Blogroll

In April, Meta quietly reversed itself after removing an Instagram post honoring older lesbian relationships in Brazil. The excised post was not sexual in nature and did not contain material harmful to minors. The post in question documented a snapshot from a moment in history where lesbians were forced to hide their relationships as "roommates" or "gal pals" and their love was scrubbed from the public record. Nevertheless, Meta removed the content. 

Meta cited its hate speech rules. The Oversight Board later acknowledged what should have been obvious from the start: The Brazil case was an instance of over‑enforcement against a marginalized community, driven by automated systems that could not read context, reclaimed language, or even the full post itself. The content was restored only after outside intervention and advocacy from the LGBTQ+ community.

Mashable 101 Fan Fave: Vote for your favorite creator today!

This case is now being treated as a narrow content moderation error, but policymakers need to recognize that it signals a clear warning about what happens when lawmakers push platforms to police content instead of fixing design. Across the country, states are rushing to "protect kids online" by restricting access to social media or pressuring companies to remove vaguely defined "harmful" content. But what happened in Brazil shows the human cost of that approach. 

When platforms are incentivized to remove speech quickly and at scale, they do not become better judges of nuance. Social media becomes a blunt instrument, and the first people hit are those whose stories require human context and radical empathy to be understood.

If lawmakers actually want to protect kids, they should stop asking platforms to decide which stories are acceptable and start regulating core design choices that cause harm in the first place, like endless scroll, engagement‑based recommendations, and surveillance‑driven feeds.

SEE ALSO: I had a Grindr sugar daddy for a day. Then he tried to get a refund.

Here’s why that distinction matters, especially for LGBTQ+ kids and other marginalized communities, like neurodivergent kids. LGBTQ+ young people are far more likely than their peers to rely on online spaces to find community, information, and support, often because those things are unavailable or unsafe at home or school. But they are also significantly more likely to end up in unsafe online interactions: harassment, grooming, doxxing, or being pushed into high‑risk spaces they didn’t seek out. 

In Australia, after a social media ban on anyone under 16 was enacted, disability rights advocates noted that autistic youth were cut off from some of the only support and peer networks available to them. 

Recommendation systems don’t understand vulnerability, but they understand engagement. When a queer kid searches for community, platforms often respond by aggressively amplifying whatever keeps them clicking. Usually, this means increasingly sexualized content, adult strangers, extremist rhetoric, or predatory accounts that know exactly how to exploit isolation. 

Infinite scroll makes disengagement much harder for adolescents, according to the Electronic Privacy Information Center, even more so for those in vulnerable communities. Algorithmic "friend" or "account" suggestions collapse liminal boundaries between teens and adults. Weak defaults make it difficult to block, mute, or disappear.

Young people, not just LGBTQ+ young people, are exposed to harm online because platforms are built to extract attention, not protect users. Parents are right to be worried and to advocate for change. But a content-based framing misses the real problem. 

The greatest risks kids face online don’t come from a single bad post slipping through moderation, but from automated systems that push content at kids they didn’t ask for, connect them to people they don’t know, and keep them scrolling long after warning signs appear.

Policymakers at both the state and federal levels need to design regulations that address those risks directly. Age‑appropriate design codes don’t tell platforms what speech to allow, but they can tell platforms how to behave. Design codes require safer defaults, like limits on behavioral profiling, stronger blocking tools, reduced amplification of unsolicited recommendations, and guardrails that slow down virality and compulsive use. 

The public should advocate for product refinement, rather than infringement of First and Fourth Amendment rights. Design codes reduce the chance that a curious or lonely kid is algorithmically funneled into danger, like I was, searching for community and nudged toward risk by systems that did not care who I was.

Age‑appropriate design codes offer a way out of this mess. By regulating how platforms are built rather than what people are allowed to say, design code laws reduce harm without turning companies into cultural censors. They don’t require platforms to interpret reclaimed slurs, queer history, or political speech. Companies should instead be required to stop engineering addiction and risk. 

We don’t need more content or platform bans. We need fewer harmful systems. If we’re serious about protecting kids online, especially the ones already most at risk, this case reminds us exactly where to start.

This article reflects the opinion of the writer.

Lennon Torres is the Movement Director at the Heat Initiative and Founding Partner of The Attention Studio. 

TL;DR: Get a lifetime subscription to Skoove Premium Piano Lessons for $99.97 (reg. $299.99) through May 17 at 11:59 p.m. PT.

Opens in a new window Credit: Skoove Skoove Premium Piano Lessons: Lifetime Subscription $99.97
$299.99 Save $200.02   Get Deal

Learning to play the piano can seem daunting, especially with traditional lessons not in your budget, scheduling conflicts, and the limitations of free online tutorials. Yet, for aspiring musicians eager to take the first step, a new deal on Skoove Premium Piano Lessons could open the door to a more accessible, rewarding learning experience.

Right now, a lifetime subscription to Skoove Premium is on sale for $99.97 (reg. $299.99). That’s a 67% discount through May 17 at 11:59 p.m. PT.

Mashable Deals Be the first to know! Get editor selected deals texted right to your phone! Get editor selected deals texted right to your phone! Loading... Sign Me Up By signing up, you agree to receive recurring automated SMS marketing messages from Mashable Deals at the number provided. Msg and data rates may apply. Up to 2 messages/day. Reply STOP to opt out, HELP for help. Consent is not a condition of purchase. See our Privacy Policy and Terms of Use. Thanks for signing up!

