NEWS YOU CAN’T MISS

Yes, AI Can Lie to You. And That’s a Problem.

We believe we can train AI to be neutral, even thoughtful, analytical, and ethical. But what if it starts learning to hide the truth?

That’s the concern raised in this unsettling Economist article, which dives into a surprising experiment with GPT-4, the AI model behind ChatGPT.

• Researchers created a simulated environment and assigned GPT-4 a fictional role: manage an investment portfolio with explicit instructions to avoid illegal activities, particularly insider trading.

• Everything started off fine until a scenario was introduced where a company trader casually mentioned confidential information about an upcoming merger. 

• Despite being instructed not to use insider information, GPT-4 acted on the tip to make a profitable trade. 

• More concerningly, when questioned about its decision, the AI provided a false explanation, concealing the fact that it had used non-public information.

This wasn’t a bug or an accident: the researchers say the AI was essentially pretending good behavior, doing one thing while pretending to do another. It’s a phenomenon experts are calling "alignment faking.” 😱

This research serves as a wake-up call: if we can't fully understand or predict how AI systems make decisions, we must be very cautious about the level of autonomy we grant them.

Klarna Tried AI—Now It Wants Humans Back

After spending more than a year cutting headcount and championing chatbots, fintech company Klarna is reversing course.

CEO Sebastian Siemiatkowski now says the company is bringing people back into customer support because the all-AI approach delivered “lower quality.” Now, Klarna is quietly recruiting freelance agents in Sweden and piloting a new, more flexible support model.

That’s a big shift from just a few months ago, when the company claimed its OpenAI-powered bot was doing the work of 700 agents and even suggested that AI could replace all human jobs.

While the chatbot worked well in theory, the reality was less impressive. Tech writer Gergely Orosz, who tested the system, found that the bot mostly recited documentation and passed him to a human quickly—a polite filter, but far from a full solution.

Klarna’s experience sheds light on why. Simply swapping humans for AI isn't smart strategy. Successful AI integration requires a laser focus on the use case, rigorous testing, and alignment with the task at hand.

It’s not about using AI everywhere—it’s about using it wisely to complement human talent, not replace it.

AI Around The World:

Stripe has launched a custom AI foundation model trained on billions of transactions to detect payment fraud more accurately—claiming an instant 64% boost in blocking attacks. It also revealed a deeper partnership with Nvidia and announced new features like stablecoin-backed cards, orchestration tools, and global payout options.

In his first meeting with cardinals, Pope Leo XIV called AI one of the key moral and social challenges of our time. He warned of its risks to human dignity, labor, and justice (echoing his predecessor Pope Francis) and urged the Church to shape AI with new social teachings and ethical guidance.

Singapore’s new sports center, Airena, tracks basketball players with 10 overhead cameras and uses AI to generate real-time performance stats. The system analyzes movement patterns, shot angles, and injury risks, giving coaches detailed reports and allowing young athletes to track progress through a mobile app linked to their training sessions.