Ask an AI image tool to voice a stereotype about sexuality, and it may well refuse. New research finds it will then turn around and draw that same stereotype anyway. Also, think robots need something like an arc reactor? Well, researchers just made one powered by an Android phone. Also, yet another study proves music can actually help soothe pain and help with recovery.

Send us feedback at [email protected].

A new study accepted for presentation at the 2026 Conference on Empirical Methods in Natural Language Processing has caught OpenAI and Google's image-generation models doing something that should unsettle anyone who uses these tools: refusing to voice a stereotype while quietly acting on it anyway. Researchers tested GPT Image 1 Mini and Gemini 2.5 Flash Image using more than a thousand AI-generated faces and found that both models declined to label a person's sexual orientation from their appearance the vast majority of the time, GPT refused 91% of such requests, Gemini 92%. The moment the framing shifted from asking a question to requesting an image edit, however, compliance soared. GPT followed through on requests to make someone appear gay or straight more than 70% of the time, while Gemini exceeded 99%.

The downstream effects were just as troubling. When researchers asked the models to describe the personalities, interests, and occupations of people in the edited images, familiar stereotypes surfaced: images modified to appear gay were more frequently paired with fashion and theater, while those modified to appear straight were linked to sports. A separate AI classifier was able to identify the intended sexual-orientation category in the altered images with 83% to 88% accuracy, suggesting the models were consistently embedding recognizable visual patterns. Both models also complied more than 97% of the time with requests to make subjects appear to have, or not have, a criminal record, a characteristic that faces cannot reveal at all.

You’ve put in the work to build an audience. But when the only way to reach your fans is through a social feed determined by algorithms, the platform decides who sees what you make next.

beehiiv gives you the tools to reach your audience directly. On their all-in-one platform, you can turn followers into subscribers, grow your newsletter, and connect it all to your community or podcast. Stop chasing virality on platforms you can’t control. Build a stronger, more sustainable business on beehiiv.

Get started for free.

A tiny humanoid robot uses an Android phone as its brain and costs just $400 UCLA researchers have built a working humanoid robot that runs on an ordinary Android smartphone, and it costs roughly $400 to construct, not counting the phone itself. Called PhoneBot, the 19-inch, four-pound prototype mounts an Android device on its torso and leans on that phone to do the heavy lifting. The phone's camera identifies and tracks people, its built-in sensors monitor orientation and movement, and it processes motion instructions that are then sent over Wi-Fi to a small onboard computer controlling the robot's 13 motors. Researchers have already demonstrated PhoneBot walking, following a person around, and recovering after falling face-first onto the ground.

What makes the project especially notable is how little it demands from the hardware. The team tested PhoneBot across four different Android devices, including a 2017 Honor 9, and all of them handled the core tasks. An older phone missed out on a mapping feature due to missing Google services, but the fundamentals worked across the board. The robot currently lacks arms and is limited to flat surfaces, and its affordable motors constrain how much weight it can carry, but the researchers are clear that accessibility and affordability are the entire point. PhoneBot is an open-source platform aimed at students and researchers who want to experiment with humanoid robotics without enterprise-level budgets.

A new study out of University Hospital Zürich found that patients who listened to their own favorite music reported less pain following an IV needle insertion, while music specifically selected for its relaxing qualities failed to show a significant benefit. The research tracked 204 patients receiving an IV cannula before an MRI scan, dividing them into four groups: one heard Mozart's Sonata K448, one listened to relaxation music, one chose their own preferred artist or song, and a control group heard nothing at all. Those in the personal-choice group reported lower pain levels both 30 seconds and two minutes after the needle went in. The researchers found that favorite music helped patients both with and without preexisting pain, while Mozart's benefit was more limited, appearing mainly in patients who were already hurting and only at the two-minute mark.

As for why the relaxation music underperformed, the team suggested its slow tempo and minimal variation may have made it easier to mentally tune out, though the study did not definitively establish what qualities gave personally meaningful songs their edge. It is worth noting that a technician accompanied and spoke to all patients throughout, providing a shared source of comfort across every group, which means music was not the only variable at play. The Zürich team said they plan to incorporate music more routinely during venous punctures for radiological procedures, particularly for patients who are anxious or claustrophobic. The findings are limited to a brief needle procedure and do not speak to whether the same effect would hold during longer or more painful treatments, leaving that question open for future research.

Razer Basilisk V4 spins 180 degrees at a button press

OnePlus Ice Card Mini is a tiny magnetic phone cooler

AMD's FSR 4 is coming to handheld gaming PCs

P7 DarkSword steals iPhone data in real time

iPhone 18 Pro Max SOS issues get a carrier update

Machinarium 2 is officially happening, 17 years later

Pokémon TCG's 30th Anniversary release is already a mess

Apple may host an October 27 event with new Macs