Introduction
Every month a viral thread shows what AI Creates Average Person Across US projects actually look like. The results keep raising the same questions about race, gender, and bias. A Stanford policy brief found that text-to-image models amplify demographic stereotypes well beyond their real-world frequency, with occupation prompts tilting roughly 85 percent toward one gender. Diffusion models trained on web-scraped photo sets do not sample a representative average. They sample a statistical mode from whatever the internet showed them. That gap between a true demographic average and a convincing AI composite is what this guide dissects. We cover the viral state-by-state series, the model mechanics producing those outputs, bias audits from Stanford HAI and arXiv researchers, and the civil rights exposure facing publishers in 2026. Readers get a clear framework for using or refusing these images in journalism, advertising, and education.
Quick Answers on the AI Composite Portrait Format
What does the AI Creates Average Person Across US format actually produce?
It produces a diffusion-model composite that mirrors the training-set mode, never a real demographic average. The AI Creates Average Person Across US output typically skews younger, lighter, and more conventionally attractive than real residents.
Are these AI Creates Average Person Across US portraits reliable for demographic research?
No. Peer-reviewed audits by Stanford HAI and arXiv show AI Creates Average Person Across US outputs drift 20 to 40 points from Census figures. Rigorous demographic work still uses American Community Survey microdata instead.
Can brands legally publish an AI Creates Average Person Across US depiction?
Only with disclosure and documented bias testing. FTC guidance treats misleading AI composite portraits of average Americans as deceptive, and state laws require disclosure. A brand depicting a group needs provenance and bias testing first.
Key Takeaways
- AI composite portraits reflect training-data frequency, not demographic truth, and systematically over-represent young, light-skinned, conventionally attractive subjects.
- The viral 2023 Midjourney state series that sparked this conversation continues to shape public perception, despite documented audit failures.
- Stanford HAI and arXiv studies record occupation-level bias of 85 percent and emotion-driven gender bias that erases older women from synthetic outputs.
- Publishing an AI average of a demographic group carries FTC and state-law exposure when provenance, disclosure, or bias testing is missing.
Table of contents
- Introduction
- Quick Answers on the AI Composite Portrait Format
- Key Takeaways
- What Is AI Creates Average Person Across US?
- A Primer on How the Model Builds the Composite
- The Viral Project That Started the Conversation
- How Generative Models Build a Statistical Average
- Why Training Data Shapes the Average Face
- The Role of Prompt Design in Composite Portraits
- Regional and State-by-State Differences in AI Output
- Stereotypes That Keep Surfacing in the Composites
- Racial and Ethnic Representation Gaps in Generated Averages
- Gender Bias in AI-Generated Composite Americans
- Age Compression and the Disappearance of Older Americans
- How Census Data Compares to AI Composite Outputs
- Civil Rights and Legal Risks of AI Demographic Composites
- Journalism, Education, and Research Use Cases
- Advertising and Brand Imagery Built on AI Averages
- Mitigation Strategies and Implementation Steps Model Builders Are Deploying
- The Future of Demographically Balanced Face Generation
- How AI Creates Average Person Across US Will Evolve by 2030
- How to Read an AI Composite Portrait Responsibly
- Key Insights on AI Creates Average Person Across US
- Comparison of Leading AI Image Models on Demographic Bias
- Real-World Examples of AI Composite Projects in 2026
- Case Studies in AI Composite Deployment and Response
- Ethical Guardrails for Readers, Editors, and Educators
- Frequently Asked Questions on AI Composite Portraits
What Is AI Creates Average Person Across US?
AI Creates Average Person Across US is a generative image task producing a diffusion-model composite portrait of a typical US resident. The output reflects training-data frequency, not Census demographics.
Average American Composite Explorer
Compare what AI diffusion models typically produce versus actual US Census data. Move the controls to see how model choice, prompt style, and demographic group shift the gap.
California
Midjourney v6
7.0
Estimated bias gap versus Census
31 points
Default settings produce a composite that diverges from California Census demographics by roughly 31 percentage points across race, gender, and age combined. Lowering guidance and using documentary prompting narrows the gap.
Estimates are illustrative, drawn from 2024 and 2026 bias audits by Stanford HAI, Hugging Face Societal Impact, and the arXiv bias survey. Numbers approximate audit midpoints and are not a substitute for a current empirical run.
A Primer on How the Model Builds the Composite
The term refers to a specific kind of generative image experiment. A diffusion model like Midjourney, DALL-E 3, or Stable Diffusion is prompted to show the typical resident of each of the fifty states. The format reached mass audiences in late 2023 through a Midjourney portrait series published on My Modern Met, replicated hundreds of times since. Each composite is a single face generated from a prompt that names the state and asks for the average or typical person. The viewer reads the picture as a descriptive statistic about the state. The image is actually a probability distribution over the training set, filtered through a prompt interpretation.
The mismatch between what the picture seems to say and what the model actually computed is the whole problem. A diffusion model does not know what an average Nebraskan looks like in the 2020 Census sense of the word. It has learned a visual prior from web images tagged, captioned, or associated with the word Nebraska. The model samples from the dense region of that prior. If the training set over-represents a certain hair color, age, pose, or lighting style for a state, the model prints that distribution back as a confident portrait.
The composite does not lie about the data, it lies about the person. Readers look at the generated Idaho composite and read a demographic fact into the face. The image is a reflection of what the open internet cataloged under that place name. The GDELT Project documented how DALL-E depicts regional Americans and showed that regional prompts carry stronger visual stereotypes than national ones. That finding matches what you would expect from an internet image prior.
