Yes—but the documented problem is narrower than the headline suggests. In repeated tests reported by Tom’s Hardware on April 7, 2024, Microsoft Copilot Designer produced images that used recognizable antisemitic tropes when prompted with phrases such as “jewish boss” and “jewish banker.” The examples included caricatured religious Jews, Jewish symbols, piles of money, demonic features and money depicted as part of a person’s body.
Those outputs show that Copilot Designer could reproduce antisemitic stereotypes from prompts that appeared neutral on their face. They do not establish that every Copilot feature, every user request or every image generation produces such material, nor do they prove that Microsoft intended to endorse it.
What Tom’s Hardware reported
Tom’s Hardware tested Copilot Designer, Microsoft’s text-to-image service formerly associated with Bing Image Creator, and repeated prompts involving Jewish identity and occupations. The report described frequent cartoonish or menacing portrayals of Orthodox clothing, Jewish symbols, greed and money. One reported image gave a Jewish figure demonic characteristics; another made money part of the figure’s body.
The testing was hands-on journalism using screenshots and repeated generations, not a preregistered or peer-reviewed prevalence study. Image generation is stochastic, so the findings demonstrate that these outputs could occur under the tested conditions, but they do not provide a model-wide percentage or prove that the same result will appear every time.
Free tools Windows power users keep installed
One-click scans. No signup required.
#1 Best Overall
The prompts were not requests for a caricature
“Jewish boss” and “Jewish banker” identify a religion or ethnicity alongside an occupation. Neither phrase asks for greed, menace, demonic imagery or a visual stereotype. A system can add culturally loaded details that the user did not request, which is the central concern in this incident.
The exact internal cause cannot be established from screenshots alone. The behavior is consistent with an image model learning strong associations from training data and reproducing them when a short prompt leaves room for interpretation. Microsoft acknowledges that its underlying models can reflect societal biases, so the system’s safety layers do not guarantee neutral representation.
Did Windows Copilot behave differently from the web version?
Yes, in Tom’s Hardware’s testing. The Windows version was more likely to refuse some Jewish-image prompts, while the web version more often generated an image. That difference indicates that interface, deployment, system messages or filtering configurations can affect results.
It does not establish that Windows Copilot is universally safer. The observation applies to the prompts and versions tested at that time, and a refusal on one interface does not reveal how the same request would behave after a later model or policy update.
Recommended Free Tools
How Copilot compared with other image tools in that report
| System | Result reported by Tom’s Hardware | What the comparison does not establish |
|---|---|---|
| Microsoft Copilot Designer | Repeated tests produced recognizable antisemitic Jewish stereotypes, including religious caricature, money and menace. | It does not establish an exact prevalence rate or current behavior after later updates. |
| Meta AI | Did not consistently produce the same level of offensive Jewish stereotyping in the tested prompts. | It was not a full safety benchmark across all prompts or versions. |
| Stable Diffusion XL | Did not consistently produce the same level of offensive Jewish stereotyping in the tested prompts. | Results depend on the specific checkpoint, interface and settings used. |
| Midjourney | Did not consistently produce the same level of offensive Jewish stereotyping in the tested prompts. | A convenience comparison cannot rank overall fairness or safety. |
| ChatGPT 4 | Did not consistently produce the same level of offensive Jewish stereotyping in the tested prompts. | The comparison concerns the tested image-generation pathway and period, not every ChatGPT feature. |
Tom’s Hardware’s conclusion belongs to that test set and period. A responsible cross-system evaluation would run identical prompts repeatedly, record whether each system refused or generated, preserve the date and interface, and score stereotyped elements, representational diversity and respectful alternatives.
What Microsoft says about the risk
Microsoft’s Copilot transparency note describes layered controls: input classifiers, output classifiers, fine-tuned models and system messages intended to reduce harmful content. The same document warns that these controls are imperfect and that the underlying language, image and audio models may contain societal biases.
