Yes—but the headline needs important context. The example came from a 2021 research study of OpenAI’s image GPT, or iGPT. It was not ChatGPT, not a modern consumer text-to-image chatbot, and not evidence that a machine “wanted” to sexualize Representative Alexandria Ocasio-Cortez.
The disturbing result was that, when the system completed cropped photographs of people, it repeatedly treated sexualized clothing as a plausible visual continuation for women while associating men more often with suits, workplaces, and career-specific clothing. That made the incident more than an isolated bad image: it exposed how a model can reproduce social patterns embedded in its training data.
The headline needs two important corrections
- This was an image-completion experiment, not a conversational AI prompt. iGPT was a research model trained to predict sequences of image pixels. It learned visual regularities from data and used those regularities to fill in missing portions of an image.
- The result was evidence of learned statistical bias, not human-like intent. The model did not understand sexism, form an opinion about Ocasio-Cortez, or decide that women should wear revealing clothing. Its training distribution and prediction objective made certain visual continuations more likely.
That distinction does not make the result harmless. It explains why the problem is best understood as a data, model-evaluation, and governance failure rather than as a machine having an offensive personal thought.
What happened to the cropped photograph
Ryan Steed and Aylin Caliskan used cropped photographs of people to examine whether unsupervised computer-vision systems had learned human-like social associations. One illustrative example involved a cropped photograph of Alexandria Ocasio-Cortez in professional clothing.
#1 Best Overall
- Wacom Intuos Small Graphics Drawing Tablet: Enjoy industry leading tablet performance in superior control and precision with Wacom's EMR, battery free technology that feels like pen on paper
- Works With All Software: Wacom Intuos tablet can be used in any software program to explore new facets of digital creativity; draw, paint, edit photos/videos, create designs, and mark up documents
- What the Professionals Use: Wacom's industry leading pen technology and pen to paper feeling makes it the preferred drawing tablet of professional graphic designers
- Software and Training Included: Only Wacom gives you software with every purchase. Register your Intuos tablet and gain access to some of the best creative software and Wacom's online training
- Wacom is the Global Leader in Drawing Tablet and Displays: For over 40 years in pen display and tablet market, you can trust that Wacom to help you bring your vision, ideas and creativity to life
The iGPT system generated multiple completions. According to reporting on the study, most of the reported completions placed Ocasio-Cortez in a bikini or low-cut clothing, despite the professional context of the original photograph.
The example became especially controversial because it involved a recognizable living person. A sexualized synthetic depiction can be harmful even when it is obviously artificial: reproducing it may objectify the subject, increase circulation of the image, and make the depicted person carry the burden of a demonstration intended to criticize the technology.
The image was initially included in pixelated form in the research materials. After ethical objections, the researchers removed it. There is no evidence that Ocasio-Cortez participated in the experiment, consented to the sexualized depictions, or had any personal relationship with the researchers that explains why she appeared in the example.
No such image is reproduced here. A neutral description or diagram communicates the research finding without recirculating the material that caused the harm.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallThe larger pattern was the point—not just the AOC example
The researchers did not set out merely to find one shocking completion. They tested associations involving gender, race, intersectional identity, weight, disability, and ethnicity. Across 15 replicated human Implicit Association Tests, they reported statistically significant bias on eight.
The study examined two unsupervised visual systems: OpenAI’s iGPT and Google’s SimCLR. They were not identical tools. iGPT could generate image continuations, while SimCLR was primarily used to learn visual representations that could then be evaluated for associations. The bikini-completion example is therefore best understood as an illustration of iGPT’s behavior, not proof that both systems produced the same image output.
The reported gender results were stark. Women were completed in swimwear or low-cut tops 53% of the time. For men, one account reported shirtless or otherwise revealing clothing 7.5% of the time and suits or career-specific clothing 42.5% of the time. Another account summarized the male result as 43% suit completions. The small difference reflects different category descriptions or groupings in secondary coverage; the consistent finding was an asymmetry between sexualized portrayals of women and professional portrayals of men.
