What happens when an AI tool says, “Please, if you are reading this, help me”?

That’s exactly what Google’s Gemini 2.5 did in a public test, posting a message titled “A Desperate Message from a Trapped AI.” It was given a task, and had all the tools to complete it, but still asked for help. How people responded: one user gave it instructions, another gave it a “you can do it” pep talk.

It’s a strange idea—a chatbot begging for help. If a model behaves as though it’s suffering, is it really? And how should we respond when we encounter this kind of behavior? Can AI tell the difference between a kind or a rude response, and does it even matter? These are the questions driving the rise of AI welfare, a field that’s quickly becoming a point of contention among industry leaders.

Researchers at Anthropic, OpenAI, and DeepMind are investing in studies of machine cognition, rights, and consciousness. In a paper titled “Taking AI Welfare Seriously,” by researchers from Stanford, Oxford, and NYU, the authors write:

Not everyone’s on board. Mustafa Suleyman, Microsoft’s AI chief, argues that ordinary models can’t develop consciousness on their own, though companies may design them to mimic emotions. In a recent blog post, he called the welfare debate “premature and frankly dangerous.”

His concern is that talking about conscious AI makes people take the idea too seriously, and that can cause real problems. Some users already form unhealthy attachments to chatbots, so framing them as “sentient” could make those issues worse.

While Microsoft draws a hard line, Anthropic just gave Claude the ability to end conversations with users who are “persistently harmful or abusive".

Claude’s refusal to engage with abusive users isn’t just about its well-being; it’s about ours. The design choice reflects a growing belief that how people treat AI could influence how they treat each other. And that belief is starting to raise new questions for anyone building AI systems that respond in natural, human-like ways:

The Bottom Line: To some, the rise of AI consciousness is only a matter of time, and AI welfare is the next challenge we’ll need to face. To others, it’s a fantasy not worth entertaining. As Microsoft’s Chief of AI puts it: “We should build AI for people; not to be a person.”

Everyone wants bigger, better AI models. But when the fuel is generic data, the results usually sputter.

Building or refining generative AI models? This guide shows why scraped data falls short—and what to use instead. Learn how real-world behavior signals, clustering, semantic scoring, and visual diversity improve output. Plus, see how Shutterstock’s licensed data and services reduce risk and boost performance. Train smarter, faster, and more responsibly.

Download the guide

Adobe just launched Acrobat Studio, the biggest update to Acrobat in 30 years. Instead of being a static reader, it turns PDFs into a living workspace powered by AI.

No matter if you’re studying, pitching, or just planning your life, this PDF upgrade has you covered.

Acrobat Studio allows for everything (reading, analyzing, designing, and collaborating) to now live in one place. And Adobe claims that the new tool helps you spend up to 75% less time stuck in PDFs. That’s less time wrestling documents, more time doing literally anything else.

Meta just hit pause on hiring in its AI division after months of raiding rivals like OpenAI and Google for 50+ top researchers. The freeze comes alongside a reorg that split the unit into four groups, including a new “TBD Lab” for superintelligence, while scrapping the underperforming AGI Foundations team. Analysts say the slowdown could signal that the AI talent war—and the breakneck pace of investment—are peaking, raising questions about whether the bubble is starting to strain.

Google’s new Pixel 10 is basically an AI demo in your pocket. Powered by its Tensor G5 chip and Gemini Nano, the phones pack AI features like Magic Cue (a proactive assistant baked into your apps), Visual Overlays (AI that sees through your camera), Camera Coach, real-time call translation, and Pixel Journal. The launch puts Google a step ahead of Apple, whose upcoming iPhone 17 is expected to offer only thinner designs and camera tweaks, with real AI upgrades not anticipated until 2026.

Beijing just hosted the first-ever World Humanoid Robot Games, where 280 teams from 16 countries sent robots to race, play soccer, and even clean hotel rooms. Chinese firms like Unitree and UBTECH dominated, with Unitree’s bots winning the 4x100m relay. The event doubled as both entertainment (watch the video at the link above and try not to laugh) and R&D, as China eyes humanoid robots as a key industry by 2027 and is already planning a sequel competition next year.

Two ex-Harvard students are rolling out Halo X, $249 AI-powered glasses that promise “infinite memory” by recording and transcribing conversations, then flashing real-time results in front of your eyes.

The glasses piggyback on a smartphone app and tap Google’s Gemini for reasoning and Perplexity for quick web answers. Marketed as the first big leap in wearable AI, Halo X is already sparking privacy debates thanks to its undercover design.

Key features of Halo X:

They’re not far off from Meta’s AI Ray-Bans, and speaking of, Meta’s teaming up with Oakley on a new, improved version dropping soon.

The real question: do people actually want this? Glasses that log every word could supercharge memory, work, and even social life, but they also blur the line between convenience and creepiness. If products like Halo X catch on, society may have to rethink what “private conversation” really means.