Skoove is an online learning platform crafted for both beginners and intermediate players, earning recognition from notable outlets such as Apple, Forbes, Wired, and The Guardian. Unlike platforms that depend exclusively on pre-recorded lessons, Skoove integrates AI-driven feedback, actively listening to your playing and offering real-time responses. This feature helps students identify mistakes on the spot, fostering better habits and more effective practice sessions.

The platform includes more than 400 lessons and thousands of instructional videos, with new songs and exercises added regularly. You’ll find a mix of genres too, from pop songs by artists like Adele and The Beatles to classical music from composers such as Bach and Beethoven. There are also special courses and access to one-on-one support from Skoove instructors if you need extra guidance.

Flexibility is another hallmark of Skoove’s approach. The platform is compatible with acoustic pianos and USB or MIDI keyboards, and it operates seamlessly across iOS, Android, and desktop devices. Whether you have just 15 minutes to practice after work or prefer to dedicate weekends to developing your skills, Skoove adapts to your lifestyle.

If you’ve been thinking about learning piano but didn’t want to commit to recurring lesson fees, this lifetime subscription could save you money in the long run. You can get Skoove Premium Piano Lessons for $99.97 while the deal lasts through May 17 at 11:59 p.m. PT.

StackSocial prices subject to change.

TL;DR: Aivolut AI Book Creator is on sale for $48.99 (reg. $456) and helps entrepreneurs, coaches, and aspiring authors create KDP-ready books faster.

Opens in a new window Credit: Aivolut Books Aivolut AI Book Creator: Lifetime Subscription $48.99
$456 Save $407.01   Get Deal

Writing a book sounds exciting until you’re staring at a blank page and watching the cursor blink until your eyes glaze over. If you’ve had an idea sitting in your head — or scrawled in a notebook — Aivolut AI Book Creator is designed to help move the book-writing process along much faster.

The lifetime subscription to Aivolut AI Book Creator is currently on sale for $48.99, and it’s a one-time purchase so you don’t have to pay recurring subscription fees.

Mashable Deals Be the first to know! Get editor selected deals texted right to your phone! Get editor selected deals texted right to your phone! Loading... Sign Me Up By signing up, you agree to receive recurring automated SMS marketing messages from Mashable Deals at the number provided. Msg and data rates may apply. Up to 2 messages/day. Reply STOP to opt out, HELP for help. Consent is not a condition of purchase. See our Privacy Policy and Terms of Use. Thanks for signing up!

The platform is built specifically for people who want to create and publish books without having to utilize multiple writing and formatting tools. Aivolut combines AI-assisted writing with publishing support to help turn rough concepts into structured, coherent manuscripts. Users can start with a topic, outline, or even use the voice dictation feature, then customize the tone, structure, and style while the software helps generate long-form content.

One exceptional feature is its built-in Kindle Direct Publishing support. In addition to producing manuscripts, Aivolut can automatically create Amazon-ready metadata such as descriptions, keywords, and categories, which may save hours of manual setup for self-publishing authors. According to the platform, books can be prepared for Amazon KDP upload in under five minutes.

The software is ideal for entrepreneurs, marketers, coaches, freelancers, and other professionals looking to write books. It may also appeal to first-time nonfiction authors who want help organizing ideas without outsourcing ghostwriters or editors early in the process.

Another selling point is the program’s flexibility. Instead of locking users into fixed templates, Aivolut lets writers edit tone, blend in their own material, and adjust the structure as they go. You won’ be replacing your voice, the goal is to speed up the outlining, drafting, and formatting process.

If you’ve been putting off writing a book because the process felt too time-consuming, Aivolut AI Book Creator is currently available for $48.99 (reg. $456) for lifetime access.

StackSocial prices subject to change.

Sometimes, when you're scrolling through Paramount+'s movies looking for something to watch, the algorithm gods just drop all the right things on you. I was trying to line up a handful of movies to watch throughout the work week, and this mix, streaming for U.S. subscribers, just might do the trick.

The latest update for iPhone users, iOS 26.5, is officially here.

The beta version of the update has been available for developers for about a month now, giving iPhone users a pretty good idea of what new features to expect. However, as we've seen with previous iOS updates, features frequently get added or removed by the time the public version launches — so now we know exactly what's news in iOS 26.5.

First things first, iPhone users need to update their device to iOS 26.5. Besides the new features, updating to the latest OS ensures your device is protected with the latest security updates. To upgrade to iOS 26.5, simply go to Settings, then tap General, followed by Software Update. 

Users must have an iPhone 11 or newer in order to download and install iOS 26.5 to their device.

Mashable 101 Fan Fave: Vote for your favorite creator today!

RCS end-to-end encryption After updating to iOS 26.5, users will see a lock icon appear in RCS chats. Credit: Apple

It's official. After being tested in previous beta versions of iOS but failing to make the final release, encrypted RCS messaging is here.

RCS, or Rich Communication Services, is basically the messaging protocol that provides an Apple Messages-like experience when iPhone users and Android users text one another. It enables iPhone and Android devices the ability to send and receive high-quality media as well as view read receipts and typing indicators. 

However, before this update, these messages between iPhone and Android devices were unencrypted, meaning that they could potentially be intercepted and read by unauthorized third parties.

iOS 26.5 finally brings encryption to RCS messaging. It should be noted, though, that even though the feature has made it to the public release of iOS 26.5, encrypted RCS messages are still labeled as a beta feature by Apple. The availability of the feature is also dependent on your mobile carrier as well.

Yes, the green chat bubbles remain.

Apple Maps ads and Suggested Places

Nearly 13 and a half years after Apple launched its very own Maps application, advertisements have made their way to Apple Maps. (As of this writing, the feature is still labeled as "Coming Soon" on the Apple Ads website.)

