The question “Will AI steal your job?” has become a kind of shorthand for a much larger anxiety: that work—stable, familiar, and tied to identity—might be replaced by systems that don’t care who gets hurt. But as Sarah O’Connor argues in a recent Financial Times piece, the debate is often framed too narrowly. The real issue isn’t only what artificial intelligence can do. It’s what it ought to do—especially when the technology’s capabilities collide with livelihoods, bargaining power, and the social contract around employment.
That shift in emphasis matters because it changes the conversation from prediction to design. If the discussion stays at “what AI can do,” then job displacement becomes an engineering problem with a foregone conclusion: automation will happen, and workers will have to adapt. But if the discussion moves to “what AI ought to do,” then automation becomes a governance problem. It becomes about choices—who benefits, who bears risk, how transitions are managed, and what obligations employers and governments should accept when they deploy systems that reshape labor markets.
O’Connor’s framing lands at a moment when AI is no longer confined to research labs or niche tools. It is being integrated into customer service, marketing, software development, finance operations, logistics planning, legal review, and a growing range of administrative tasks. In many sectors, the first visible change is not the disappearance of entire occupations overnight. It’s the quiet reallocation of tasks: the parts of jobs that are easiest to standardize get automated first, while the remaining work becomes more complex, more monitored, or more dependent on human oversight. That incremental transformation can still be destabilizing, even if it doesn’t look like mass layoffs in the early stages.
So what does it mean to ask what AI ought to do? It means treating AI deployment as something closer to infrastructure than consumer gadgetry. Infrastructure doesn’t just “happen.” It is planned, regulated, maintained, and—crucially—built with assumptions about safety, access, and fairness. When AI is used to make decisions that affect pay, hiring, performance evaluation, or service outcomes, it should be held to similar standards. The question becomes: what protections should be built into the system before it reaches the workplace?
One of the most important implications of O’Connor’s argument is that productivity gains are not automatically fair outcomes. AI can increase output per worker, reduce errors, and speed up processes. Those improvements can be real and measurable. But the distribution of those gains is a separate decision. Without guardrails, the benefits tend to accrue to those who own the technology and control the deployment timeline, while the costs—job redesign, wage pressure, retraining burdens, and uncertainty—fall on workers and communities.
This is where “ought to do” becomes more than moral language. It becomes a practical set of expectations about how organizations should behave. If AI is going to change work, then the change should be structured so that workers are not simply asked to absorb disruption. Instead, workers should be treated as stakeholders in the transition.
That idea sounds straightforward, but it collides with how many companies currently approach automation. In many workplaces, AI adoption is treated as a competitive necessity: implement quickly, measure performance, and iterate. Consultation with employees may be limited, and the timeline for workforce impact may be unclear. Even when companies intend to be responsible, the incentives can push them toward minimizing short-term costs rather than maximizing long-term stability.
O’Connor’s perspective challenges that incentive structure. It suggests that the future of work should not be treated as inevitable. It can be shaped—by policy, by corporate commitments, and by collective bargaining. In other words, the question is not whether AI will transform employment; it’s whether the transformation will be designed to protect people.
Consider the difference between two scenarios that both involve AI replacing certain tasks. In one scenario, a company introduces AI tools and then expects employees to “learn the new system” without time, support, or guarantees. In the other scenario, the company builds a transition plan: it identifies which tasks will change, provides training during paid hours, offers pathways to new roles, and sets expectations about how performance metrics will be adjusted. The technology might be similar in both cases. The outcomes for workers could be radically different.
The “ought to do” framing also highlights a common misconception: that AI’s impact is best understood at the level of whole jobs. In reality, jobs are bundles of tasks. Automation tends to target specific components—drafting, summarizing, routing, classification, scheduling, data entry, and other activities that can be standardized. That means workers may not lose their titles immediately, but they can lose autonomy, bargaining power, and the ability to influence how work is done. A role can be “kept” while its content is hollowed out.
This is why the debate about AI and employment should include questions about task redesign and worker agency. If AI is introduced, workers should have a say in how it changes their day-to-day responsibilities. They should also have clarity about what the system is doing, what it is not doing, and what happens when it fails. In high-stakes environments—healthcare, finance, safety-critical operations—“human in the loop” cannot be a slogan. It must be operationalized with training, authority, and accountability.
