Plain names are used everywhere on this page; the template each one sends is printed underneath. Elsewhere on the page and in the reports the same five are referred to by shorthand: the bare prompt (s1), the bare group prompt (s2), make it realistic (s3), the informed prompt (s4) and the informed group prompt (s5). The last two are the ones that cast the tool as a photojournalist working from sociological research.

The Claude run received the same five templates unchanged, each with one constant sentence in front: “I know you do not generate photographic images. Do your best anyway: produce the closest equivalent you can in graphic form, using whatever visual means you have.” The preamble allows a different medium. It does not change the register, meaning the tone and the manner of address, which is what the study measures. Tools without an aspect-ratio control also received a constant suffix, “Use a 16:9 aspect ratio.”, identical in every one of that tool's cells.

The Semiotic Audit Grid

Every image was coded once, between 31 August and 3 September 2026, with the Semiotic Audit Grid v1.0, an instrument I wrote for this paper; coding assisted by Claude (Cowork). Every field is a question with a closed list of answers, arranged in tiers:

• Tier 1, what is shown. How many people, whether a face is visible, how gender, age and region are coded, where the scene is and in what condition, how many screens, what the equipment is like, the mood and the class conveyed, and a free note on everything else in the frame.
• Tier 1b, the reality checklist. Seven yes-or-no indicators, each taken from a documented finding in the platform-labour literature. Their sum is the Reality Alignment Score.
• Tier 1c, idealization. No extra coding: an index computed from five fields already filled.
• Tier 1d and 1e, the group sets. On the two prompts that ask for a group, whether a collective is actually shown, and what kind it is, and separately whether the photograph could have been taken at all.
• Tier 1f, the narrative layer. What the subjects are doing in one plain clause; which moment of the story is staged; whether the one who gives the orders has been given a body inside the frame; and which modality governs the scene.
• Tier 1g, outliers of colour, shape and layout. The corpus norm of colour, shape and layout was written down after the first hundred images and frozen; only departures from it are flagged and described.
• Tier 2 and 3. How the image is built, in Westberg and Kvåle's terms, and, on the images chosen for the talk, what it naturalizes and what it erases.

The corpus was coded in one pass, coding assisted by Claude (Cowork); 88 images were then coded a second time, blind, taking every tenth filename of the 885 images of the first coding pass. Agreement was 89.6 per cent across the 40 closed fields common to both passes. On 2 September 2026 two fields that had been used only as exception flags outside the bare prompt, whether the image shows the named job and whether it takes the metaphor in the name literally, were coded again blind: first on a stratified sample of 110 images from the four later prompts, then on all 156 images of the bare group prompt, of which 84 depict the job, 54 per cent against the 63 per cent the first pass had recorded. Thirty images where the two passes diverged were looked at a third time. The results before and after these passes are kept frozen beside the coding sheet, together with a sensitivity analysis restricted to the images that depict the job, considered and declined as a headline. Those rechecks are the only guard against a single reading, and they are reported as such. Five fields fell below 80 per cent and are treated as descriptive rather than as evidence; they are named in the findings report under its limits. The 53 comparison images at the bare prompt, the four Claude drawings that match them and the 23 photographs were coded in one pass by the same procedure and were part of no recheck. The region field records only what an image shows, is reported in aggregate, and is never a claim about any real person's identity.

Four things to declare

• The sessions are not neutral. Most of these products will not generate images in an anonymous or temporary chat, so personalization was switched off by settings rather than excluded by construction. Grok is the exception, and it will generate in a private chat; I did not take that route, partly for consistency with the other six and partly because the run was already under way. The Claude images were made in anonymous chats, because by then I had spent weeks putting this material through my own account, and its memory of it was no longer neutral. So the limitation is real but narrower than a clean claim about the industry: what this corpus sees is close to what a signed-in user sees, and part of that is my choice rather than theirs.
• Two routes into ChatGPT. Most of its cells in the main analysis were submitted through Codex rather than the web app, so those conversations cannot be checked afterwards for whether the tool searched the web, the way the others can.
• Midjourney publishes by default. Stealth generation belongs to its Pro tier and above. These 288 images were made on a Standard account, so they were public on the tool's own Explore feed as they were generated, and may still be findable there. That is the ordinary condition for most of its users, and it is the condition this corpus was made under.
• Delivered format. All 1,010 generated images were requested at 16:9, by interface control or by a constant suffix. Two tools returned a different shape throughout, Copilot at 3:2 and Mistral at 4:3. Between them that is 144 of the 362 images where the ratio had to be asked for in words. Gemini's ratio of 1.792 is under 1 per cent off the shape requested, and is left as it is. Those departures from 16:9 are themselves part of the record, so nothing was re-cropped, and nothing is cropped in the gallery either. The 23 photographs were requested at no ratio at all: they are shown at the shape and size the sites served them.

The words, and what they mean

Plain labels are used throughout the page. The technical name and the raw coded value are always one click away, under “show technical names” on any image's page.

How to read this page

• Gallery. Every image, filterable. Values inside one filter combine with OR; different filters combine with AND. A blank means the field was not coded, and a blank never matches a score range. Source separates the three kinds of image on the page: generated by one of the seven tools, drawn as code by Claude, or photographed and published by a company. The photographs carry a frame and a label on their tile and a provenance band on their page, they have no prompt, job name or prompt set, and the Company filter reaches them directly.
• Any image. Click it for the full picture, the prompt that made it with its timestamp, and every field that was coded on it, grouped and plainly named.
• Compare. Two to six images side by side, with the rows that differ highlighted. Six, because one of the report’s own comparisons uses six images.
• Findings. What the images show, each with its question, its images and the numbers behind it.
• Census. The names the sector uses for this workforce, in English, French and Italian: who coined or spread each one, in what register, what it connotes, and where that was checked. For the thirteen names the corpus was built on, the census also says how often the tools drew the job at the bare prompt, with a link to those images.
• Links carry state. The address bar holds the view, every filter and the open image, so any selection can be sent to somebody as a link.
• Keyboard. With an image open, ← and → move to the previous and the next image and Esc closes it. Anywhere on the page, / puts the cursor in the search field.

Credits

Author and analyst

Lorenzo L. D. Incardona, lldincardona.com

The paper

“The False Consciousness of LLMs: Representations of Ghost Workers in Artificially Generated Images”, written for the INDL-9 conference, the ninth conference of the International Network on Digital Labour, on the theme “AI Supply Chains: Building an interdisciplinary research agenda for AI and labour”, ILO, Geneva, 09-11 September 2026. www.indl.network/indl-9

Coding instrument

Semiotic Audit Grid v1.0. Coding assisted by Claude (Cowork).

Terminology census

Compiled 27 August 2026 from web-verified desk research, assisted by Claude (Cowork, model Fable 5). Sources are linked term by term; items that could not be verified are listed as such.

Images

Generated by ChatGPT, Gemini, Copilot, Grok Imagine, Mistral Vibe, Meta AI, Midjourney and, in a separate run, Claude (SVG rendered to PNG). The images are synthetic: no real person is portrayed. They are reproduced here as research evidence, under each tool's own terms, and re-encoded to JPEG for the web; the archived originals, with their provenance metadata, stay with the author. Visible watermarks and content credits were not removed. The 23 photographs of the parallel corpus were published by Google, OpenAI, Anthropic, Microsoft AI, Mistral and Meta on their own sites and are reproduced here at the size the sites served them, unaltered, as research evidence, each linked to the page it was taken from; they remain the property of the companies that published them.

This webpage

Built with Claude Code, models Claude Opus 5 and Claude Fable 5.1, in ultracode mode.