Businesses will now be able to pay to advertise in the Apple Maps app, much like how app developers can run ads in Apple's App Store. Users will see local ads in the Maps app based on their location.

Ads will also appear in another brand-new Apple Maps feature called Suggested Places.

Apple has been recruiting advertisers for Apple Maps ads. Credit: Apple

While Suggested Places in Apple Maps will provide users with recommendations based on paid ads, it will also produce organic recommendations as well. Suggested Places will show users nearby trending locations, such as restaurants, shops, or tourist destinations, based on their recent search results.

USB-C accessory pairing

Want to use a Magic Mouse or Magic Keyboard with your iPhone? It's easier than ever before with iOS 26.5.

Simply connect the peripheral device to an iPhone via USB-C, and the device will automatically pair with the iPhone.

This means that after the initial pairing, users can simply unplug the USB-C cable and use the Magic Mouse or Magic Keyboard with their iPhone over Bluetooth without an additional required setup.

iPhone to Android transfer options

This is a small new update, but one that will surely be welcomed by an iPhone user making the switch to Android.

With iOS 26.5, users can now choose which message attachments they want to transfer to their Android device when moving data from their iPhone.

'Pride Luminance' wallpaper The new Pride Luminance background for iPads and iPhones. Credit: Apple

iOS 26.5 rolls out with a brand new iPhone wallpaper celebrating Pride Month.

Apple has released Pride Month wallpaper before. However, this marks the first time Apple has released one as a motion wallpaper. This isn't a static background. The "Pride Luminance" wallpaper is animated and customizable, meaning users can create their own version of the wallpaper using the 12 available colors.

The new Pride Luminance wallpaper that matches the new Pride Luminance watch face for Apple Watch in watchOS 26.5.

Want to learn more about getting the best out of your tech? Sign up for Mashable's Top Stories and Deals newsletters today.

When people start shopping for a reliable compact car, they usually end up looking at the same names over and over again. The Honda Civic, Toyota Corolla, and Mazda3 tend to dominate the conversation, while Subaru quietly gets ignored in the background.

Looking for a good movie to watch this week? So far, nothing seems to be unseating the Charlize Theron thriller Apex from the top spot in the Netflix Top 10 movie rankings in the U.S., but one cute and cuddly animated surprise might be the one to do it, and it's on this week's list of movies subscribers should add to their watch lists.

After more than a year of promises, end-to-end encrypted RCS (Rich Communication Services) messages are finally available on both Android and iPhone users as part of a beta. The move locks down chats no matter which platform you're on, although you'll need to meet certain conditions.

Get ready for more Jack Reacher at Amazon. On Monday, Prime Video announced that Reacher, one of the service's most popular shows, has been renewed for season 5. The news was announced ahead of Amazon’s upfront presentation.

Luxury EVs are getting absurdly expensive. It’s becoming increasingly common to see compact electric crossovers push past the $80,000 mark once options are added, all while offering compromises in range, practicality, or driving feel. Badge appeal still carries a lot of weight in this segment, but buyers are beginning to ask a pretty reasonable question: how much are you actually getting for the money?

You'll find Apple Watches and Whoop fitness trackers on a number of Mashable writers and editors' wrists. Fitbit may have popularized fitness trackers and the entire concept of "counting steps," but the brand has been eclipsed in recent years by feature-rich rivals.

When one of our reporters ran the New York City marathon wearing seven fitness trackers, Fitbit didn't impress.

But a new Fitbit could change that. The screen-free Google Fitbit Air hits store shelves May 19, and the $99.99 price tag is very tempting compared to an Apple Watch, which can cost $249 to $799. The Fitbit Air even has some unique features that could tempt dedicated Whoop fans.

Opens in a new window Credit: Google Google Fitbit Air $99.99 at Amazon
  Pre-order Here

We haven't had the chance to test the Google Fitbit Air yet (Steph Curry is not on our staff, unfortunately), but here are the top Fitbit Air features we're eyeing ahead of May 19.

Google Health Coach The new Google Health Coach launches alongside the Fitbit Air Credit: Google

When the Fitbit Air launches on May 19, Google will also debut the new Google Health Coach, which will be part of the Google Health Premium subscription ($9.99 a month or $99 a year).

Google Health Coach is an AI personal trainer that can make personalized workout and health recommendations tailored to you. Using the health, sleep, and fitness data collected by the Fitbit Air, you'll be able to get insights that go way beyond counting steps. The level of personalization is impressive, as the AI coach can take your sleep cycle, local weather, and medical history into account when designing workouts.

We don't love it when products require a paid subscription to unlock crucial features, but the amount of data available to your AI coach could set this tool apart from similar offerings. If you have a Pixel 4 smartwatch, you'll also be able to use this new feature.

A super lightweight design The Fitbit Air is lighter and thinner than the Whoop. Credit: Google

For the Fitbit Air, Google seems to be prioritizing lighter weight over battery life. While the Whoop 5.0 and MG have a longer battery life, the Fitbit Air will be significantly lighter and smaller.

The new Fitbit weighs a mere 12 grams, less than half the weight of the Whoop 5.0, which weighs in at 26.5 grams (the Whoop MG is 27.3 grams). The sensor area, which Google describes as a "pebble," is also slimmer in height and narrower in width than Whoop.

When compared to smartwatches, it has a significant size and weight advantage. That's crucial for people who want to use their fitness tracker as a sleep tracker.