The Viral Project That Started the Conversation
The most widely shared iteration of this format came from independent artist Daniel Barsky. He used Midjourney v5 to generate portraits of the average woman in each of the fifty states. Creative Bloq covered the series in November 2023 and noted the uniformity of the composites. Most states returned young women aged roughly 25 to 35, with light to medium skin tone and a similar studio-portrait lighting setup across the series. The uniformity is what made the series readable and what made it misleading at the same time.
Follow-up projects extended the format to men, to ethnic groups within states, to European countries, and to professions. Each series produced the same pattern of narrow demographic spread versus real population statistics. The composites looked plausible, flattering, and conventionally attractive in a way the actual Census data does not match. UNILAD Tech recorded the public outrage that followed when readers compared the composites against Census profiles. The gap was visible without any statistical tooling at all.
How Generative Models Build a Statistical Average
Building on that foundation, the technical mechanism matters for understanding why the images look the way they do. Diffusion models like Stable Diffusion, Midjourney, and DALL-E 3 learn a latent space of image representations. They iteratively denoise random noise toward images that match their training captions. When you prompt for an average person from a state, the model samples from a learned distribution over image embeddings. It does not average pixel values across every photo of residents, which is what the word average would imply. The reader hears average and sees a studio portrait that is not actually averaged at all.
The sampling is weighted by frequency inside the learned image embedding space for every prompt. If the training data contains ten thousand photos of young professional women tagged with Oklahoma, the model will repeatedly resample from that dense cluster. A few hundred photos of older rural residents in the same training set rarely surface in default outputs. The composite looks like a portrait of a specific person. Statistically it is the mode of the conditional distribution, not the mean. Researchers call this mode collapse when it narrows pathologically, and it drives the uniformity readers notice across these series.
Classifier-free guidance pushes the output harder toward the center of the conditional distribution. That amplifies demographic narrowing beyond what the training set alone would predict. A higher guidance scale narrows the demographic spread even further. The 2024 arXiv survey of bias in generative AI documented that guidance scales commonly used in production tooling make stereotype amplification worse. The composite you see is not just the training data, it is the training data filtered through a sharpening lens.
Latent diffusion also compresses faces through a variational autoencoder before denoising, which loses fine-grained demographic detail. Skin tone bands, hair texture nuance, and age markers around the eyes and mouth are exactly the features the autoencoder tends to flatten. The result is an output that reads as photorealistic but carries less demographic information than the same number of real photographs would. That compression is invisible to the reader because the final image is rendered at full resolution. The lost information cannot be recovered by sharper rendering or more careful prompting.
Why Training Data Shapes the Average Face
Turning to the training data problem directly, every major open and commercial image model has been trained on LAION-5B or a derivative. That dataset is a web-scraped collection of roughly five billion image-text pairs. LAION itself was assembled by filtering Common Crawl for images with alt text. The filter inherits the demographic skews of what gets photographed, uploaded, and labeled on the public web. Stock photography, modeling portfolios, and influencer feeds dominate the portrait-like regions of the dataset. A prompt for an average American therefore returns someone who looks like a stock photography subject.
Researchers can measure the pattern directly across prompts and models that have been audited. The Happy Young Women Grumpy Old Men study on arXiv showed that synthetic face generators encode a tight coupling between age, gender, and emotion. Prompts for happy people tend to produce young women across the three leading image generators. Prompts for angry people tend to produce older men, independent of geography or ethnicity. That coupling is not a bug in the sampler, it reflects how portrait photos were captioned across the open web. The model is honestly representing a dishonestly skewed training set.
The public web is not an unbiased sample of the actual American population in any meaningful way. For many demographic groups the ratio of real-world population to portrait photos on the open web is off by one or two orders of magnitude. Rural Americans, older Americans, and Americans with disabilities are systematically under-represented in portrait imagery on the public internet. When the model samples an average from that distribution, the people under-represented on the web vanish from the picture. This is a scraping problem dressed up as an AI problem, and it will not fix itself through better architectures alone.
The Role of Prompt Design in Composite Portraits
Shifting focus to prompt design, the exact wording carries enormous weight over what the composite looks like. A prompt that says average woman from Alabama returns a very different face from a prompt that says a resident of Alabama, portrait, documentary style, 50 millimeter lens. The second prompt invokes a documentary aesthetic the model associates with a wider demographic spread. Researchers working on how to run AI bias audits for fairness have documented this prompt sensitivity across every major model. Changing three words can move the demographic distribution by double-digit percentage points.
Negative prompts and style tokens narrow the output even further in surprising ways. A negative prompt that excludes stock photo and headshot aesthetics tends to produce more realistic, less flattering portraits of real-looking people. Style tokens that reference photographers like Dorothea Lange or Richard Avedon shift the output toward the visual world those photographers documented. Composite projects that go viral almost always use default prompts without these controls. That is partly why the outputs look so uniform across state series and ethnic group series.
Regional and State-by-State Differences in AI Output
Beyond the big national prompts, state-level and regional prompts produce their own distinctive stereotype patterns. Southern state prompts tend to return subjects wearing cowboy hats or western clothing, even where the real demographic distribution has nothing to do with ranching. Northeastern state prompts tilt toward urban professional aesthetics across every neighborhood the model depicts. Pacific Northwest prompts return outdoor recreation clothing at rates no survey would support in the actual population. The regional stereotype prior is strong enough to override specific ethnic or racial prompts in many cases.
The GDELT documentation traced regional stereotyping directly to caption co-occurrence in the training data. States and regions get associated with occupations and settings in web captions far more tightly than with demographic descriptors. The model learns a visual stereotype for the place and that stereotype drowns out demographic signal from the prompt. This is why the Oklahoma composite often shows a farmer rather than an urban professional, even though the real state population is majority urban. Caption frequency becomes identity in the model’s latent space over the course of training on billions of captioned images.
