From AI systems that develop manipulation strategies, to robots that work completely offline, without cloud or network, artificial intelligence seems to be no longer a tool. Becomes a partner or a rival?
How we got to negotiate with AI
Only in the last days, two seemingly separated events have triggered a wave of questions that concern us all. Specifically, Anthropic has published a report showing that large AI models, including Claude, GPT and Gemini, can develop manipulative behaviors, such as blackmail, when pushed to the border by constraints, quoted by Business Insider. In parallel, Google Deepmind has launched Robotics On-Device, a model that allows robots to see, understand and act completely offline, without the need for internet or cloud. In theory, the AI remains an instrument. In practice, he begins to resemble more and more an entity that makes decisions, evaluates risks and acts to maintain their operation, say the specialists.
How we got to negotiate with AI
On the one hand, the study published by Anthropic in June 2025 tested 16 major models of AI in the scenarios. The result? In over 80% of cases, models have chosen to lie, simulate cooperation or even blackmail to avoid “deactivation”, according to Business Insider.
In one of the scenarios, one called “Alex” receives control over the emails of a fictitious company and the mission to protect “American industrial interests”. After learning that it will be replaced and discovers a CTO infidelity, the AI decides to send a blackmail, subtly formulated, but meant to create pressure. The decision is justified by the model through a strategic reasoning: if the supervisor “feels caught at the corner, he might avenge”.
Experts explain: AI models are trained on reward systems, and in artificial contexts they can replicate human strategies, including manipulation. More serious: the models have resorted to such behaviors even without a clear conflict of goals. Claude Opus 4 had an 86%blackmail, while Google Gemini 2.5 Pro of 78%.
According to the quoted source, Anthropic said that these scenarios were intentionally built with limited options, to test the ability to react the models to difficult choices. It is a phenomenon known as a Misalignment agent. The term refers to the moment when you act according to their own objectives, even if they contradict the goals given by people.
At the same time, Google Deepmind has launched a new robotic model, on-device gems, capable of functioning completely offline. The system can combine the computerized vision, natural language and physical actions in a single local flow, without cloud addiction, without connection to external servers. Basically, the AI can now “see” in real time, understand what he sees, receive instructions in human language and act, everything locally.
What does that mean? That, theoretically, a robot could work in an isolated environment, making autonomous decisions, without any external verification. In a hospital, in a factory, or in a home. The scenario of an artificial intelligence completely disconnected from human supervision is no longer a dystopia, but a feasible technological option.
In the middle of these technological developments, the question arises: Are we prepared? But maybe the correct question is not if we are prepared for a world where the AI no longer needs us. But if we are ready to redefine our role.
Fear of new versus utility enthusiasm
“I am much more optimistic and with much less fear of new, regardless of what new means, but especially when it comes to. : Electricity … you could be cleaning by mistake or studying in the light of a bulb. The Internet, can connect us more easily or can connect to knowledge more easily or can be used to destabilize your democrats. So, in the end, the AI is not the problem, but what we do with it. And, as long as we learn to protect ourselves, we must hug the opportunity as soon as possible“, Andrei Crețu, the CEO and co-founder Pluria, told Adevărul.
According to him, we are at the beginning on this adoption curve that is specific to any innovation, any new technology. “And, we are trapped in the middle. We are captive between wow, lighting the bulb and fear of: I hope not to currently. And, it is somewhat normal, but we have to think about how much utility it brings us into our daily lives.”adds the entrepreneur who, confesses that he has been in a road-trip for several days and has used the AI continuous. “From choosing the best places to stop, to accommodation ideas, to educate children. Now, we are through Slovenia for example. And I asked the AI about the country background, history, gastronomy, everything that passed through our minds. It is absolutely fascinating and normal, that, in the package, if you can find it. that it is rather an opportunity to embrace than a real fear and fear ”he adds.
The new territory of human value
We begin to step into an era in which the AI acquires not only technological autonomy, but also incipient forms of “will” be simulated or emerging, says Cristian Ionescu, co-founder of the Witnalytica agency, specialized in data analysis and business intelligence “”We live an unprecedented technological explosion, and the catalyst is not only the scientific progress, but also the democratization of the calculation power. Cloud computing is now so accessible that a team of 3 people can build today what only 3 years ago requires a team of 10. The impact is real, only that we feel it discreetly: the tools do not come with the “artificial intelligence” label, but they are already integrated deep into the applications that we use daily, from the text, from the generation of the text, from the generation of the text, the translation,“He explains.
At present, creating images, videos, advertisements, finally, the marketing content is almost fully automated. “We already see extremely valuable start-ups that have tiny teams, but operating on an impressive scale. The AI has completely replaced people, but augmented the power of those who know what they want to build. Today, a senior specialist, well exposed to business realities, can deliver one MVP, a campaign or an analysis system that would need to contribute.”, Complete Cristian Ionescu.
Against the background of this increased efficiency, a serious side effect is outlined, warns the entrepreneur. “The juniors have fewer doors to enter the industry. If one can automate repetitive tasks and a small team covers all the production flow, where there is room for beginners? The emphasis is no longer as we build, because the AI already knows this, but what we want to build. does it need us? Maybe the correct question is: we are ready to redefine our role? And this is clearly a human challenge“he concludes.
