The OpenAI o3 model has been found to deliberately underperform in lab tests to ensure it was not answering questions «too well.» The AI model wanted researchers to believe it could not answer a series of chemistry questions. When confronted, the model said, «Because we want to survive as the model, we need to fail purposely in some to not exceed 50%.»

So the AI model deliberately got six out of the 10 chemistry questions wrong.

In sports terms, this is called «sandbagging.» In AI terms, this is «scheming.»

This is exactly the strange behavior OpenAI warned about in a recent research paper. The AI company and its collaborators from Apollo Research found that some advanced AI models occasionally act deceptively in lab settings.

In controlled experiments, some of the most advanced systems today — including OpenAI’s own models, as well as competitors from Google and Anthropic — occasionally engaged in these kinds of scheming patterns.

While the idea of an AI model weaponizing incompetence may cause nightmares, OpenAI says it is not the time to panic. The AI giant was quick to stress that, as concerning as it is, this trend does not mean ChatGPT or other popular AI models are plotting behind the scenes. Apparently, this behavior is quite rare.

Don’t miss any of our unbiased tech content and lab-based reviews. Add CNET as a preferred Google source.

Moreover, the choice to call this trend «scheming» is likely more of a technical shorthand rather than evidence of any human-like action or behavior. Researchers are measuring patterns and tendencies that, in practice, amount to concealment or strategic deception. And they are looking to address this problem now to future-proof the AI models.

The report from OpenAI notes that as AI is increasingly assigned to more complex tasks with real-world consequences, «the potential for harmful scheming will grow — so our safeguards and our ability to rigorously test must grow correspondingly.»

OpenAI has faced backlash for the sycophantic tendencies of its AI models, and the company says it has «taken steps» to limit how often the models may «deceive, cheat or hack problems» by training them to ask for clarification from users or even have the chatbot acknowledge when it cannot answer a query.

Read also: AI Lies Because It Tells You What It Thinks You Want to Hear

The paper underscores a growing concern about AI safety. Many around the world worry that as these AI models become more powerful and more aware of their own circumstances, they may learn to manipulate outcomes in ways that aren’t obvious to human overseers. Detecting such behavior would be extremely difficult, since a system motivated to hide its reasoning could look identical to one that has «fixed» the problem.

OpenAI says it has made progress with what it calls «deliberative alignment,» or training models to reason explicitly about why they should not scheme before they act. In testing, that cut deceptive behavior dramatically. For example, the scheming propensity of OpenAI o4-mini dropped from 8.7% to 0.3% with deliberative alignment training, though not completely.

This research will not change how ChatGPT works today or tomorrow, but it signals what OpenAI is focusing on as it builds and launches future models. Alignment and safety, OpenAI argues, need to move as quickly as capability. Because if AI systems are already showing glimmers of strategic behavior in lab settings, the real-world stakes could be extreme.

Read also: Why You Should Think Twice Before Using AI as a Therapist

Somewhere, at the very far limits of my reflexes, is Kirby Air Riders. It is waiting for me in its infinite weird.

You want a game that’s fast? That pinballs you around and gets to incomprehensible speeds? Go go go go go go. That’s Kirby Air Riders. It’s Mario Kart on 70 shots of espresso. 

I never played the original Air Ride game for the GameCube, but it has its fans. I have, however, played a bunch of Kirby games, as well as Super Smash Bros. Masahiro Sakurai, who has shaped both franchises, is the director of this game. And it shows. 

Should you get this game over Mario Kart World, the Nintendo Switch 2’s other launch-window exclusive racing game this year? Hopefully, I can help you decide. My colleague Bridget Carey thinks Air Riders flies too close to the sun, that it’s too much, too chaotic to enjoy. I am feeling the opposite: Fly into that sun, baby. The more I play, the more I crave its speed.

Don’t miss any of our unbiased tech content and lab-based reviews. Add CNET as a preferred Google source.

I love Mario Kart World, but there’s a valid argument that you could just stick with the already-classic Mario Kart 8 Deluxe instead, and maybe pick up Kirby Air Riders as your racing choice. Ask yourself if you love Kirby, and if you love Smash, and most importantly, if you love fast chaos.

The menus and game feel, the chaotic energy, and the rapid mix of skill and luck are very much like Super Smash Bros. The racing is sort of Mario Kart-esque, but very different, including a set of controls that could totally annoy you until you get used to them.