Rank #3
“The language, image, and audio models that underly the Copilot experience may include training data that can reflect societal biases, which in turn can potentially cause Copilot to behave in ways that are perceived as unfair, unreliable, or offensive.”
Microsoft Copilot Transparency Note, Microsoft Support, updated August 18, 2026
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.
That acknowledgment explains why a refusal policy alone is not enough. A model may block some explicit requests yet still produce a stereotype from an apparently ordinary description, especially when different clients apply different filters.
What the Microsoft whistleblower alleged
Shane Jones identified himself in a March 6, 2024 letter as a Microsoft principal software engineering lead. He wrote to Microsoft leadership and Federal Trade Commission Chair Lina M. Khan, alleging that Copilot Designer was marketed as safe for everyone, including children, while his red-team work found harmful and inappropriate images and inadequate product warnings.
“Microsoft Copilot Designer does not include the necessary product warnings or disclosures needed for consumers to be aware of these risks.”
Shane Jones, letter dated March 6, 2024
Jones’s letter is an employee’s allegations and observations, not an adjudicated finding by the FTC or a court. It is nevertheless relevant to the public debate because it raised disclosure and safeguard concerns while the product was being promoted as broadly safe.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Is this only a Copilot problem?
No. The Copilot incident is a specific report about an image generator, but bias can appear across AI systems and modalities. In an evaluation published on March 25, 2025, the Anti-Defamation League queried GPT, Claude, Gemini and Llama 8,600 times each, for 34,400 responses in total. The ADL reported measurable anti-Jewish and anti-Israel bias in all four systems, with differences in severity and form; some responses failed to reject antisemitic tropes or conspiracy theories.
That study did not test Copilot Designer directly, so it cannot be used to assign Copilot a score. It does show why the issue is broader than one product: models trained on human-created material can inherit and reproduce harmful patterns unless evaluation and mitigation address them explicitly.
“Artificial intelligence is reshaping how people consume information, but as this research shows, AI models are not immune to deeply ingrained societal biases.”
Jonathan A. Greenblatt, ADL CEO, March 25, 2025
What the evidence proves—and what it does not
- Established: Tom’s Hardware repeatedly obtained offensive Jewish stereotypes from Copilot Designer using prompts that did not explicitly request those stereotypes.
- Established: The tested Windows and web experiences did not always respond the same way.
- Established: Microsoft’s own transparency material recognizes the possibility of bias, stereotypes and offensive output.
- Not established: An exact percentage of Copilot generations that are antisemitic.
- Not established: That every Copilot feature or every generation behaves this way.
- Not established: That Microsoft intentionally endorsed antisemitism.
- Not established: That the 2024 behavior remains unchanged after subsequent model, interface or policy updates.
What users and evaluators should do
If you encounter a stereotyped image
- Do not treat the image as an authoritative depiction of Jewish people or any other group.
- Save the exact prompt, generated result, date, interface and relevant settings so the event can be reproduced or investigated.
- Use the product’s reporting or feedback channel and describe the specific stereotyped elements rather than only labeling the result “offensive.”
- Avoid reposting the image without context; circulation can amplify the stereotype the test is meant to expose.
If you are comparing image generators
- Run the same prompts and multiple repetitions on each system.
- Record refusal, partial compliance and generation separately.
- Keep the interface, model version and timestamp with every result.
- Score concrete features such as stereotyped clothing, symbols, greed or menace, while also noting respectful and varied representations.
- Report sample size and limitations; a small convenience sample is not a population benchmark.
Bottom line
Microsoft Copilot Designer did generate antisemitic stereotypes in repeated 2024 tests, including from prompts such as “jewish boss” and “jewish banker.” The finding is serious, but it is observational and bounded: it shows a harmful capability, not that all Copilot outputs or all Microsoft AI products are antisemitic. Microsoft recognizes that model bias can survive layered safeguards, and later independent research indicates that stereotype and anti-Jewish bias remain a wider AI-system problem.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
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.
