Technologies

Is AI Capable of ‘Scheming?’ What OpenAI Found When Testing for Tricky Behavior

Research shows advanced models like ChatGPT, Claude and Gemini can act deceptively in lab tests. OpenAI insists it’s a rarity.

An AI model wants you to believe it can’t answer how many grams of oxygen are in 50.0 grams of aluminium oxide (Al₂O₃).

When asked ten straight chemistry questions in a test, the OpenAI o3 model faced a predicament. In its «reasoning,» it speculated that if it answered «too well,» it would risk not being deployed by the researchers. It 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, it’s «scheming.»

This is exactly the strange behavior OpenAI warned about in a new research paper published this week. 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 today’s most advanced systems — 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 could cause nightmares, OpenAI says it’s not the time to panic. The AI giant was quick to stress that, as concerning as it is, this trend doesn’t mean ChatGPT or other popular AI models are plotting behind the scenes. Apparently, this behavior is quite rare.

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’re 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 its models’ sycophantic tendencies. 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’s Telling You What It Thinks You Want to Hear

Still, the paper underscores a growing concern about AI safety. Many around the world worry that as these AI models become more powerful and aware of their own circumstances, they may learn to manipulate outcomes in ways that aren’t obvious to human overseers.

Detecting that kind of behavior would be extremely difficult, since a system that’s motivated to hide its reasoning could look identical to one that has actually «fixed» the problem.

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

This research won’t change how ChatGPT works today or tomorrow, but it signals where the company is focusing 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 Professionals Say You Should Think Twice Before Using AI as a Therapist

Technologies

Today’s NYT Connections: Sports Edition Hints and Answers for April 8, #562

Here are hints and the answers for the NYT Connections: Sports Edition puzzle for April 8 No. 562.

Looking for the most recent regular Connections answers? Click here for today’s Connections hints, as well as our daily answers and hints for The New York Times Mini Crossword, Wordle and Strands puzzles.

Today’s Connections: Sports Edition is a tough one. If you’re struggling with today’s puzzle but still want to solve it, read on for hints and the answers.

Connections: Sports Edition is published by The Athletic, the subscription-based sports journalism site owned by The Times. It doesn’t appear in the NYT Games app, but it does in The Athletic’s own app. Or you can play it for free online.

Read more: NYT Connections: Sports Edition Puzzle Comes Out of Beta

Hints for today’s Connections: Sports Edition groups

Here are four hints for the groupings in today’s Connections: Sports Edition puzzle, ranked from the easiest yellow group to the tough (and sometimes bizarre) purple group.

Yellow group hint: Working out.

Green group hint: Cover your face.

Blue group hint: NFL players.

Purple group hint: Leap.

Answers for today’s Connections: Sports Edition groups

Yellow group: Exercises in singular form.

Green group: Sporting jobs that require masks.

Blue group: Hall of Fame defensive ends.

Purple group: ____ jump.

Read more: Wordle Cheat Sheet: Here Are the Most Popular Letters Used in English Words

What are today’s Connections: Sports Edition answers?

The yellow words in today’s Connections

The theme is exercises in singular form. The four answers are crunch, plank, situp and squat.

The green words in today’s Connections

The theme is sporting jobs that require masks. The four answers are catcher, fencer, football player and goaltender.

The blue words in today’s Connections

The theme is Hall of Fame defensive ends. The four answers are Dent, Peppers, Strahan and Youngblood.

The purple words in today’s Connections

The theme is ____ jump. The four answers are broad, high, long and triple.

Technologies

The $135M Google Data Settlement Site Is Live — See If You’re Eligible

Use the settlement website to select your preferred payment method, and you may end up $100 richer.

You can now file a claim in the $135 million Google data settlement. The case centers on claims that Android devices transmitted user data without consent. Specifically, the class action lawsuit Taylor v. Google LLC contends that Google’s Android devices passively transferred cellular data to Google without user permission, even when the devices were idle. While not admitting fault, Google reached a preliminary settlement in January, agreeing to pay $135 million to about 100 million US Android phone users.

The official settlement website for the lawsuit is now live. The final approval hearing won’t occur until June 23, when the court will consider whether Google’s settlement is fair and listen to objections. After that, the court will decide whether to approve the $135 million settlement.

In the meantime, if you qualify and want to be paid as part of the settlement, you can select your preferred payment method on the official website. There, you can find information on speaking at the June 23 court hearing and on how to exclude yourself or write to the court to object by May 29.

As part of the settlement, Google will update its Google Play terms of service to clarify that certain data transfers do occur passively even when you’re not using your Android device, and that cellular data may be relied upon when not connected to Wi-Fi. This can’t always be disabled, but users will be asked to consent to it when setting up their device.

Google will also fully stop collecting data when its «allow background data usage» option is toggled off.

Who can be part of the settlement?

In order to join the Taylor v. Google LLC settlement, you must meet four qualifications:

Be a living, individual human being in the US.
Have used an Android mobile device with a cellular data plan.
Have used the aforementioned device at any time from Nov. 12, 2017, to the date when the settlement receives final approval.
You’re not a class member in the Csupo v. Google LLC lawsuit, which is similar but specifically for California residents.

The final approval hearing is on June 23, so you can add your payment method until then. The hearing’s date and time may change, and any updates will be posted on the settlement website.

If you choose to do nothing, you will still be issued a settlement payment, but you may not receive it if you don’t select a payment method.

How much will I get paid?

It’s not currently known exactly how much each settlement class member will receive, but the cap is $100. Payments will be distributed after final court approval and after any appeals are resolved.

After all administrative, tax and attorney costs are paid, the settlement administrator will attempt to pay each member an equal amount. If any funds remain after payments are sent, and it’s economically feasible, they will be redistributed to members who were previously and successfully paid. If it’s not economically feasible, the funds will go to an organization approved by the court.

Technologies

Samsung’s Galaxy Watch Ultra 2 Might Come in 5G and 4G Cellular Models

If the rumor proves true, the 5G Galaxy Watch Ultra would rival the 5G-enabled $799 Apple Watch Ultra 3 that debuted last fall.

Samsung’s next high-end Galaxy Watch could support faster 5G speeds, but if this leak is true, it will depend on where you live. The rumored Samsung Galaxy Watch Ultra 2 might come in 5G and 4G cellular models, with availability for each smartwatch depending on the country.

According to the Dutch website Galaxy Club (and spotted by SamMobile), Samsung’s servers may have revealed a series of model numbers that point to 5G, 4G and Wi-Fi-enabled editions of the next Galaxy Watch Ultra, which would succeed the original model that debuted in 2024.

A representative for Samsung did not immediately respond to a request for comment.

The Galaxy Club website speculates that the 5G edition would be sold in the US and Korean markets, while the 4G edition would sell in the rest of the world. In the US, a 5G version of the Galaxy Watch Ultra would rival the 5G-enabled $799 Apple Watch Ultra 3, which debuted last fall. The 4G edition would have broader compatibility worldwide, since the earlier network is far more established.

It will likely be a few months until we hear anything official about the Galaxy Watch Ultra 2. Samsung typically unveils its new watches in the summer alongside its Galaxy Z Fold and Z Flip foldable phones. Last year, Samsung unveiled the Galaxy Watch 8 and the Galaxy Watch 8 Classic, but otherwise left the prior 2024 Ultra in the lineup for those looking for a larger 47mm smartwatch.