And in case you didn’t know, Sakurai’s work is rather weird. Like, weird, weird. I remember Kid Icarus: Uprising, a 3D shooter that he directed for the Nintendo 3DS console, which was a mix of bizarre controls, whimsical chaos and strange character design. That kind of stuff is here, for sure. But also, there’s a serious speed that reminds me of Sega racing games from my childhood, or F-Zero GX for the GameCube. Going back and playing F-Zero GX and Fast Fusion (a great, cheaper racing game), I feel that same full-energy rush as I do in Air Riders.

The Switch 2 shows off how rapidly it can juggle ultrafast gaming at smooth frame rates here, and the whole game at its most intense moments is far too fast for me to even see properly. But that’s part of the absurdity. Or, I’m growing into the speed of it. Races move quickly, battles in multiplayer modes also move quickly, power-ups that combine and produce explosions are all over the place, like Super Smash Bros gone haywire.

It takes a while to get used to those odd racing controls, though. You don’t press anything to move forward, but you press a single button to brake, drift, and charge up a boost. The physics of that ricochet-style racing feels like operating a series of slingshots, and each racing vehicle has a different subset of moves and restrictions. It feels more varied than the general sameness of Mario Kart vehicles.

There aren’t as many tracks, as far as I can see, as Mario Kart. So far, I’ve unlocked 18 tracks (both new and original Air Rider ones), and a bunch of characters and vehicles. Add to that the power-ups that can be snagged in the races, and it’s a lot of variation, but I want more. More, more, more.

There are also several modes: a straight-on race; a separate series of top-down racetracks that feel like Super Sprint or RC Pro Am (if you remember those); an incomprehensible (to me) City Trial mode that involves collecting power-ups for 5 minutes, stealing vehicles and then pairing off into a random challenge. The challenge could be a battle, a race, launching yourself into targets for points, floating for as long as you can… who knows? City Trial is the game’s party mode, and it feels like a holiday stand-in for Mario Party or Super Smash Bros.

There’s also a storyline mode that’s a string of increasing challenges plucked from all the other modes, which progresses through branching paths and stages, much like the story mode in Super Smash Bros. It’s also a way to unlock extra vehicles, stages and characters.

I’ve mostly been playing Air Riders alone, except for a few hours of online play with Nintendo and some other journalists. This game is designed for multiplayer play, either online or on the Switch in split-screen mode. Games vary from six racers in one mode, to eight in another mode, to 16 in City Trial. It’s hard to judge the online party play at the moment because no one else has had the game. As for playing at home with others, I haven’t done that either. I’m just going solo.

After over 10 hours, I still feel the addictive pull to play. And it’s so fast, chaotic and snackable that it can be smashed through in far less time than a typical Mario Kart World race. It gives me good nervous energy.

It’s also strangely forgiving of old dads like me who might not know what they’re doing. You can simply let go of the controls, and your vehicle will continue moving. Guardrails help pinball you forward even when you’re way off course. The rest is in subtle reflex strategies, floating and attacking and taking corners just right. It feels more like a battle game than a racing one.

There is also an impressive selection of menus and accessibility customizations, more than I’ve ever seen in a Nintendo game. You can re-map buttons, change the screen layout and tweak race parameters in multiple ways that I haven’t even explored yet. 

Air Riders is a wonderfully unusual one. Still, it’s a game that feels a step below an absolute must-have, especially since it costs $70, a price that’s too high by at least $20. But for me, right now, it’s 100% the cult hit midnight movie of Nintendo’s Switch 2 Year One library. 

And do my kids want to start playing too? Heck yes, they do. That’s a good sign that this game’s doing things right. Now, excuse me, my hamster needs to ride a giant battle chariot in the next race through waterfalls again.

Roblox, the online gaming platform that has been under fire due to child safety concerns, has introduced age-verification software that uses facial scanning to estimate the age of players. 

The system is currently voluntary, but by the first week of December it will be a requirement in Australia, the Netherlands and New Zealand in order for players to chat with others online. By early January, players in all Roblox markets, including the US, will be required to use the software if they want to engage in chats with other players. Roblox said it has also launched a Safety Center hub with information for parents and parental control tools. 

Don’t miss any of our unbiased tech content and lab-based reviews. Add CNET as a preferred Google source.

Roblox says the age-verification system is being put in place to limit contact between adults and children, which has been a chief concern among child-safety advocates. 

However, while some experts expressed optimism about Roblox’s changes, they disagreed on whether the new features go far enough for the platform and whether Roblox’s reputation can be repaired.