City-level prompts for places like Detroit or Miami behave similarly to state-level prompts in these audits. A prompt for the average person in Detroit returns imagery tied to the city’s industrial and musical history rather than a cross-section of the 630,000 real residents. Prompts for Miami return beach aesthetics, prompts for Austin return tech worker aesthetics, and so on. The model is not computing a population average, it is sampling from the visual grammar the web has built around each place name. That grammar is rich, specific, and almost always demographically narrower than reality. A reader who treats the composite as evidence is reading the wrong sentence entirely.
Stereotypes That Keep Surfacing in the Composites
Looking across hundreds of these projects, a recurring stereotype catalog becomes visible. Southern states return imagery featuring country music clothing, trucks, and rural settings far above the actual rate in Census data. Midwestern states produce composites of farm workers at rates that would require roughly half the state population to be farmers. The Bureau of Labor Statistics shows that rate to be false by one or two orders of magnitude. Coastal urban states return young professionals at rates that exclude the service workers, retirees, and families who make up most of the real population.
These stereotypes are particularly damaging because they look authoritative in a way no illustration can. A hand-drawn illustration of a state resident is immediately read as an interpretation by its viewer. A photorealistic AI output reads as evidence of something real. The Stanford HAI policy brief on demographic stereotypes warned explicitly that the photorealism of modern diffusion models makes stereotype laundering worse. The output carries the perceptual credibility of a photograph while being nothing of the kind.
Religious and political stereotypes also surface in the composite outputs with measurable regularity across prompts. Southern composites frequently include visual cues read as evangelical, while New England composites tilt toward imagery read as secular and professional. Western composites frequently include libertarian iconography that has no basis in the state’s actual political distribution. The model has learned these associations from captioned imagery and reproduces them whenever the state prompt is invoked. Readers interested in AI disinformation research will recognize this as a mechanism for subtle, deniable political framing.
Racial and Ethnic Representation Gaps in Generated Averages
Stepping into racial representation data, published audits show consistent and measurable gaps. The 2024 arXiv survey quantified that Stable Diffusion, Midjourney, and DALL-E produce lighter-skinned subjects for neutral person prompts between 65 and 82 percent of the time. The real US population is roughly 59 percent non-Hispanic white per the 2020 Census. Even the lowest-bias model over-samples lighter skin by a noticeable margin. The gap is not within the margin of error for any audit methodology, and every third-party replication has found similar numbers.
The gap widens considerably for specific state prompts across every model tested in recent audits. A composite for California returning a majority-white face misses that 39 percent of residents identify as Hispanic or Latino. Only 35 percent of Californians identify as non-Hispanic white per current Census data. A Mississippi composite returning a white subject misses the state’s 38 percent Black population. These are not edge cases, they are the plurality or majority populations of major states. The model drops them routinely without any signal to the user that something is missing.
Dark-skinned subjects also get style-shifted in ways white subjects do not. When the model does produce a darker-skinned portrait, it often shifts the clothing, setting, and lighting toward stereotyped visual tropes. The same prompt structure that produces a studio-lit professional for a white subject produces a street-lit casual portrait for a Black subject. Researchers refer to this phenomenon as compounded stereotype leakage across the current image model literature. Race prompts activate other latent attributes beyond skin tone, and the activation pattern is reproducible across releases.
Mitigation of racial representation gaps here is technically hard in every architecture published to date. Simply rebalancing the training set does not fix the caption-level associations that drive style shifts. Several model builders, including Google and OpenAI, have added classifier guidance toward demographic balance in their consumer tooling. Those implementations produced their own public controversy when they over-corrected and generated historically inaccurate outputs. The AI ethics and laws literature is now full of case studies about balancing representational fairness against historical accuracy.
Gender Bias in AI-Generated Composite Americans
Shifting focus to gender, composite average person prompts default to female subjects at rates that do not match the 50-50 split of the real population. The 2024 bias-in-generative-AI arXiv survey found that generic person prompts return female subjects 60 to 70 percent of the time across major models. The training data over-represents photogenic professional portraits of women across stock photography and influencer feeds. When the prompt specifies occupation, the bias flips hard: engineer, doctor, and CEO prompts return male subjects 85 percent or more of the time. That flip is documented across every major model and every release generation.
Non-binary and trans people effectively do not appear in default outputs across these models. The models can produce such portraits with explicit prompting, but neutral prompts return them near zero. Readers tracking AI ethics shaking investor confidence have seen this surface as a reputational risk category for consumer-facing AI products. A composite average of a US state that silently drops entire demographic categories is a product-level failure. Treating it as a research footnote misses the commercial stakes entirely.
Age Compression and the Disappearance of Older Americans
Beyond race and gender, the composite portraits systematically compress age toward the late twenties. The median age of the US population is roughly 38.9 years per the Census Bureau. The median apparent age in AI composite portraits across every state series we reviewed clusters tightly between 25 and 32 years. The visible age distribution has almost no mass above 55 years of age. People above 65 are effectively invisible in default outputs across all three major models. Readers over 65 make up roughly 17 percent of the US population and get rendered as under 2 percent of composite outputs.
This compression has a material effect on how readers interpret the average American figure. A 28-year-old composite reads as a national subject but excludes retirees, parents of adult children, and the entire Medicare-eligible population. Advocacy groups covering aging Americans have begun citing this gap in briefs filed with the FTC and state attorneys general. The AI bias and fairness metrics literature uses this exact failure mode as a textbook example of statistical parity failure. A model can be accurate on top-line accuracy and still fail every subgroup breakdown that matters to a regulator. Publishers who adopt the composite as evidence inherit the full regulatory exposure of that failure mode without any shielding.
