When AI won’t stay still, law and work struggle

AI won’t – From policing software that changed its name after bans to workplace systems that hollow out the meaning of jobs, the problem isn’t just what AI can do—it’s that nobody can even pin down what AI is. That fluid definition spills into legal risk, corporate behav

On a single day, a city can draw a line around “predictive policing” software—and then watch the company step back, rebrand, and claim it was never offering the thing the city tried to stop.

Paola Lopez describes that kind of scramble in Merkur (Germany): a chatbot. a smart fridge. a predictor of payment-default risk. an automated translator. a self-driving car. an email spam filter. and an earthquake predictor are all placed under the same umbrella term. “AI.” The category resists stable definition. The systems themselves shift quickly as newer models supersede older ones. For lawmakers, that moving target becomes more than a philosophical headache—it creates concrete legal problems. If “AI” covers a broad and shifting range. from generative models to more limited algorithmic tools. how do you regulate what you can’t clearly name?.

Lopez frames the dilemma like a trap with two exits that both fail. Regulators risk being too narrow, missing the emerging systems that keep arriving. Or they go too broad and lump together fundamentally different technologies. As the debate intensifies, she adds, firms are deliberately downplaying their use of AI.

PredPol illustrates the tension. It had been a market-leader in predictive policing until the city of Santa Cruz—where it is based—banned the use of such technologies. The company responded by changing its name to Geolitica and claiming that it had never offered predictive tools in the first place. Lopez points to the cycle that can follow: at first. everyone wants to “use AI” and everything “has AI. ” because it’s easy to ride the AI hype wave. Then, once regulation kicks in, “nobody will want to have been associated with AI.”.

That urgency around definition lands differently once AI enters the workplace—not only by automating tasks, but by shifting what work itself is worth. Lisa Herzog distinguishes four dimensions of work value being undermined by AI: didactic, community-building, meaning-creating, and political.

In the didactic dimension, work is supposed to be a space where people acquire and refine skills. Herzog says that when AI automates complex tasks. it can limit opportunities for learning through practice. making expert knowledge harder to gain. That can also hit motivation: it was “precisely the opportunity to practise and develop certain skills” that attracted someone to a profession in the first place.

Work is also social. It integrates people who might never meet otherwise, building shared culture. But algorithmic management isolates workers, Herzog argues, making it harder to develop a sense of community.

Then comes meaning. Herzog describes human action as “structurally polysemic. ” where even smaller. more tedious tasks can feel worthwhile when you know they contribute to a broader goal. On AI-managed platforms. the connection is stripped away: numerous mini jobs are outsourced to temporary workers who lack knowledge of the ultimate purpose. Labour can start to feel like navigating “an obstacle course of tiny. intricate hurdles” rather than achieving something of real value.

And finally. workplaces can be “important loci of politicization.” Herzog ties this to the ability to talk about working conditions and workers’ rights—conversation that fosters political consciousness and action. By restructuring labour into individualized tasks and pitting zero-hours workers against each other in a race to pick up jobs. AI negates that collective dimension.

The thread running through these descriptions is the same one Lopez begins with: AI’s shape keeps changing. but the consequences do not pause. A category that refuses a fixed definition can be stretched wide in public debate and narrow in legal response; a technology that arrives in work systems can remake the value of jobs even as it changes what it’s called.

The article turns next to how that ambiguity is engineered into everyday experience. Max Beck writes about “engineered anthropomorphism,” arguing that large language models are increasingly human-like. They produce responses that feel conversational, empathetic, and self-aware. Beck insists the quality isn’t incidental: design choices build it in. Even the decision to present interactions in the form of a ‘chat’—rather than node-based workflows or command-line tools—is deliberate. The display of generated tokens as flowing text, reminiscent of human typing, also matters.

Beck traces how LLMs are made. A base model is trained on vast text corpora to generate statistically plausible language. and at that stage the “form of the response” is determined purely by probability theory from the training data. which “is not always conversational.” Fine-tuning adapts the model to more specific tasks and improves relevance and fluency.

Then comes ‘Reinforcement Learning from Human Feedback’ (RLHF). where human evaluators rank and compare outputs. rewarding those that appear helpful. polite. or friendly. Beck says this process embeds human communicative norms into responses. giving the model its “personality. ” and producing a style that often mimics emotional awareness.

That anthropomorphic style, Beck argues, has financial advantages and isn’t going away. Human-like systems are easier and more pleasant to use, increasing “stickiness” and prolonging interaction time. Use-time, he writes, is “the currency of all interactive platforms.”

So the question that starts as a legal one—what counts as AI. when it changes shape. and how to regulate something that keeps moving—ends up as a lived one. In the workplace Herzog describes, where learning, community, meaning, and political agency are undermined, the stakes aren’t abstract. And in the marketplace Lopez reports—where firms can rebrand after bans and claim they were never doing what they were said to be doing—the category’s instability becomes a shield.

Between law, labour, and interface design, the “new infinity” that Birger P. Priddat describes starts to feel less like a slogan and more like a warning. Priddat argues that economics has long multiplied possibilities of access. expanding into new “field regimes.” In the twenty-first century. with physical resources exhausted. the field expands internally into human behaviour itself: Google. Meta. and others define private data as “raw” and “ownerless” until processed by their algorithms. And what comes next. Priddat suggests. is the biological—remaking the world through AI-assisted solutions such as solar geoengineering. heat-resistant corals. plastic-eating bacteria. and lab-grown proteins.

But the line that lingers is his description of direction: the transition from the unconscious destruction of planetary systems to their conscious composition of life. If that is the horizon, then the immediate task is painfully concrete. People are already being categorized. measured. reorganized—and the name of what’s doing it keeps shifting under everyone’s feet.

artificial intelligence AI regulation predictive policing Santa Cruz PredPol Geolitica workplace automation meaning of work algorithmic management LLMs RLHF anthropomorphism cultural identity technology and society