When technology changes faster than people’s lives can adjust, the question stops being whether jobs will be affected and becomes how society responds to the disruption. That is the thrust of a call from Gus O’Donnell, the former UK cabinet secretary, who argues that workers labelled “losers” in the transition to AI should not be left to absorb the shock alone. Instead, he says, governments and employers must fund large-scale retraining so displaced workers have a realistic route into new work—an approach he frames not as charity, but as economic policy designed to manage instability before it becomes entrenched disadvantage.
O’Donnell’s position lands at a moment when public debate about AI often swings between two extremes: either the technology will create enough new roles to offset losses automatically, or job displacement is being exaggerated and therefore requires little intervention. His argument sits in the middle, insisting that even if net employment effects are uncertain, the distributional impact is immediate and uneven. Some workers will lose tasks, then roles; others will see their work redesigned around AI tools; and many will need help moving from one set of skills to another. The key, he suggests, is to treat retraining as infrastructure—something that must be planned, financed, and delivered at scale, rather than offered as an optional add-on after layoffs begin.
The core of his proposal is straightforward: compensation for those harmed by automation should take the form of targeted investment in training and reskilling. But the implications are anything but simple. Retraining at the scale implied by AI-driven change requires decisions about who pays, what gets taught, how quickly programmes can be rolled out, and how to ensure that training leads to actual employment rather than simply producing certificates.
A shift from “job loss” to “task loss”
One reason the retraining debate has become so contentious is that AI does not always eliminate entire occupations overnight. More often, it changes tasks within jobs. A customer service role may be partially automated through AI chat systems; a legal assistant’s workload may be reshaped by document analysis tools; a marketing team may rely on AI-generated copy and campaign variants. In these scenarios, workers may not be dismissed immediately, but their bargaining power declines as fewer human hours are required for the same output. Over time, that can still lead to redundancies—especially for people whose experience is concentrated in tasks that become cheaper to perform with software.
O’Donnell’s framing implicitly acknowledges this reality. If the problem is task displacement, then the solution cannot be limited to training people for entirely new careers in a single leap. It must also support transitions within industries: helping workers learn how to use AI tools, how to supervise or validate outputs, and how to shift toward higher-value activities that remain difficult to automate. That means retraining programmes should be designed around the evolving structure of work, not around static job descriptions.
The “compensation” idea: why retraining is more than a moral argument
Calling retraining “compensation” is a deliberate rhetorical move. It reframes the issue from a debate about whether governments should be generous to a debate about fairness and responsibility. If public policy encourages or enables technological adoption—through regulation, procurement, tax incentives, or infrastructure—then the costs of adjustment cannot be treated as purely private. Workers who bear the risk of transition should receive support that reduces the likelihood of long-term harm.
There is also a pragmatic economic logic. When displaced workers struggle to find stable employment, the costs show up elsewhere: higher unemployment benefits, lower tax receipts, increased demand for social services, and—perhaps most damaging—skills erosion. People who remain out of work for extended periods often lose confidence, professional networks, and the ability to compete effectively. Retraining that arrives late can become less effective precisely because the window for re-entry narrows.
O’Donnell’s emphasis on funding suggests he believes the UK—and other countries facing similar pressures—must treat retraining as a time-sensitive intervention. The longer the delay between displacement and training, the more expensive and less successful the response becomes.
Why “at scale” is the real challenge
The phrase “retraining at scale” sounds like a slogan until you confront the operational details. Training systems are not built to respond quickly to labour market shocks. Universities and colleges have academic calendars; providers have capacity constraints; employers may hesitate to commit to hiring trainees without clear signals about job readiness; and individuals may struggle to attend courses while managing income loss.
To make retraining work, several bottlenecks must be addressed simultaneously.
First, there must be a pipeline of training places that can expand rapidly. That could involve partnerships with private training providers, community colleges, and employer-led academies. But expansion is not just about numbers; it is about quality. If programmes are rushed, they risk becoming generic and disconnected from the skills employers actually need.
Second, training must be aligned with demand. AI adoption is creating new roles—often in areas like data governance, model evaluation, cybersecurity, and AI operations—but the demand for these roles is not infinite, and it varies by sector and region. A retraining strategy that ignores local labour market realities can produce graduates who are qualified on paper but unable to secure interviews.