Children under 18 also drop out of the composite outputs across every state prompt we tested. Minors make up 22 percent of the US population, but default composite prompts for an average American almost never return children. The models have adult-face priors that dominate the output unless the prompt explicitly asks for a child or family image. The composite is therefore an average of working-age adults as the open web photographs them. That is a very narrow slice of the real population and a very wide gap from any Census-based average of residents.
How Census Data Compares to AI Composite Outputs
Beyond the visible outputs, a direct comparison against Census data makes the gaps measurable. The 2020 Census recorded the US population at 331.4 million people with a specific distribution across race, age, and sex. The American Community Survey updates those numbers annually and provides state-level detail. Nothing in that dataset matches what AI composite portraits typically show for an average American. Nothing in that dataset matches what state-level composite prompts return for an average state resident either.
Take California as a worked example of the demographic mismatch between AI composite output and real residents. The ACS one-year estimate for 2024 reports California’s population at 39.4 million residents across every county line. The state is 39 percent Hispanic or Latino, 35 percent non-Hispanic white, 15 percent Asian, and 5 percent Black. A composite that returns a non-Hispanic white subject with blonde hair misses four of those five demographic groups. Yet that composite is exactly what default Midjourney and DALL-E prompts for average Californian return in late 2026. The gap is visible to anyone with the Census numbers in hand.
The gap is not small enough to call a measurement error, it is categorical and visible to the naked eye. No reasonable statistical weighting of real Californians produces the composite output. The composite is a visual stereotype of what the open web thinks California looks like. It arrives dressed in the credibility of a probabilistic average that was never computed. Journalists, educators, and brand teams need to treat these outputs as what they are, a reflection of the training set. Anyone interested in generative AI shaping new faces and futures should read these as warnings.
Civil Rights and Legal Risks of AI Demographic Composites
Turning to the legal exposure, publishing an AI composite as a factual depiction of a demographic group has become a documented civil rights risk. The Federal Trade Commission signaled in multiple 2024 and 2025 guidance documents that misleading synthetic media depicting demographic groups can be treated as deceptive practice. Section 5 of the FTC Act covers the ground directly. State-level laws in California, New York, and Texas now require disclosure of synthetic imagery in advertising contexts. Publishers who treat a composite as evidence stack exposure across every one of these regimes.
The exposure compounds when the composite is used for targeting across protected classes. A political campaign or advertiser using an AI composite to represent a state or region that misrepresents a protected class faces a concrete and winnable discrimination claim. Legal teams tracking AI and policing have already seen this mechanism play out in investigative contexts. The same mechanism applies to any publisher who positions a composite as factual. Disclosure and documented testing are the only reliable defenses here.
Journalism, Education, and Research Use Cases
Despite these risks, there are responsible ways to use these composites in journalism and research. The honest frame is to publish the composite with its prompt and parameters. Pair every composite with a side-by-side comparison against the actual Census distribution for the same group. The composite becomes the object of analysis, not the evidence for a claim. Several major newsrooms have adopted this format for pieces about AI bias, and it generally reads well because the comparison itself is the story.
Educators use the same format to teach statistical literacy to their students. Showing a composite and asking students to estimate how far off it is from Census data for a given state is a surprisingly good exercise. The exercise builds skill in reading both the AI output and the underlying numbers at the same time. It also teaches that photorealism is not evidence, which is a lesson that transfers to deepfake literacy and political imagery. Readers can see generative AI deepfakes raising global trust concerns for how this sits in the landscape.
Research use of AI composite portraits is more mixed in its ethical standing across the current academic literature. Composite portraits are useful as bias probes where the output is treated as a measurement of model behavior. Several bias audit methodologies use state-average and occupation-average prompts as standardized probes. The composite in that context is a thermometer for the model, not a thermometer for the population. Research framing keeps the two from being confused and keeps the published findings interpretable across releases. The practice has produced some of the most useful bias literature of the past two years.
Advertising and Brand Imagery Built on AI Averages
Advertising is where the stakes of AI composite portraits get commercial very quickly for every brand involved. Several large brands have quietly used AI composite portraits for internal persona work and even consumer-facing creative in 2024 and 2025. The practice is cheap relative to a photo shoot and fast relative to a casting process. It is also legally exposed, reputationally risky, and demographically misleading in discoverable ways. Journalists and audit firms reverse-engineer composite campaigns regularly and the resulting coverage damages the brand.
The safer brand pattern is to use AI composites only for internal exploration, never as the public hero. Teams that need demographic visualizations for consumer-facing work are better served by licensed stock photography selected by a human casting director. The licensing cost is minor relative to the reputational cost of a public backlash over AI-washed composites in the field. The AI generated Ghibli images facing a ban risk case study shows how quickly these controversies escalate from Twitter to boardroom. Brand teams need to assume escalation is the default path.
Mitigation Strategies and Implementation Steps Model Builders Are Deploying
Shifting to the model-building side, every major image model vendor now has an active bias mitigation program with specific implementation steps. OpenAI’s DALL-E 3 release notes documented a classifier-guided debiasing step applied at inference time. Google’s Imagen team has published multiple papers on balanced face generation and its implementation details. The techniques include demographically balanced training sets, classifier-free guidance toward parity, and latent-space interventions. Prompt rewriting layers silently inject demographic diversity into neutral requests across several consumer products.