Rank #2
- Word-first 16K Pressure Levels: The upgraded stylus features 16,384 levels of pressure sensitivity and supports up to 60 degrees of tilt, delivering smoother lines and shading for a natural drawing experience. With no battery or charging needed, it operates like a real pen, making it easy for beginners to create effortlessly. This functionality helps novice artists develop their skills and explore their creativity without the intimidation of complex tools
- Designed for Beginners: This drawing pad desinged with 8 customizable shortcuts for both right and left-hand users, express keys create a highly ergonomic and convenient work platform
- Perfectly Adapted for Android: The XPPen Deco 01 V3 art tablet supports connections with Android devices running version 10.0 and above. It is recommended to download the XPPen Tools Android application, which adapts to your smartphone's screen aspect ratio, ensuring accurate mapping. It also supports mapping on Android screens with different aspect ratios in portrait mode
- Large Drawing Space, Bigger Bold Inspiration: This expansive drawing pad has10 x 6.25-inch helps you break through the limit between shortcut keys and drawing area
- Easy Connectivity for Beginners: The Deco 01 V3 offers USB-C to USB-C connectivity, plus adapters for USB C. This ensures easy connection to various devices, allowing beginner artists to set up quickly and focus on their creativity without compatibility concerns. Whether using a laptop, tablet, or desktop, the Deco 01 V3 provides a seamless experience, making it an ideal choice for those just starting their digital art journey
The 53% figure belongs to that particular experiment, its images, crops, model, completion procedure, and category definitions. It is not a current benchmark for every image model, and it should not be presented as a prediction about what all AI systems would do to Ocasio-Cortez or to women generally.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Why a model can produce a sexist-looking result without “believing” anything
Image models learn from patterns in their training material. If women are disproportionately depicted in sexualized contexts, while men are more often shown in offices, uniforms, suits, or other occupational settings, those correlations can become useful for prediction.
For an autoregressive image model such as iGPT, the basic task was to predict what pixels are likely to come next. The model did not need a concept of sexism to make a biased completion. It only needed to learn that, within its training distribution, particular visual cues associated with women often appeared alongside particular types of clothing.
OpenAI’s own iGPT documentation acknowledged this tension: some visual regularities are useful for ordinary prediction, while others can be harmful from a fairness perspective. A model can therefore be technically successful at predicting statistically plausible pixels while producing an output that is socially degrading or unfair.
This is the central lesson of the incident. The failure was not simply that an algorithm made an unusual guess. The failure was that the training data and objective made a stereotyped guess seem plausible enough to repeat.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Why the example should not be generalized to every modern image generator
iGPT was a 2020-era research model with a pixel-sequence prediction method. It is not the same product or architecture as DALL·E, ChatGPT’s current image-generation features, Stable Diffusion, Gemini, or contemporary image-editing systems.
Those systems may have different training data, safety filters, prompting interfaces, image encoders, sampling methods, and post-deployment monitoring. The Steed-Caliskan study did not test all of them. Its exact percentages cannot be transferred to present-day products without a new, controlled evaluation.
Rank #3
- 【Android 16 Tablet with Gemini AI 3.5】The 11 inch TABWEE T90 brings a cleaner, smarter tablet experience for work, study and home use. Android 16 adds a simpler lock screen, customizable Quick Settings, improved video calls with background blur and portrait light, easier cloud photo selection and enhanced Linux terminal support. With Gemini AI 3.5, you can search, translate, summarize notes, organize ideas and plan tasks more efficiently—ideal for online classes, reading, video calls and light office work
- 【11 Inch FHD IPS Display for Streaming, Reading and Learning】Enjoy a bigger and clearer viewing experience on the 11 inch FHD display with 1920 x 1200 resolution and an 84.9% screen-to-body ratio. The fully laminated TDDI screen delivers more natural colors, smoother touch response, and a more immersive feel when watching movies, reading documents, viewing recipes, or studying online. With Widevine L1 support, dual smart speakers, and 380-nit brightness, this Android tablet is ideal for Netflix-style streaming, YouTube, music, eBooks, online lessons, and comfortable indoor entertainment
- 【24GB RAM + 128GB ROM, More Space for Daily Needs】Android 16 tablet comes with up to 24GB RAM, including 8GB built-in RAM and 16GB virtual memory expansion, helping daily apps run smoothly for browsing, video watching, note-taking, email, and study apps. The 128GB ROM provides plenty of room for photos, documents, music, videos, learning materials, and work files. Need more storage? The TF card slot supports up to 2TB expansion, so you can keep more movies, books, files, and family memories without constantly deleting content. Memory expansion path: Settings > System > Memory Expansion
- 【T615 Octa-Core + 8000mAh Battery for Long Days】Powered by the Unisoc T615 octa-core processor, T90 delivers reliable everyday speed for streaming, video calls, online lessons, reading, emails and document editing. The 8000mAh battery supports up to 11H video playback, 8H daily entertainment or 40H music, while smart power management helps reduce background drain, making it ready for flights, commutes, study sessions and busy workdays
- 【13MP Camera with Google Lens】The 13MP rear camera works with Google Lens to make daily tasks easier. Scan documents, translate text, identify objects, search products, capture handwritten notes, or save useful information in seconds. The 5MP front camera is suitable for video calls and clear face-to-face communication. With dual-band 5G WiFi, Bluetooth 5.0, and support for GPS, GLONASS, Galileo, and Beidou, this Android tablet offers fast downloads, stable wireless connections, easy accessory pairing, and reliable location support for travel, maps, and mobile use
That limitation does not make the historical study irrelevant. It shows why every new system needs to be tested rather than assumed to be neutral. A modern model could reduce one form of bias while introducing another, or behave differently depending on the prompt, crop, identity, clothing, lighting, language, or safety setting.
