Humans in the Loop
Wrapping up our generative AI working group
To conclude MIT’s working group on generative AI and the work of the future, we’re releasing a paper summarizing our findings from three years of research on companies’ experiments with generative AI.
Here is the executive summary and here is the full paper (with abstract below).
How can generative AI lead to better jobs? In January 2026, approximately half of American workers reported using AI, but how new technologies have affected the quality of their jobs remains largely unclear. Drawing on a study of more than twenty companies across four major industry groups – healthcare and life sciences; retail; finance, insurance, and real estate; and manufacturing – this paper identifies patterns in how organizations have experimented with generative AI; how those experiments have changed the roles of workers; and how organizations can support high-quality jobs as they integrate new technologies.
The applications of generative AI among the companies we studied were directed toward three common challenges. The bottleneck problem is where workers are responsible for a growing number of simple tasks that get in the way of higher value-added work. Generative AI tools have been aimed at relieving these bottlenecks by speeding up near-routine tasks. The cafeteria problem emerges in a process that requires workers to consult experts from various domains and integrate their input into a product, document, or idea. Organizations have looked to generative AI to predict what those domain experts might have said based on what they have produced in the past. The learning curve problem represents the extra time and effort novices require to complete a complex task in a new domain. Generative AI tools have been directed to help workers to perform as if they had more experience – and to develop expertise – in new domains.
Across the applications of generative AI addressing these challenges, there has been a shift in the core tasks that professional and technical workers are being asked to perform. Where generative AI tools are being deployed, workers are increasingly asked to perform supervisory control tasks as the “human in the loop” overseeing and analyzing a process rather than executing the process manually. Although supervisory control tasks may be new to workers in law or healthcare, the “human in the loop” concept is not new. A range of occupations from airline pilots and manufacturing technicians to utility operators are supervisors of automated systems, and there are guidelines for how workers in these roles can thrive that can inform the use of generative AI.
We draw six lessons for how to capture the benefits of these technologies for productivity and knowledge creation without reducing the quality of work or the level of skill among workers who come to rely on AI.
Minimize drudgery. Although some workers in surveys report enjoying their routine tasks, there are many workers who would prefer problem solving and creative tasks to the mundane and routine. Some organizations have ventured into applications of generative AI that aim to augment tasks that workers enjoy – and from which they derive value – setting up resistance from the workers affected. The more straightforward applications of generative AI can focus on the bottlenecks that both workers and employers can agree would be better if they were gone.
Promote learning. There is early evidence that using generative AI can lead to forms of mental offloading, where a worker performs a task using generative AI, but does not retain the knowledge of the task that they performed. Learning at work can unlock better career opportunities for employees, and deliver higher productivity for employers. There is evidence that generative AI technologies can help employees learn by pinpointing where there are gaps in their knowledge and providing relevant information. However, ensuring that generative AI tools are used for learning and not for offloading will require guardrails that steer workers in the former direction.
Preserve teamwork. Generative AI applications can decrease individuals’ reliance on others for expert advice. Even in the cases where generative AI tools can enable an individual to do the work of a team, it may be important to preserve teamwork for other reasons, including mentorship, collective learning, and trust building among people who might rely on one another for other tasks.
Interface design: As employers have grown to understand the capabilities of generative AI tools, they have tended to buy tools from vendors rather than build their own custom applications. Even when organizations use tools built for their industry, they have a choice in how their workers engage with generative AI: they can customize the interfaces that their workers use. Good interface design can ensure that generative AI promotes learning and attention rather than skill atrophy. The principles for good interface design have sought to maximize a user’s situational awareness – their ability to understand what is happening and why – as well as to manage their mental workload.
Domain expertise: There is some indication that generative AI tools can help users extend their capabilities. However, this should not be mistaken as a replacement for domain expertise. The users of AI tools must still be able to provide the underlying context, interpret their results, and improve them as they improve their understanding of the background domain. In short, domain expertise can still complement generative AI technologies, but the expertise itself must be sufficient to validate and scrutinize what generative AI tools provide.
Accountability: Generative AI tools can create a moral hazard for workers using them. A worker may have an incentive to use generative AI tools as a shortcut to produce information and perform work that appears valid and impressive, even if that work contains errors or masks a lack of underlying understanding or capability. One way for organizations to address the moral hazard challenge is to establish accountability for workers using AI tools to increase their incentives for learning and raise their costs of making an error. Employers can introduce random errors or tests of their employees to audit performance and reward workers for their vigilance.
If you’d like to follow up or stay in touch, please write to workofthefuture@mit.edu.
Strong piece. But “human in the loop” is often a transitional phrase, not a destination. Once models improve, many firms will quietly move from humans in the loop to humans on the loop - and eventually humans blamed by the loop.
The real task is not inserting people at checkpoints. It is designing organisations where human judgment actually retains authority, skill, and learning value. Otherwise oversight becomes theatre. 
Human-in-the-loop isn’t a feature request—it’s the governance layer that makes deployment safe, reliable, and socially acceptable. Great synthesis of real experiments.