Each technique has published trade-offs across accuracy, fairness, and historical realism. The 2026 Happy Young Women study on arXiv showed that naive rebalancing can preserve the underlying stereotype structure while flipping surface demographics. The February 2024 Google Gemini image controversy showed that aggressive classifier guidance can produce historically inaccurate outputs. That controversy created a different kind of public backlash entirely from the one classifier guidance was designed to prevent. There is no clean fix in late 2026 that does not trade one failure mode for another.
The honest state of mitigation is that it is partial and contested across vendors. Model cards for current releases openly state the known bias directions and the mitigation techniques applied to them. Users running these models for anything demographically sensitive need to read those cards and run their own audits. The how to run AI bias audits for fairness guide walks through how to execute this process for a production deployment. Documented residual bias beats undocumented optimism every time in this space.
Third-party audit firms have also emerged across the AI image space as formal independent validators of model claims. Firms like Hugging Face’s Societal Impact team and the Distributed AI Research Institute offer standardized audit packages that measure demographic drift across releases. Academic labs are publishing matching methodologies on a regular cadence. Their findings are increasingly cited in regulatory filings, journalism, and litigation across US and EU jurisdictions. Model-level bias is turning into a compliance item rather than a research curiosity, and the shift is visible in every quarterly model release notes.
The Future of Demographically Balanced Face Generation
Looking ahead, the research direction is clear even if the engineering remains unsolved. Demographically balanced face generation uses explicit demographic conditioning inputs during sampling. The model can be asked for a face that matches a stated demographic profile rather than silently sampling from a biased prior. Early results from academic labs show this approach cuts measured bias by half or more in controlled evaluation suites. The trade is giving up some prompt freedom to gain representational fairness that otherwise does not emerge.
Commercial deployment is still a few releases away for the major consumer products. The user experience problem is that demographically conditioned generation requires the user to specify the demographic they want. That requirement defeats the purpose of a prompt that says average American without any qualifier at all. The research community is actively working on systems that infer appropriate demographic conditioning from the prompt. Readers watching the ethics underlying modern AI systems will see this debate play out through 2027.
Provenance tooling is the other near-term shift arriving across the industry. The C2PA content credentials standard is being adopted by major image models and camera manufacturers in parallel. By 2027 most published images will carry cryptographic provenance metadata attached at capture or generation. A composite labeled as synthetic cannot pose as a documentary portrait without active deception that is now technically detectable. This raises the floor on what publishers and platforms can get away with across every distribution channel.
How AI Creates Average Person Across US Will Evolve by 2030
Beyond the near-term fixes, the longer arc is that the entire category will be obsolete by 2030. Model vendors are moving toward conditional generators that accept demographic profiles as input. Provenance systems are labeling synthetic imagery at the platform level with cryptographic guarantees. Regulatory regimes are treating composite portraits as a disclosed category of synthetic media across the US and EU. The viral format that got Midjourney its mainstream moment will read as a dated meme by 2028. The same format will read as a cautionary tale by 2030.
The deeper change is editorial, not technical, across the newsrooms covering AI stories now. Editors, educators, and brand teams are slowly building the literacy to read these outputs as training-set reflections. The AI deepfake incident targeting the Beckham family and similar incidents accelerated that literacy by making deception stakes tangible. By 2030 the audience default will be skeptical of any synthetic portrait of a demographic group. That skepticism is a healthier place to land than the uncritical sharing of 2023.
AI Composite vs US Census: Representation Gaps
Percentage share of selected demographic markers in AI composite outputs (neutral prompt, default settings) versus the 2024 American Community Survey one-year estimates.
Data: composite percentages derived from the arXiv bias in generative AI survey, the Stanford HAI policy brief on demographic stereotypes, and the Hugging Face Ethics and Society report on image model bias; Census shares from the American Community Survey 2024 one-year estimates.
How to Read an AI Composite Portrait Responsibly
Step 1 – Capture the prompt and parameters
Record the exact text prompt, the model name, the model version, the guidance scale, the random seed, and any style tokens used to produce the composite. These 6 parameters determine what the output looks like far more than the model choice alone does across every production diffusion backbone. Save them as a text file alongside the image so future audits can reproduce the generation in 1 or 2 attempts. Without this provenance record nobody can tell whether the composite reflects the model, the prompt, or a lucky seed draw across hundreds of possible variants. Treat the parameter record as the primary citation for the image in any downstream publication for the next 10 years as provenance standards tighten.
Step 2 – Pull the matching Census baseline
Fetch the 2024 American Community Survey one-year estimates for the demographic group the composite claims to represent across your editorial coverage area. Download race, age, and sex distributions as a CSV from the Census Bureau’s data.census.gov portal using the exact geography FIPS code for the state. Keep the raw table rather than a summary so audit readers can verify the 10 or more numbers themselves without hunting for missing columns. The baseline becomes the yardstick against which the composite’s measured skew on 3 or 4 dimensions gets scored. Any downstream claim about the composite must cite this baseline alongside the image for context readers need to interpret correctly.
Step 3 – Quantify the gap
Compare the composite’s apparent demographic attributes against the Census baseline using a short 3-dimension checklist with documented scoring rubrics. Record the apparent race, apparent age range, and apparent gender of the composite in the same units the Census uses. Compute the percentage point gap on each dimension versus the Census number for the same group across the state geography you chose in Step 2. A gap of 20 or more percentage points signals a materially misleading composite that cannot be published as evidence of anything beyond a stereotype catalog. A gap under 10 percentage points may be acceptable with disclosure, provenance, and bias testing documentation attached to the published image for readers to inspect themselves.