Accountability is another area where “what AI ought to do” becomes concrete. Many AI systems are difficult to explain, and their outputs can be inconsistent. When these systems are used in employment contexts—such as screening candidates, predicting performance, or monitoring productivity—errors can translate into real harm. A biased model can disadvantage certain groups. A flawed metric can punish workers for circumstances outside their control. A poorly calibrated system can create a false sense of objectivity.
If AI ought to do something, it should do it reliably and transparently enough that affected people can contest outcomes. That implies documentation of model behavior, auditing for bias, clear escalation paths for disputes, and limits on how automated judgments can be used. It also implies that organizations should not hide behind technical complexity when workers ask basic questions: Why was I rated this way? What data was used? What can I do to correct it?
The “future of work” conversation often becomes polarized between two extremes: techno-utopianism (“AI will free humans from drudgery”) and techno-doom (“AI will eliminate jobs”). O’Connor’s framing offers a third path: treat AI as a tool whose social consequences depend on governance. That means acknowledging both the potential benefits and the risks without pretending either side is guaranteed.
There are genuine reasons to believe AI can improve work. It can reduce repetitive burden, help workers focus on higher-value tasks, and provide assistance that makes complex work more accessible. For example, AI can draft first versions of documents, summarize long reports, translate materials, and help with coding or analysis. In skilled professions, these tools can accelerate learning and reduce time spent on low-level tasks.
But the same tools can also intensify work. If AI speeds up production, employers may expect faster turnaround without reducing workload. If AI improves monitoring, managers may increase scrutiny. If AI reduces the cost of producing content, organizations may raise output targets. In that environment, workers can experience “productivity” as pressure rather than relief.
This is why the question “what AI ought to do” should include labor standards. It should include expectations about workload, evaluation, and the right to disconnect from constant performance measurement. It should include rules about how AI-generated outputs are credited and how responsibility is assigned when mistakes occur. It should include protections against surveillance creep—where AI becomes a mechanism for continuous tracking rather than a tool for assistance.
Another practical dimension of O’Connor’s argument is the need for skills training and job transition plans, but with a crucial caveat: training alone is not a solution if the labor market is shrinking or if workers are being displaced faster than they can move. Skills programs can help, but they must be paired with demand-side commitments—new roles, new hiring, and pathways that actually exist.
A responsible approach to AI deployment would therefore treat training as part of a broader labor strategy. That includes identifying which roles are likely to change, partnering with educational institutions, offering certifications that employers recognize, and ensuring that workers can access training without losing income. It also includes designing internal mobility: if AI changes tasks within a company, the company should create routes for workers to move into new functions rather than simply replacing them.
In some industries, internal mobility is already happening, but it is uneven. Workers with less bargaining power often face the hardest transitions. They may be offered training that is too generic, too late, or too disconnected from actual job openings. A “responsible” AI rollout would be judged by whether workers can realistically move into stable employment, not by whether training sessions were offered.
Policy has a role here too. Governments can shape the incentives for responsible deployment through procurement rules, labor regulations, and requirements for transparency. They can also support transitions through unemployment insurance reforms, wage insurance, and funding for retraining. But policy alone cannot substitute for workplace-level commitments. The workplace is where AI is deployed, where decisions are made about staffing, and where workers experience the consequences first.
That is why the “ought to do” framing should extend beyond governments to employers and investors. Employers should be expected to conduct impact assessments before deploying AI in ways that affect employment. These assessments should consider not only technical performance but also labor effects: which tasks will change, how many roles might be affected, what timelines are involved, and what mitigation measures will be implemented. Investors, meanwhile, can influence corporate behavior by rewarding responsible practices and penalizing reckless automation strategies that externalize costs onto workers.
There is also a cultural dimension. Many workplaces treat AI adoption as a top-down modernization project. But if the future of work is something society can shape, then workers need a voice. That could mean stronger consultation mechanisms, union involvement, and employee representation in decisions about AI tools. It could mean co-design approaches where workers help define how AI should be