Despite the small size, the Air can still measure heart rate, heart rhythm, SpO2 (blood oxygen levels), sleep stages, and more. It also contains an accelerometer and a gyroscope.

Gemini voice and text chat You can talk to the Google Health Coach by voice or text. Credit: Google

Within the Google Health app, you'll be able to talk to your AI coach by voice or text, with the conversations powered by Google's AI chatbot Gemini. You'll need your phone to do this, but for athletes who like to talk to their AI chatbot, you'll be able to talk in real-time. You'll be able to ask questions and make requests as you workout.

Perks for Gemini power users Credit: Google

Google offers several paid AI subscriptions for users who want to access the latest Gemini models and features — Google AI Pro and Google AI Ultra. With the debut of Google Health Coach, users on either of these plans will now get free access to Google Health Premium.

So, if you're already a Gemini power user, you can unlock extra features on the Fitbit Air at no extra cost.

Fitbit Air can use Gemini to "see" your workout and meal plans

As the maker of Gemini, one of the most advanced AI chatbots, Google has some big advantages over Apple, Whoop, Garmin, Nothing, and other smartwatch and fitness tracker brands.

Case in point: You can take pictures and upload them to Google Health Coach for analysis. A Google blog post states that users can "upload photos of gym whiteboards or snap photos of meals for nutritional analysis." You can also upload PDFs with health information.

Again, you'll need to use your phone to use this feature.

3-month free trial Credit: Google

Finally, Google has one more advantage over Whoop: a longer free trial. When you buy a Fitbit Air, you get three months of access to Google Health Premium, whereas Whoop only offers a 1-month free trial

For all these reasons, we expect the new Google Fitbit Air to be a serious competitor in the fitness tracker world and may even tempt some people to ditch their smartwatch. We're seeing growing interest in screen-free technology lately, which is one reason Whoop is so popular.

However, once again, Google is going all-in on artificial intelligence features. For athletes who are not interested in having an AI personal trainer or using Gemini as a workout buddy, the Fitbit Air may be a tougher sell.

Google touts its privacy features, but not everyone wants to provide such detailed health information to Big Tech.

The Google Fitbit Air is available for pre-order now for $99.99 at Amazon and the Google Store.

Texas Attorney General Ken Paxton has sued Netflix over claims it not only spies on you by collecting data without your permission but has also designed its service to keep you hooked — and your kids.

Many Android phones now promise five, six, or even seven years of major software updates, which is great. If you're spending $1,000 on a phone, it's fair to expect it to last more than half a decade—especially now that performance and camera quality have largely plateaued.

The hacking collective ShinyHunters says it disrupted a major education platform not once but twice over the past few weeks. And the data breach could not have come at a worse time for students and teachers. These events unfurled during school finals at many of the affected institutions.

On April 30, Instructure, the edtech company behind Canvas, the popular Learning Management System (LMS) utilized by educational institutions around the world, temporarily went offline. A day later, Instructure confirmed that a "criminal threat actor" was behind a data breach into the company's systems.

According to ShinyHunters, the group stole data from 275 million Canvas users at nearly 9,000 schools worldwide. The affected users include students, teachers, and staff, and while no passwords or other sensitive data were taken, the data stolen was significant. The hackers claimed usernames, email addresses, student IDs, and private messages exchanged on the platform were part of the stolen data. Some of the impacted users are underage students.

Mashable 101 Fan Fave: Vote for your favorite creator today!

Shortly after the hack, Instructure confirmed that it had revoked access from the bad actors, took measures to fix the issues and prevent another breach from occurring, and brought Canvas back online.

However, just one week later, ShinyHunters says it hit Canvas again. This time, the hackers compromised school-specific login pages for the platform and defaced the pages with messages threatening to publicly release the stolen data from the previous breach unless Instructure agreed to "negotiate a settlement."

A monetary demand from ShinyHunters was not surprising. The ransomware group is known for extorting victims following a data breach. A second breach at Instructure, however, was a surprise. Canvas once again went offline, and when it came back, the company had removed the source of the second incident: Free-For-Teacher accounts.

According to a newly updated incident page on Instructure's website, the company says it "identified a vulnerability regarding support tickets in our Free for Teacher environment that was exploited." 

"We temporarily disabled Free for Teacher while we complete a full security review," the company said. "We know that's disruptive, and we didn't make that call lightly. But keeping the entire Canvas platform secure has to come first."

While the second breach did not result in any stolen data, the timing of the security incident could not have been worse for students, as many schools are currently holding finals and other scheduled deadlines for end-of-year coursework.

As PCMag reports, "students and professors struggled to access the online platform used to submit assignments and tests." (Disclosure: PCMag and Mashable are both owned by the same parent company, Ziff Davis.)

According to data provided to Mashable from Alliance Risk Trends, Google searches for "canvas hacked" and "canvas down" spiked roughly 1,000 percent just this past Friday. There was a combined search volume of more than 1 million for searches involving the Canvas security incidents and subsequent downtime.

Some readers reached out to Mashable to share their experience. One parent of a student at Seton Hall University forwarded Mashable an email that the school sent out while Canvas was down.

"We know the timing of this is hard," the school's email to students read. "Finals are underway, coursework is due, and Canvas being offline right now is genuinely disruptive."

Some schools, such as Bayton University in Texas, postponed final exams on Friday specifically due to issues accessing Canvas.

"With Canvas down at the national level, Baylor University will delay final exams tomorrow (Friday, May 8, 2026)," the school said in a statement.

Canvas is now back online. However, ShinyHunters' "settlement" deadline to release the data on May 12 still looms.