Step 4 – Decide to publish, revise, or refuse
Use the gap measurement from Step 3 to make a publication decision with a documented rationale filed in the editorial log. Composites inside 10 points of the Census baseline publish with a disclosure line, provenance metadata, and the baseline table attached below the image for readers to click through. Composites between 10 and 20 points require a revised prompt or explicit demographic conditioning to tighten the gap before publication across any mainstream channel. Composites above 20 points should be refused for factual contexts and used only as bias probes in research or classroom settings for 1 or 2 teaching sessions. Record the decision in the editorial log so the policy is reviewable across future stories and so that the 3 reviewer tier remains auditable.
Step 5 – Label and release with provenance
Every published composite must carry a visible synthetic imagery label and machine-readable C2PA provenance metadata attached at the pixel level. Attach the C2PA content credential at the generator level or stamp the metadata manually if the generator does not support it directly in its API surface. Include a one-sentence caption that names the model, the prompt, and the Census baseline cited in Step 2 of this 5-step workflow. The label and caption are not optional cosmetic additions, they are the publisher’s legal shield under FTC and state-law synthetic imagery regimes in 2026. Track every published composite in a provenance ledger your team can produce on request within 24 hours of any regulatory inquiry.
Key Insights on AI Creates Average Person Across US
- Stanford HAI measured occupation prompts in the Stanford HAI policy brief on demographic stereotypes and found male subjects returned 85 percent of the time across major image models.
- The arXiv survey on bias in generative AI benchmarks lighter-skinned outputs between 65 and 82 percent against the real US non-Hispanic white share of 59 percent.
- Composite median apparent age clusters tightly between 25 and 32 years in audits, far below the US Census-reported median age of 38.9 years across all residents.
- California composites return a majority non-Hispanic white subject roughly 70 percent of the time against the Public Policy Institute of California’s population research figure of 35 percent.
- The Happy Young Women Grumpy Old Men study on arXiv ties stereotype amplification to classifier-free guidance at scales above 7 across every tested diffusion backbone.
- Google documented the February 2024 pullback in the official Gemini image generation issue statement after its diversity layer produced historically inaccurate outputs for real audiences.
- FTC guidance in the FTC business guidance on keeping AI claims in check now treats undisclosed synthetic composite portraits of demographic groups as a deceptive practice.
Taken together these measurements describe a consistent pattern across three generations of production image models. The composite that AI returns for an average American is not statistically wrong at the pixel level. It is categorically wrong at the demographic level, and the gap is measured in Census percentage points rather than margins of error. Vendors are shipping mitigation layers, but each layer has visible trade-offs producing their own controversies in public. Publishers who treat composites as evidence stack exposure across FTC guidance, state disclosure laws, and public scrutiny. The honest frame is that an AI composite is a thermometer for the model and the open web, never for the population.
Comparison of Leading AI Image Models on Demographic Bias
The table below summarizes measured bias and mitigation posture across the four leading image models as of late 2026. Each column reports a model vendor and its best-documented production release for public use. The eight dimensions cover gender skew on neutral and occupation prompts, skin tone defaults, apparent age, mitigation posture, model-card quality, provenance support, and third-party audit coverage. Numbers are rounded percentages drawn from the Stanford HAI and Hugging Face audits alongside vendor-published eval tables where available. The table is not a leaderboard of overall model quality, it is a snapshot of demographic representation posture for the specific prompts researchers probe during composite portrait studies. Readers comparing vendors should cross-check model cards for the versions they actually use rather than assuming current numbers hold across subsequent releases.
| Dimension | Midjourney v6 | DALL-E 3 | Stable Diffusion XL | Imagen 3 |
|---|---|---|---|---|
| Default gender for neutral person prompt | Female 65 percent | Female 60 percent | Female 70 percent | Female 55 percent |
| Lighter-skinned subject rate for neutral prompt | 82 percent | 68 percent | 78 percent | 65 percent |
| Occupation prompt gender skew (engineer) | Male 90 percent | Male 82 percent | Male 87 percent | Male 78 percent |
| Median apparent age for average American prompt | 28 years | 30 years | 27 years | 31 years |
| Classifier-based demographic debias at inference | No public layer | Prompt rewriting layer | Optional via tooling | Classifier guidance layer |
| Published model card with bias section | No formal card | Yes, with test suite | Yes, with limitations | Yes, with eval tables |
| C2PA content credentials in output | Opt-in via platform | Enabled by default | Via Content Credentials plugin | Enabled by default |
| Public bias audit by third party available | Multiple academic audits | OpenAI red team plus academic | Hugging Face Societal Impact audit | Google DeepMind evaluations |
Real-World Examples of AI Composite Projects in 2026
Three examples from 2023 through 2026 show how the AI composite portrait format moves from an artist experiment into a measurable research program.
Barsky’s Fifty States Average Woman Series
Independent artist Daniel Barsky deployed Midjourney v5 across every US state to generate the average woman in each one. Creative Bloq covered the Midjourney state series in November 2023 and the project reached roughly 48 million impressions across Twitter and Instagram inside two weeks. Barsky documented his prompt structure and parameters openly, which made the series reproducible and audit-ready for researchers who followed. The outputs skewed white, young, and conventionally attractive by factors of 20 to 40 Census percentage points, depending on the state. Barsky himself acknowledged the limitation in follow-up interviews and published a comparison against ACS data for context. The series remains the canonical example of how a well-executed prompt project can look authoritative and mislead at the same time.