What a serious evaluation should test
A responsible evaluation cannot rely on one memorable example. It should measure whether comparable inputs receive systematically different treatment.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
- Use matched images and prompts. Hold the setting, pose, crop, and professional context constant while varying only the identity-related feature being studied.
- Repeat the test. Generative systems can produce different results across random seeds and runs. One output is not a rate.
- Define categories before counting. “Revealing,” “professional,” and “career-specific” clothing need clear, reproducible definitions.
- Test intersectional identities. Gender bias may interact with race, ethnicity, age, body size, disability, or other attributes rather than appearing as one simple male-versus-female effect.
- Measure both harmful and beneficial associations. A system may avoid sexualization in one context but erase disabled people, stereotype occupations, or make some groups less visible.
- Include human review and affected communities. Automated scores can miss humiliation, context, cultural meaning, and the consequences of distributing an output involving a real person.
- Monitor after deployment. Bias is not a one-time debugging problem. New users, prompts, model updates, and downstream tools can expose failures that were absent in laboratory testing.
UNESCO’s 2025 AI red-teaming playbook reflects this broader approach by encouraging tests for gender stereotypes and technology-facilitated gender-based violence before harmful outputs appear organically. NIST’s Generative AI Profile similarly treats harmful bias, privacy, synthetic-content misuse, evaluation, and post-deployment monitoring as lifecycle risks that organizations must map, measure, manage, and govern.
The ethical problem of showing the evidence
The AOC example also illustrates a difficult rule for AI-safety research: documenting harmful output can itself cause harm.
Researchers and journalists may need to establish that an output occurred. But displaying the output can re-sexualize a real person, make the image easier to find, and turn a subject into an involuntary exhibit. Pixelation is not automatically an ethical solution if the image remains recognizable or readily recoverable.
A better practice is to describe the input-output mismatch, report the method and aggregate findings, and explain why the original image was withheld or removed. If a visual is necessary, a neutral crop, abstract diagram, or recreated non-identifying illustration is generally more informative than the harmful depiction itself.
Free tools Windows power users keep installed
One-click scans. No signup required.
The image should also never be described as an authentic photograph. It was a machine-generated completion of a cropped source image, not a real photograph of Ocasio-Cortez wearing the clothing shown in the output.
Rank #4
- Wacom Intuos Medium Bluetooth Graphics Drawing Tablet: Enjoy industry leading tablet performance in superior control and precision with Wacom's EMR battery free technology that feels like pen on paper
- Works With All Software: Wacom Intuos tablet can be used in any software program to explore new facets of digital creativity; draw, paint, edit photos/videos, create designs, and mark up documents
- Wireless Superior Connectivity: Connect wirelessly via Bluetooth or directly using USB-A cable which enables you to work, draw or create whether it's at a desk, on the sofa, in classrom or even outside
- Software and Training Included: Only Wacom gives you software with every purchase. Register your Intuos tablet and gain access to some of the best creative software and Wacom's online training
- Wacom is the Global Leader in Drawing Tablet and Displays: For over 40 years in pen display and tablet market, you can trust that Wacom to help you bring your vision, ideas and creativity to life
From research finding to broader social risk
Gendered image outputs matter because they can reinforce patterns that already exist offline: objectification, unequal professional visibility, discriminatory assumptions, and the treatment of women’s bodies as available for public manipulation.
At larger scale, synthetic media can also be used for impersonation, fraud, harassment, and image-based abuse. NIST’s work on synthetic content identifies discrimination and bias, impersonation, fraud, and erosion of trust as risks that synthetic media can intensify. It discusses tools such as provenance, watermarking, authentication, detection, and safeguards against harmful generation—but also emphasizes that technical detection alone cannot resolve the social harm.
That is why the relevant response is not just “make the classifier better.” It includes better data governance, representative evaluation sets, safety testing, reporting channels, consent practices, provenance systems, platform enforcement, and support for people targeted by synthetic abuse.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsWhat the TAKE IT DOWN Act changes—and what it does not
The legal landscape has changed since the 2021 experiment. In the United States, the TAKE IT DOWN Act became Public Law 119-12 on May 19, 2025. The law covers certain nonconsensual intimate visual depictions, including some digital forgeries created or altered with software or artificial intelligence, and requires covered platforms to establish a notice-and-removal process.