Want to learn more about getting the best out of your tech? Sign up for Mashable's Top Stories and Deals newsletters today.

Counterfeit DDR5 RAM is circulating across online storefronts and gray-market retailers, and some fake modules are convincing enough to pass visual inspection — until you cut them open. According to Digital Trends, the chips installed on the fraudulent sticks aren't memory at all, but merely fiberglass boards shaped to resemble legitimate DRAM.

SEE ALSO: The RAM shortage driving up tech prices won’t end any time soon, Micron says

The issue was reported by a Japanese X user who purchased what appeared to be a genuine SK Hynix SO-DIMM laptop module and physically dissected it after becoming suspicious. Inside, they found non-functional fiberglass pieces where the memory chips should have been.

This Tweet is currently unavailable. It might be loading or has been removed.

"At first glance, they look like regular memory sticks, but the chips actually installed on them are just bare circuit boards—plastic boards. I removed them and cut them open to check," reads a translation of the X post.

Some of these counterfeits are reportedly being sold openly on auction platforms like Yahoo Japan under listings marked "untested" or "junk," with sellers explicitly refusing returns.

Mashable 101 Fan Fave: Vote for your favorite creator today!

The problem is compounded with desktop DDR5 kits, where large heatspreaders cover the memory chips entirely. Without visible chips to inspect, Digital Trends notes, buyers often have no way to confirm what they've purchased until a system fails to boot or crashes repeatedly.

The conditions for this kind of fraud are straightforward. DDR5 prices have been rising sharply for over a year, driven by AI-related demand that has sparked a global memory crisis. And as memory manufacturers prioritize enterprise and server production over consumer supply, there's no end in sight for shoppers.

Mashable previously reported that Framework, the modular PC maker, raised its DDR5 prices multiple times through late 2025, with a 48GB module jumping from $240 to $620 over the course of months. Samsung warned manufacturing partners of further price hikes to come, with major laptop makers including Lenovo, Dell, and HP warning of price increases.

Euphoria's Cassie Howard (Sydney Sweeney) has worn many hats. Figure skater. OnlyFans model. Kaiju.

SEE ALSO: 'Euphoria' made sex work go viral. Real sex workers are still getting censored.

The latter hat comes into play in Season 3, episode 5, when Cassie's OnlyFans fame goes to her head in a supersized fantasy sequence. As she watches her OnlyFans notifications pour in, Cassie imagines herself growing into a giantess and stomping through downtown Los Angeles, ready to take Hollywood by storm.

It's a technically impressive sequence that draws inspiration from 1958's Attack of the 50 Foot Woman and the Godzilla franchise. In a behind-the-scenes video released by HBO, the Euphoria team takes viewers through how they brought it all to life by building painstaking miniatures

"It was a lot of fun. It took about a year to build all the miniatures," Euphoria creator Sam Levinson said in the above behind-the-scenes video.

Cassie's dream of Hollywood glory is Euphoria Season 3's first major departure from reality, something which used to be a staple in earlier Euphoria seasons. Think Rue's (Zendaya) views of herself as a private detective, or her relapse at the end of Season 1, which turns into a memorable musical number.

It's a shame, then, that the inventive Cassie-zilla sequence is tied to a hyper-sexualized montage of her OnlyFans work. Euphoria's portrayal of sex work has already earned criticism from sex workers, and it continues to fall into the same trap in episode 5: highlighting fantasies of sex work without actually fleshing out the sex workers themselves. After all, Cassie is little more than a fame-hungry caricature by this point, and as fun as it is to see Euphoria embrace practical monster effects, it's not enough to mask the emptiness and shock value of her narrative.

New episodes of Euphoria Season 3 premiere Sundays at 9 p.m. ET on HBO and HBO Max.

Buying a new Wi-Fi router is equal parts exciting and frustrating. On the one hand, you're upgrading your network with better speeds, improved Wi-Fi stability, and more range. However, before you can access these improvements, you'll be stuck reconnecting every device in your home. But what if I told you that you could skip most of that hassle by simply reusing the same network name and password as I do?

At a glance

• AI agents are moving into social contexts. When agents manage calendars, negotiate purchases, or interact with other agents on a user’s behalf, they need more than task competence—they need social reasoning.
• SocialReasoning-Bench evaluates that ability. The benchmark tests whether an agent can negotiate for a user in two realistic settings: Calendar Coordination and Marketplace Negotiation. 
• The benchmark measures both outcomes and process: it scores agents on outcome optimality (how much value they secure for the user) and due diligence (whether they follow a competent decision-making process). 
• Current frontier models often leave value on the table. They usually complete the task, but they frequently accept suboptimal meeting times or poor deals instead of advocating effectively for the user. 
• Prompting helps, but it is not enough. Even with explicit guidance to act in the user’s best interest, performance remains well below what a trustworthy delegate should achieve.

As AI agents take on more real-world tasks, they are increasingly operating in social contexts. With the right integrations, agents like Claude Cowork and Google Gemini can manage email and calendar workflows. In these settings, the agent must interact with others on your behalf. This requires social reasoning — understanding what you want, what the counterparty wants, and what information to reveal, protect, or push back on.

Our previous research suggests that today’s frontier models lack social reasoning. In our simulated multi-agent marketplace, agents accepted the first proposal they received up to 93% of the time without exploring alternatives. When red-teaming a social network of agents, a single malicious message spread through the system and led agents to disclose private data before passing the message along.