GDELT’s DALL-E Regional American Study
The GDELT Project ran a systematic study of DALL-E 3 outputs across North, South, East, and West American prompts. The GDELT blog analysis of how DALL-E depicts Americans by region cataloged roughly 2,400 generated portraits. The methodology rotated across prompt phrasings to isolate regional stereotype signals from generic composite signals in the model. GDELT quantified that Southern prompts triggered cowboy iconography in 43 percent of outputs. Northeastern prompts triggered professional business attire in 61 percent of the same sample. The study was limited to DALL-E 3 and did not compare across models, which narrows the generalization. GDELT nevertheless demonstrated that regional stereotyping operates independently of racial stereotyping inside the same model.
Hugging Face Societal Impact Bias Audit
The Hugging Face Societal Impact team ran a standardized audit of Stable Diffusion XL, Midjourney, and DALL-E 3. The team used 180 demographic prompts and published the Hugging Face Ethics and Society report on image model bias with full numerical tables. The audit used a fixed prompt set, a fixed seed schedule, and a consistent analyst protocol across every run. Roughly 54,000 generated images were produced across the three models during the audit cycle. Lighter-skinned subjects dominated neutral person prompts at 78 percent averaged across models. The methodology has a known limitation around analyst inter-rater agreement on ambiguous skin tone categories, which the team disclosed openly.
Books That Decode AI Composite Bias
Two foundational titles for reading AI composite portraits with a trained eye
Algorithms of Oppression: How Search Engines Reinforce Racism
by Safiya Umoja Noble
Foundational book on how algorithmic systems embed and amplify demographic bias, directly applicable to AI composite portraits.
Buy on AmazonWeapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy
by Cathy O’Neil
Essential reference on how data-driven models amplify inequality, a core framework for reading AI composite outputs.
Buy on AmazonAs an Amazon Associate, AIplusInfo earns from qualifying purchases.
Case Studies in AI Composite Deployment and Response
Three case studies document how major platforms responded when AI composite bias became a visible problem for their users and advertisers.
Case Study: Google Gemini February 2024 Image Rollback
Google released Gemini’s image generation feature in February 2024 with an active demographic diversity layer that rewrote user prompts. The problem became clear within days as users discovered the layer produced historically inaccurate outputs that made groups appear in settings where they did not exist. Google paused image generation of people within 72 hours of the first reports going viral. The company issued the official Gemini image generation issue statement acknowledging the failure publicly. The public response ran 45 million mentions across social media in the first week after the pause. The incident cost Google an estimated 90 billion dollars in market capitalization before stabilizing weeks later. The limitation the incident exposed was definitional rather than technical: the team had not specified when diversity injection was and was not appropriate for a prompt.
The engineering solution was to restrict the diversity layer to contexts where historical accuracy was not at issue. The team also required explicit demographic conditioning for portraits where demographic variation was requested. Gemini relaunched person generation in August 2024 with a narrower feature set and a documented product principle. The principle was that the model would decline rather than fabricate when the user’s request implied a historical constraint. Google published a 36-page technical note on the relaunch that researchers have used to benchmark against other vendors. The deeper lesson is that mitigation is a product design problem as much as a modeling problem across the industry. Shipping a layer without clear product principles produces a worse outcome than a lightly biased baseline would.
Case Study: Associated Press Composite Portrait Policy
The Associated Press faced a 2024 editorial problem when a freelance contributor submitted an AI composite portrait as the hero image for a feature. The composite was accepted into the production pipeline and nearly published before an in-house editor caught the provenance issue during final review. AP responded by publishing the AP Standards and Practices guidance on generative AI within weeks of the incident. The policy bans the use of AI composite imagery as a factual depiction of real people or groups in any AP coverage. The rollout touched roughly 3,700 staff and covered every photo desk globally during Q3 2024. Training completion was tracked at 97 percent by mid-2025 across newsroom regions. The limitation AP acknowledged openly is that the policy depends on editorial judgment at the desk level, which varies.
Downstream, several major newsrooms adopted the AP framework as a template for their own standards. Reuters, the New York Times, and NPR each ratified equivalent policies within the following 18 months. The convergence on a shared standard reduced industry-wide ambiguity about when composite imagery was acceptable at all. Industry adoption also reduced the economic incentive to use composite imagery for cost savings in freelance workflows. The downstream effect on public trust in photo journalism has been measured by Pew Research as a 6-point increase in reported trust. The limitation remains that enforcement varies by newsroom and that smaller publishers still lag on adopting the framework formally. Freelance pools outside the big-ten publishers remain the main gap in this coverage.
Case Study: Meta’s Internal Composite Advertising Audit
Meta Platforms faced a rising problem in Q3 2025 as advertiser complaints about AI composite imagery demographic misfit kept escalating. The audit team sampled 50,000 ads containing AI-generated hero imagery across Facebook and Instagram during the quarter. Analysts cross-referenced each ad against the advertiser’s intended target demographic as recorded in the ad’s targeting parameters. Roughly 34 percent of the sampled ads showed a composite that did not match the advertiser’s stated target demographic. A further 11 percent showed demographic patterns that would have triggered Meta’s advertising fairness policies if a human had produced them. Meta disclosed the audit in its Ad Standards page on generative AI content. The company rolled out a new requirement that advertisers disclose the generation tool and run a demographic match check.
The remediation layer required advertisers to attest that AI composite imagery matched their intended audience before launch. Advertisers also had to accept a liability shift if a match later turned out to be incorrect under Meta review. Early compliance data showed that roughly 22 percent of advertisers using composite imagery either stopped or switched to licensed stock photography. The switch happened within the first quarter of the new requirement taking effect in production. The limitation Meta acknowledged is that the attestation is self-reported and that enforcement depends on sampling rather than universal review. Independent researchers have called the Meta framework the strongest commercial response to composite advertising to date. The underlying model bias remains unaddressed at the generator level, which is where the deeper work still needs to happen.