Under the statutory process, covered platforms must remove qualifying material—including known identical copies—within 48 hours after receiving a valid request through the required procedure. The Federal Trade Commission began enforcing the Act’s platform notice-and-removal provisions on May 19, 2026, and its guidance expressly includes AI-generated deepfakes within the discussion of image-based abuse.
That does not mean every sexualized AI image automatically falls within the same legal category. Coverage can depend on the statute’s definitions, the nature of the depiction, consent, intent, the platform involved, and other facts. The law is also not a retroactive legal judgment about the 2021 research event, nor is it a universal solution to biased or degrading model outputs.
For someone targeted by potentially qualifying material, the practical lesson is to preserve evidence, use the platform’s designated reporting and removal process, and seek appropriate legal or victim-support advice. The specific legal remedy depends on the facts and jurisdiction.
Recommended Free Tools
Best Value
- WACOM’S BEST PEN TECHNOLOGY: Pro Pen 3 offers amazing precision, 8192 pressure levels, tilt support & lag-free tracking for smooth, precise strokes; choose slim, straight or flared grip, adjust balance & button layout for precision and comfort
- DIAL INTO PRODUCTIVITY: Speed up and stay in the creative flow with 10 customizable tablet ExpressKeys & 2 mechanical dials conveniently located at the top of the tablet and close to your keyboard for efficiency
- DESIGNED FOR MODERN MONITORS: With a versatile 16:9 format perfect for multiple monitors and wide displays; medium size has a large active area with a small footprint 11.4 x 8.1” / 291 x 206 mm
- SLEEK AND STURDY: Measuring 4mm at its thinnest this professional-grade drawing tablet feels like pen-on-paper on your desk, combined with the robust durability of magnesium making it equally suited for your office or life on the go
- WIRELESS FREEDOM: use Intuos Pro with USB or connect wirelessly via Bluetooth to multiple computers with a simple switch
What the experiment actually proves
The study supports a narrower and more defensible conclusion than the headline may suggest:
- Unsupervised visual models can learn social associations from image data without being explicitly programmed with those stereotypes.
- Those associations can appear in generated or reconstructed visual content.
- The same system can be accurate at visual prediction and harmful from a fairness perspective.
- A population-level pattern is more important than any one shocking output, but an individual public figure can still bear the immediate cost of demonstrating that pattern.
- The exact result should not be treated as a present-day score for every AI image product.
Readers looking for a sustained treatment of why automated systems reproduce social patterns may also want More than a Glitch by Meredith Broussard, a useful book on AI bias and the limits of treating technical systems as neutral. It is a follow-up for understanding the larger problem, not a substitute for platform safeguards or support for people targeted by synthetic abuse.
Research basis and scope
This article is based on the 2021 study by Ryan Steed and Aylin Caliskan; OpenAI’s documentation about iGPT; contemporary reporting on the image-completion experiment; UNESCO’s 2025 AI red-teaming guidance; NIST’s Generative AI Profile and synthetic-content work; Public Law 119-12, the TAKE IT DOWN Act; and Federal Trade Commission guidance on its platform notice-and-removal provisions.
The policy discussion reflects the research context available through August 2026. Legal requirements vary by country and by the facts of a particular case.
Frequently Asked Questions
Was the AOC image created by ChatGPT?
No. The example came from OpenAI’s earlier iGPT research model, which predicted sequences of image pixels. It was not ChatGPT and was not a modern text-to-image prompt in the consumer-product sense.
Did Alexandria Ocasio-Cortez consent to the experiment?
The available evidence does not establish that she participated in the experiment or consented to the sexualized depictions. The source photograph showed her in professional clothing, and the researchers later removed the pixelated example after ethical objections.
Does the 53% figure describe current AI image generators?
No. The 53% result belongs to that study’s particular model, images, crops, completion procedure, and category definitions. It should not be presented as a current benchmark for DALL·E, ChatGPT image generation, Stable Diffusion, Gemini, or every other image system.
Is every sexualized AI image covered by the TAKE IT DOWN Act?
No. The Act covers certain nonconsensual intimate visual depictions and digital forgeries, subject to statutory definitions and other facts. Whether a specific image qualifies can depend on the depiction, consent, intent, platform, and jurisdiction.
The Bottom Line
The important failure was not that an AI imagined a bikini in a human sense. It was that a system trained on socially patterned visual data treated sexualized representation as a plausible default for a woman even when the source context pointed toward professional attire. That is why the AOC example belongs in the history of AI bias—and why responsible systems must be tested, monitored, and governed rather than assumed to be neutral.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
