At the start of the call, Springboards' CEO said something that alone sums up the moment: 'We often say we're a self-loathing AI company. We don't know if we really like what we're doing.' The quip, echoed by the idea of an equally self-critical AI journalist, is not just startup posturing. It is the symptom of a tension that data now measure precisely.

The numbers behind the paradox are stark. According to the Pew Research Center, more US adults expect AI to have a negative impact on them personally and on society than expect a positive one, with pessimism strongest among the young. More than half of people worldwide say AI products and services make them nervous, according to a Stanford University report. In a May Gallup poll, 71% of US adults said they would oppose construction of a new AI data center in their area, far more than the 53% who would oppose a nuclear power plant. And in an NBC poll in March, AI was less popular than ICE.

Yet usage keeps growing. ChatGPT hit a billion monthly users in May, according to Sensor Tower, while Google DeepMind's Gemini follows closely with 950 million users in July. Half of US adults now use a chatbot, more than double the share in 2023, and one in four do so every day. Across OECD countries, more than a third of adults have used generative AI tools in the last three months. The picture is of two conflicting truths: many people cannot stand AI and, at the same time, cannot get enough of it.

The most plausible explanation is not that there are two distinct groups. If that were the case, the Venn diagram would already be almost a circle. Rather, the hostility is not aimed at the technology itself, but at the relentless drive of companies to push it into every possible domain. Those promising the biggest social and economic upheaval in generations cannot be surprised if part of the public reacts with distrust. The social media parallel is useful but imperfect: there, leaving a platform meant losing content and connections, leaving only the choice between adapting or starting over. With AI there is still room to maneuver, because regulation and open-source alternatives are reshaping the market before lock-in becomes definitive.

Opposition to data centers, in particular, is an infrastructural signal more than a technological rejection. It is not a protest against the electricity needed to run models; it is a demand not to bear externalities without a voice. This shifts attention toward more distributed architectures, where inference can happen closer to data and with less dependence on a few cloud providers. For those evaluating on-premise or self-hosted deployments, the signal is twofold: regulatory pressure can make building new centralized data centers more expensive and slower, while the availability of open-source models makes it possible to test alternatives without being locked to a single vendor.

I don't want to be naive: trillion-dollar companies are hard to move. But what comes next is not as inevitable as they often imply. When I asked Springboards' CEO why develop a new model if he doesn't really like AI, he replied that there is no walking back from LLMs, but you can still make them do something different. It is the most honest closing: AI that clearly says what it can and cannot do, without posturing as a technology about to take over the world. If that becomes the new standard, the survey paradox might resolve itself.