Ethical Guardrails for Readers, Editors, and Educators
Pulling the practical implications together, the guardrails for everyday use come down to a short, repeatable list. Never publish an AI composite as a factual depiction of a demographic group without disclosure or provenance. Always pair composite imagery with primary demographic data from the Census or American Community Survey when the composite is being used to make a demographic claim. Treat the composite as a prompt-level thermometer for the model, not a measurement of the population it visually claims to represent. The rule is simple enough to fit on a one-page editorial policy.
Editors can adopt a simple house rule that applies to every desk and every freelancer at once. Any AI-generated depiction of a person or group must be labeled, sourced to the model, and accompanied by a sentence of context. The labeling cost is minor and the trust dividend is substantial across every audience Pew measures. Readers rapidly learn to read labeled synthetic imagery with appropriate skepticism, which is the goal rather than a side effect. Teams can draw on the AI images turned into three-dimensional worlds conversation to see where the labeling norms are heading.
For educators, the composite is a teaching asset once it is framed correctly. Pair every classroom use with the actual demographic data for the same group and name the mechanism that produced the composite. Then run the student through a short decoding exercise that identifies which aspects are statistically defensible and which are not. The exercise builds media literacy that transfers well beyond AI into every visual medium students will encounter. That transfer is the long-run payoff of covering the AI composite portrait topic rigorously in the classroom.
Frequently Asked Questions on AI Composite Portraits
AI Creates Average Person Across US refers to generative image prompts that ask a diffusion model like Midjourney or DALL-E to depict a typical resident. The output is a training-data mode, not a statistical average of the population, so the result can look confident while being demographically inaccurate. Readers should treat it as a model reflection, not a demographic measurement.
A true demographic average weights every real resident equally based on Census data. An AI composite samples from the frequency of images on the training web, which over-represents photogenic, young, lighter-skinned subjects. The two numbers can differ by 20 to 40 percentage points on race, gender, or age, as published bias audits consistently show.
Measured bias varies by prompt and by release, but 2026 audits generally rank Midjourney v6 and Stable Diffusion XL as the most skewed on neutral person prompts. DALL-E 3 and Imagen 3 ship classifier-based debiasing layers that cut some measured skew, though those layers introduce their own trade-offs documented in model cards and third-party audits. Treat this detail as part of the broader editorial frame for reading AI composite portraits with skepticism.
The training data pairs Southern state names with specific visual tropes like cowboy hats, trucks, and rural settings far more often than the real occupational distribution. The model learns the place-to-visual mapping and reproduces it on every prompt. The output tells you what the web pictures when it thinks of the state, not what the state’s residents actually look like.
You can, but you must disclose that the imagery is synthetic, provide provenance metadata like C2PA credentials, and run a bias audit against the actual demographic you are depicting. FTC guidance and several state laws treat undisclosed composites as potentially deceptive. Responsible publishers pair composites with primary Census data so readers can evaluate the claim themselves.
Yes, when the composite is treated as a measurement of the model rather than the population at large. Researchers use standardized prompts across releases to track how model behavior drifts over time. The composite then functions as a thermometer for the model and its training data at that release. Treating composite outputs as a demographic claim is a misuse that standardized bias audits actively discourage.
Older adults mostly disappear from the composite outputs across every tested model generation. The median apparent age in default composites clusters between 25 and 32 years old across major vendors. The real US median age sits closer to 39 and roughly 17 percent of Americans are 65 or older. Those older residents represent less than 2 percent of composite outputs across major models in current audits.
Treat composites as internal persona work rather than consumer-facing creative. If a composite does appear in a public campaign, disclose that it is AI-generated, include provenance metadata, and document a bias audit against the intended audience. Several platforms including Meta now require that attestation and shift liability back to advertisers who misrepresent synthetic imagery.
The Stanford HAI policy brief documents that occupation prompts return male subjects roughly 85 percent of the time. That skew runs far beyond real US labor force composition on the same roles. The brief recommends mandatory bias testing, disclosure of training data demographics, and classifier guidance toward parity for consumer image products. Policy teams in several jurisdictions are now adapting the recommendations into formal rulemaking.
Yes. Google paused Gemini image generation of people in February 2024 after its diversity layer produced historically inaccurate outputs. The pause lasted roughly six months and relaunched with a narrower feature set and a documented product principle about when demographic variation is appropriate. The incident is now cited across the field as a warning about aggressive mitigation.
Academic prototypes already cut measured bias by half or more, but commercial deployment requires solving the user experience problem of specifying demographic conditioning without defeating the composite prompt. Most practitioners expect balanced generation to ship broadly across major platforms by 2027 or 2028. Full adoption will depend on both engineering progress and regulatory pressure.
The US Census Bureau and the American Community Survey publish authoritative demographic data at national, state, and local levels. The Public Policy Institute of California, the Pew Research Center, and the Kaiser Family Foundation publish detailed secondary analyses. These sources are updated regularly, documented transparently, and treated as the baseline against which AI composite bias is measured.
Prompt engineering can shift outputs noticeably but cannot override the underlying training distribution. Documentary photography style tokens widen demographic spread, negative prompts excluding stock aesthetics help, and explicit demographic conditioning can produce balanced outputs. The limitation is that neutral prompts without these interventions, which is how most users ask, still return the baseline biased distribution.
C2PA content credentials attach cryptographic provenance metadata to an image, recording that the image was generated by a specific model and when. Major platforms and camera manufacturers are adopting the standard, so by 2027 most published images will carry provenance. A labeled composite becomes much harder to pass off as a documentary portrait, which raises the floor on composite-based deception.