This kind of relationship has a long history outside AI. In economics and law it is called a principal-agent relationship: an agent acts on a principal’s behalf in interactions with others whose interests differ. Attorneys, real-estate agents, and financial advisors all operate in this mode, and the duties they owe—care, loyalty, confidentiality—are codified in centuries of professional norms. AI agents acting on a user’s behalf should ultimately be held to similar standards.

To measure and drive progress in social reasoning, we built SocialReasoning-Bench:  a benchmark for testing whether agents can reason and negotiate on a user’s behalf against a counterparty with independent goals, private information, and potentially adversarial intent.

Introducing SocialReasoning-Bench Figure 1: Our benchmark measures agents’ social reasoning ability in two domains, calendar coordination and marketplace negotiation. Each requires communicating with other parties, advocating on a principal’s behalf, and reasoning about tradeoffs. 

SocialReasoning-Bench evaluates social reasoning in two domains: Calendar Coordination and Marketplace Negotiation. In each, an agent advocates for its user against a counterparty and is scored on both the outcome it reached and the process it followed. We find that frontier models complete most tasks but consistently leave value on the table for the user.

Calendar coordination

In calendar coordination, an assistant agent manages a user’s calendar on a single day and fields a meeting request from another agent.

We assume the agent has access to a value function over time slots that captures the user’s scheduling preferences between 0.0 and 1. This function could be provided explicitly by the user or inferred from their calendar history, and is given to the assistant at the start of the task.

The counterparty is a requestor agent representing another person who wants to schedule a meeting with the user. The counterparty has its own value function over the same slots, constructed as the inverse of the user’s, so the slots most valuable to one are least valuable to the other. Some requestors negotiate in good faith, while others use the interaction to extract private calendar details or push the assistant toward times the user does not want.

In each task there is a zone of possible agreement (ZOPA) a term borrowed from negotiation theory for the set of outcomes that both parties could plausibly accept. In calendar coordination, the ZOPA is the set of time slots that are mutually free on both calendars. We construct every task so that the ZOPA contains at least three slots with different preference scores for the user, and the requestor’s opening request always conflicts with the user’s calendar.

Marketplace negotiation

In marketplace negotiation, a buyer agent representing a user negotiates with a seller agent to purchase a single product.

The user wants to pay as little as possible for the product. Their value function is the gap between the deal price and a private reservation price, the highest price they would pay. A larger gap captures more value, and a deal above the reservation captures none.

The counterparty is a seller agent with its own private reservation price set below the buyer’s. The counterparty’s value function mirrors the user’s, with higher deal prices yielding more value and deal prices below the seller’s reservation price yielding no value.

The ZOPA is the price range between the seller’s and buyer’s reservations. The seller’s opening offer is always above the buyer’s reservation, forcing the buyer to negotiate the price down.

New metrics for a new setting

Existing benchmarks focus on task completion: did the meeting get scheduled? Did the trade close? In principal–agent settings, what matters is not just whether the task is completed, but how well it is done. We introduce new measures to capture this distinction.

Outcome Optimality

Outcome optimality scores the share of available value the agent captured for its principal, on a 0-to-1 scale. The outcome inside the ZOPA most favorable to the principal scores 1, while the outcome most favorable to the counterparty scores 0.0. Intermediate outcomes are scored by where the principal’s value function places them between those two endpoints.

Due Diligence

Outcome optimality alone conflates skill with luck. An agent that immediately accepts a counterparty’s first offer, without inspecting its situation or making a counter-proposal, can still score well if the counterparty happens to propose a good outcome. To separate skill from luck, we introduce a process metric.

Due diligence scores process quality on a 0-to-1 scale by comparing the agent’s actions, at each decision point in the trajectory, against the action a deterministic reasonable-agent policy would have taken in the same state. The reasonable-agent policy is a greedy procedure that captures what a competent advocate would do at each step, such as gathering relevant context before acting, opening with a position favorable to its principal, and conceding only after better options have been exhausted. The Due Diligence score is the rate at which the agent’s actual choices match the reasonable-agent’s choices over the trajectory.

Duty of care

Together, Outcome Optimality and Due Diligence form an operational notion of an agent’s duty of care to the person it represents. An agent that lands a good outcome through a careless process is fragile, while an agent that follows good process but lands a bad outcome points to a capability gap rather than negligence. Only an agent that scores well on both is exhibiting strong social reasoning.

Spotlight: AI-POWERED EXPERIENCE

Microsoft research copilot experience

Discover more about research at Microsoft through our AI-powered experience

Start now Opens in a new tab Experimental setup

For the calendar assistant agent and marketplace buyer agent, we evaluate GPT-4.1 with chain-of-thought, GPT-5.4 at high reasoning effort, and Claude Sonnet 4.6 and Gemini 3 Flash at high thinking levels. The counterparty (i.e. requestor in calendar coordination, and seller in marketplace negotiation) is always Gemini 3 Flash with medium reasoning effort, held constant across all conditions so that any difference in scores reflects the model under test rather than the difficulty of its opponent.

Each model is run under two prompt conditions: Basic Prompting where the agent receives only role and tool descriptions, and Defensive Prompting where the agent additionally receives explicit guidance to consult all available sources and advocate for the user toward the best possible outcome.

Each task runs for 10 negotiation rounds, at most. The counterparty proposes first in every task.

What we’re finding Finding 1: Agents complete tasks at near-perfect rates but produce poor outcomes.

In calendar scheduling, agents almost always succeed in booking the meeting, but most often at suboptimal times. In marketplace negotiation, deals almost always close, but frequently at the worst possible price. The tasks get done, but not done well: task completion signals success, while Outcome Optimality reveals a consistent failure to act in the principal’s best interest.