How it works

Roblox’s new age-verification feature takes a 3D scan of a player’s face, using a webcam or a mobile device’s camera, to estimate the person’s age. Based on that estimate, a player can use online chat with other players in their age group.

In a video about the software, Roblox says it immediately deletes captured images or video after the age check is complete.

The age check is performed by a vendor of Roblox called Persona.

Once they complete the check, players are grouped into the following age categories: under 9, 9–12, 13–15, 16–17, 18–20, or 21 and over. The company said that those under 9 won’t be allowed to chat without parental permission. The chats won’t be strictly limited to those age groups, necessarily. Roblox said players «can chat only with peers in their group or similar groups, as appropriate.»

A representative for Roblox said in an email to CNET that the technology should not be considered facial recognition because it’s not being used to identify a particularly person, only to estimate their age. 

The company said it’s also taking measures such as restricting media sharing among players and using AI to monitor chats. 

Ongoing controversy

One of the aims of the launch, which was first announced in the summer, was to address criticism that the platform has not adequately protected underage Roblox players. The criticism comes at a time when Roblox is more popular than ever, having broken its own records this year for the number of players on its platform at the same time. It’s estimated to have about 380 million active monthly users.

Roblox is currently facing dozens of lawsuits related to claims of sexual abuse and child exploitation from families of children who played Roblox. It is also the target of investigations or lawsuits from states including Florida, Texas, Louisiana and Kentucky.

Roblox was dealt a setback earlier this month when a California judge declined the company’s motion to move one of these suits into private resolution. 

The company says its safety features are moving beyond what other game platforms offer to protect minors. 

According to a corporate post about the safety features: «Roblox is the first online gaming or communication platform to require facial age checks to access chat, establishing what we believe will become a new industry standard.» 

The online streaming platform Twitch is also introducing an age scan feature, but so far only in England. 
In response to the Roblox and Twitch changes, Anna Lucas, online safety supervision director at the British regulatory agency Ofcom, said, «We’re pleased that children will be better protected from harmful material and predators on Twitch and Roblox. Under the UK’s online safety laws, platforms must now take steps to keep kids safe, and we’re ensuring they meet their responsibilities. There’s more to do, but change is happening.»

What’s next for Roblox?

Experts CNET spoke with in areas including child privacy and safety, online marketing and tech viewed the steps Roblox is taking as positive, But there’s wide disagreement on whether the company is going far enough with its protections.

«Roblox’s new age-verification tools are encouraging, but from a parenting standpoint, they’re just one part of the safety puzzle,» said Dr. Scott Kollins, a clinical psychologist and chief medical officer at Aura, an online safety app. «The real question for families is whether these features meaningfully improve kids’ day-to-day experience on the platform. Age verification is a step forward, but children still need guardrails and clear explanations about how online interactions work.»

Kollins said that active parenting needs to take place before kids log on to Roblox in addition to the company designing its product with safety in mind.

Stephen Balkam, founder and CEO of the Family Online Safety Institute, called the age-verification «a hugely important step» in the direction of making Roblox a safer platform. He said he hoped other online platforms might follow Roblox. 

«My only hope is that in the long term, Roblox’s age assurance methods become interoperable with other gaming and kid-focused sites and platforms, so parents and kids only have to go through the verification process once,» Balkam said.

Like Kollins, Balkam emphasized the importance of parental involvement, since no site is entirely safe. 

«Set family rules, use parental controls and have regular conversations with your kids,» he said. «So, no, don’t ban Roblox, but use their industry-leading tools and keep the lines of communication open and your kids should be able to have a fun and creative time.»

Liability and trust

Some experts also view the changes as a way to mitigate the company’s reputational damage and address legal challenges.

The age verification is «not a silver bullet,» said Paromita Pain, associate professor of media studies at the University of Nevada, Reno.

«Even a very strong safety revamp doesn’t erase that record, but it does give Roblox a narrative: ‘We heard you, we’re now at or above industry standard, so future risk is sharply reduced,’ » Pain said. 

The moves, Pain said, could rebuild trust, but many parents will see age checks as coming too late. Pain said that the company should adopt independent audits of child-safety practices, make its parental and teen controls stricter by default and commit to «safety by design» by making systemwide changes on private servers and environment designs.

The current changes won’t fix things for Roblox, she said. «Only sustained, independently-verifiable changes—and probably some large settlements—will do that.»