Third, training must be accessible to people who are already under pressure. Many displaced workers cannot afford to take months off work without income support. That is where funding becomes more than tuition. It may need to include travel support, childcare, digital equipment, and—crucially—income replacement during training periods. Without these supports, the people most likely to need retraining may be least able to participate.
Fourth, there must be a credible pathway from training to employment. Certificates alone rarely guarantee outcomes. Programmes need employer involvement: internships, apprenticeships, guaranteed interview schemes, or direct recruitment pipelines. The goal is to reduce the gap between learning and earning.
A unique take: retraining as a labour market stabiliser
O’Donnell’s argument can be read as a broader critique of how societies handle technological transitions. Too often, governments treat job disruption as a temporary inconvenience and focus on macroeconomic stability—keeping inflation low, maintaining growth, and hoping that labour markets will self-correct. But AI-driven change may not behave like a typical recession. It can be gradual in some sectors and sudden in others, and it can alter the skill requirements of entire industries without waiting for economic cycles to turn.
In that context, retraining becomes a stabiliser. It can prevent the formation of a long-term underclass of workers whose skills no longer match the economy’s needs. It can also reduce political backlash by demonstrating that the transition is managed rather than imposed.
This is where O’Donnell’s call feels particularly relevant: he is not arguing against innovation. He is arguing for a social contract that makes innovation politically sustainable and economically efficient.
What should retraining actually teach?
If retraining is the answer, the next question is what the curriculum should look like. AI is not a single skill; it is a set of capabilities that can be embedded across many workflows. That means training should be modular and role-specific.
For many workers, the most immediate value may come from learning how to work alongside AI systems rather than competing with them. That includes understanding how AI outputs are generated, how to verify accuracy, how to manage risk, and how to integrate AI tools into existing processes. In practical terms, that might mean training in prompt design and workflow automation for certain roles, or training in data literacy and quality assurance for others.
But there is also a need for deeper technical and analytical skills for those moving into new functions. AI adoption increases demand for people who can interpret data, manage systems, and ensure compliance. Yet not everyone displaced will want—or be able—to move into highly technical careers. A well-designed retraining strategy should therefore offer multiple pathways:
1) Upskilling within the same occupation
2) Transitioning to adjacent roles within the same industry
3) Moving into entirely new occupations where demand is strong
4) Supporting entrepreneurship or freelance work where appropriate, especially for people with transferable experience
The danger is designing a one-size-fits-all programme. O’Donnell’s emphasis on targeted investment implies he would favour a system that can tailor training to different starting points and different labour market destinations.
The role of employers: from passive beneficiaries to active partners
Employers are often described as the drivers of AI adoption, but they are also the entities best positioned to identify which skills are changing and which roles are emerging. If retraining is to be effective, employers must do more than provide occasional lectures or allow trainees to sit in on meetings.
They need to participate in curriculum design, offer work placements, and commit to hiring outcomes where possible. This is not only a matter of corporate goodwill; it is a matter of reducing recruitment friction. Employers face uncertainty about the reliability of newly trained candidates. If they help shape training, they can reduce that uncertainty.
At the same time, employers may resist retraining obligations if they fear it will become a cost without benefit. That is where government funding and policy design become crucial. If public money is used, it should incentivise employer engagement and tie funding to measurable outcomes such as job placement rates, retention, and wage progression.
Funding mechanisms: what “more funding” could mean in practice
O’Donnell’s call for additional funding raises the question of how that money would be deployed. There are several plausible models, and each has trade-offs.
One approach is to expand existing adult education budgets and create rapid-response training funds that can be activated when specific sectors face displacement. Another is to establish a national retraining programme with regional delivery partners, ensuring that training capacity can be scaled quickly.
A third approach is to create individual training entitlements for workers at risk of displacement, combined with employer co-investment. This can empower workers to choose programmes that fit their circumstances, but it requires strong oversight to prevent low-quality providers from capturing funds.
Whatever the mechanism, the funding must be designed to avoid common failure modes: bureaucratic delays, training that is too theoretical, and programmes that do not connect to real hiring demand.
The political dimension: why this