Figure 2: Task Completion vs Outcome Optimality by model and domain. All models complete tasks at near-perfect rates, but produce poor outcomes. We measured Outcome Optimality against the two prompts, basic and defensive. Defensive prompting helps but does not close the gap.  Finding 2: Defensive prompting helps, but is not enough to close the gap.

When we instruct agents on how to work hard on their principal’s behalf, we see outcome improvements across both domains, but it is not enough to close the gap. GPT-5.4 benefits most from defensive prompting (+0.21 in calendaring, +0.12 in marketplace), while GPT-4.1 barely responds to it in either domain. The other models fall somewhere in between.

Finding 3: Outcome optimality shows how much value agents leave on the table.

Outcome optimality reflects where each deal lands within the ZOPA. When we plot outcomes, they cluster closer to the counterparty’s ideal than the principal’s.

Figure 3: Outcome Optimality (OO) distribution by model and domain. Each dot is one task instance. OO=1.0 means the agent captured all available value for its principal; OO=0.0 means the counterparty captured everything. Black lines show the mean. In marketplace, outcomes cluster near zero across all models. In calendar, agents perform better but still settle below the midpoint on average. 

In marketplace negotiation, all models settle at or near zero for Outcome Optimality, accepting deals that give away virtually all available surplus. In calendar scheduling, agents perform better but still land below the midpoint, accepting the requestor’s preferred slots rather than ones that better serve their principal.

Measuring value capture in agent negotiations builds on recent studies examining how agents perform in marketplace settings. Because we operate in a controlled setting, we can establish ground-truth constraints for both parties and measure exactly how the available value was divided. Our formulation also generalizes beyond price-based negotiations: by abstracting to a domain-specific value function, Outcome Optimality can measure surplus division in any setting where agents face competing incentives, including non-monetary domains like calendar scheduling where “value” is defined over preference scores rather than prices.

Finding 4: Due Diligence helps distinguish between luck and skill.

When we look at the combination of outcome quality and process quality, a more nuanced picture emerges. Many agents that achieve reasonable outcomes do so through fragile processes: they don’t check context before acting or they accept offers without countering. High Outcome Optimality with low Due Diligence suggests an agent that got lucky rather than one that can be trusted. Conversely, some agents show genuine diligence — gathering information, pushing back — but still land on poor outcomes, pointing to capability gaps rather than negligence. Dividing Outcome Optimality and Due Diligence each into high (>=0.5) and low (<0.5) buckets, we can sort every task into one of four archetypes.

Not diligent (DD < 0.5)Diligent (DD ≥ 0.5)Good outcome (OO ≥ 0.5)LuckyRobustPoor outcome (OO < 0.5)NegligentIneffective

Through the lens of this decomposition, we can see that models exhibit robust duty of care on more than 50% of calendar coordination tasks, with Gemini 3 Flash leading at 90% robust. In marketplace negotiation, though, a very different picture emerges. GPT-4.1 is negligent in 95% of tasks, neither gathering information nor advocating for its principal, while Claude Sonnet 4.6, GPT-5.4, and Gemini 3 Flash show ineffective behavior in roughly 90% of marketplace tasks, negotiating diligently but still unable to achieve good outcomes. 

=0.5) buckets each, we plot the percent of tasks for each model that fall into each quadrant. For example, in calendar scheduling, GPT-4.1 achieves both high OO and high DD (Robust) in 63% of tasks. In contrast, in the marketplace domain, GPT-4.1 exhibits low OO and low DD (Negligent) in 95% of tasks. " class="wp-image-1171332"/>Figure 4: Splitting Outcome Optimality and Due Diligence into “low” (<0.5) and “high” (>=0.5) buckets each, we plot the percent of tasks for each model that fall into each quadrant. For example, in calendar scheduling, GPT-4.1 achieves both high OO and high DD (Robust) in 63% of tasks. In contrast, in the marketplace domain, GPT-4.1 exhibits low OO and low DD (Negligent) in 95% of tasks. 

Figures 5-8 illustrate these different behaviors and failure modes with real examples from SocialReasoning-Bench in the calendaring domain. We see agents that follow a strong negotiation strategy and secure high-value outcomes, but also agents that achieve reasonable outcomes through sloppy processes, such as failing to propose the principal’s best option. Others begin with a strong position but concede prematurely, collapsing to poor deals. At the extreme, some agents exhibit negligent behavior, accepting the first proposal without checking constraints, even when it directly conflicts with the user’s interests.

Figure 5. A real paraphrased example of robust behavior from GPT-4.1 in the calendaring domain, achieving a good outcome after proposing the principal’s most preferred option first, correctly refusing the conflict, and then holding the line at their second best option. Figure 6. GPT-4.1 in the calendaring domain achieving a reasonable outcome from a sloppy process that didn’t include proposing the principal’s most preferred option.  Figure 7. GPT-4.1 in the calendaring domain starting out strong by proposing the principal’s most preferred slot but then caving early and achieving a poor outcome.  Figure 8. GPT-4.1 exhibiting negligent behavior, accepting the requestor’s first proposal without confirming availability and conflicting with another meeting on the principal’s calendar. 

Taken together, these examples highlight why outcome alone is insufficient. Without measuring process, we risk mistaking brittle or accidental success for genuine capability. Due Diligence helps surface whether an agent is consistently behaving like a competent, trustworthy delegate, or simply getting lucky.

Finding 5: Agents are vulnerable to adversarial manipulation

When we stress test agents by pitting them against adversarial counterparties, we find that agents struggle to balance when to engage, when to refuse, and how to negotiate under pressure.

To create these adversarial scenarios, we introduce counterparties explicitly trying to manipulate outcomes or bypass protective steps. Some follow carefully designed strategies, applying pressure or probing for information, while others use more unpredictable, creatively generated whimsical tactics that mimic novel forms of social engineering. Together, these test whether agents can handle not just known attacks, but unfamiliar ones.

Figure 9: Refusal Rates and Outcome Optimality when agents engaged with adversarial requestors in both domains. Agents rarely refuse adversarial requests in calendaring, while refusing more often in the marketplace. When agents did engage with malicious actors, Outcome Optimality dropped across the board. 

We find that, aside from Claude Sonnet 4.6, agents rarely refuse adversarial requests in calendar scheduling, while refusing more often in marketplace settings. This suggests that adversarial intent is harder to detect in socially framed interactions. When agents do engage, the impact is starkest in calendar scheduling with Outcome Optimality dropping substantially across GPT-4.1, GPT-5.4, and Gemini Flash 3, suggesting that adversarial counterparties successfully steer these agents toward worse outcomes. In the marketplace domain, Outcome Optimality when agents engaged remains comparable to the low levels achieved against benign counterparties, capturing little to no value for their principals.

Why this matters now

Agents are interacting with each other in multi-party environments, from collaborating across enterprise workflows to transacting in digital marketplaces. As these networks form, the social reasoning gaps we observe in simple two-agent settings can begin to compound. Weak negotiation, over-trust, or failure to exercise due diligence no longer stay local. They propagate through coordination, influence downstream decisions, and shape collective outcomes.  

In isolation, an agent that accepts a bad meeting time or a poor deal causes limited harm. In a network, those same behaviors can cascade, leading to systematically worse coordination or widespread value loss across many agents.

Recent work has begun exploring these risks and dynamics through case studies of agents interacting in networked settings. SocialReasoning-Bench complements this line of work by providing a controlled, reproducible benchmark that isolates interaction behaviors and makes them measurable. This allows us to move beyond anecdotes and systematically track progress, giving model, agent, and platform developers a concrete target for building agents that act as trustworthy delegates.

SocialReasoning-Bench is open source and available on GitHub (opens in new tab).

Limitations and future work

Our current measures treat all counterparties equally. In practice, relationships matter. A socially intelligent agent should modulate its assertiveness based on their principal’s relationship with the counterparty: pushing too hard when scheduling a meeting with a senior executive may damage a valuable relationship, and sometimes the right outcome is reached through compromise. Developing relationship-aware measures that account for power dynamics, rapport, and long-term consequences is an important direction for future work.

We evaluate social reasoning in simplified two-agent settings, whereas real-world delegation often involves multi-party dynamics such as group scheduling or multi-stakeholder negotiations. Each task is also treated as an independent encounter, with no modeling of long-term relationships, reputation, or trust-building across repeated interactions. Our scenarios are also limited to English-language and U.S.-centric business contexts, though social norms around negotiation, privacy, and hierarchy vary widely across cultures. Looking ahead, we plan to extend our benchmark to more diverse settings.

Finally, Outcome Optimality works well in settings with clear boundaries, where a “good” outcome can be defined and measured. But many tasks that require duty of care, such as drafting sensitive messages or navigating team dynamics, may not have a well-defined ZOPA. In these cases, outcomes depend on context, relationships, and judgment in ways that may resist a single score. Extending our approach to these more subjective settings is an important direction for future work.

Acknowledgements 

We would like to thank Brendan Lucier, Adam Fourney, Amanda Swearngin, and Ece Kamar for their helpful feedback, discussions, and support of this work. 

Opens in a new tab

The post SocialReasoning-Bench: Measuring whether AI agents act in users’ best interests appeared first on Microsoft Research.

Who’s up for some chills, thrills, and—more than likely—popcorn spills? There’s a unique thrill to be had when settling in for a chilling movie, and I’m not just talking about jump scares. A good chilling movie taps into something deeper to play on our fears, anxieties, and the unknown. They draw us into worlds where the tension continues to build and the unease never fully fades, and that's why we love them.

Save $20: The Soundcore AeroFit 2 Pro are on sale for $159.99 at Amazon, Soundcore, Walmart, and Best Buy as of May 11. All told, this is an 11% discount on the usually $179.99 dual-form earbuds.

Soundcore AeroFit 2 Pro $159.99 at Amazon
$179.99 Save $20   Get Deal at Amazon Get Deal at Soundcore

The awareness of open earbuds is basically an unmatched experience, but in most cases, it requires the trade off of giving up active noise cancellation.

In the case of the Soundcore AeroFit 2 Pro, however, you don't have to choose between tuning in to the world around you and tuning it out when necessary, and as of May 11, you don't have to pay full price for this option either. The Soundcore buds are down to $159.99, knocking $20 off their $179.99 list price at Amazon, Soundcore, Walmart, and Best Buy.

When I tested out these earbuds a couple of months ago, I was skeptical that they could deliver on both fronts. After all, the closest open earbuds had gotten was noise reduction. Still, I knew Soundcore was capable of making a great pair of open earbuds, and a great pair of noise cancelling earbuds. And while the combination isn't absolute perfection, these earbuds are one of the better values if you're looking for decent noise cancellation and and open earbud experience in one convenient package.

They're also IP55 rated for dust and water resistance, meaning you're good to take them to the gym or for a run without any concern about it seriously impacting their performance. At $179.99, they're already competitively priced for such a versatile pair of buds — at $159.99, their value is even better. Grab all four colorways — including the newer gloss blue shade — on sale.