The high road, the low road – which approach to AI at work?

Those on the 'high road' anticipate employment gains and market share growth from AI; those on the ‘low road’ are expecting job losses and higher output with less labour.

Emeritus Professor David Peetz
Laurie Carmichael Distinguished Research Fellow, Carmichael Centre
Centre for Future Work

Emeritus Professor David Peetz sets out the challenges and opportunities for organisations in the choices they make about how they introduce AI at work. He argues that engaging workers and allowing these workers a say in how and when and why AI is introduced is good not only for workers, but also for the productivity and profitability of firms and for society. David proposes a new institutional framework to regulate AI at work.

Cory Doctorow described how, to understand the way workers are affected by AI, it is useful to refer to a concept from automation theory, that of the Centaur and its counterpart, the ‘Reverse Centaur’.

A Centaur was a mythical Greek creature, half man, half horse. In automation theory, a Centaur is a worker who is ‘assisted by a machine (a human head on a strong and tireless body).1 A Reverse Centaur is a machine that uses a human being as its assistant – a frail and vulnerable person puppeteered by an uncaring, relentless machine’.2

Whether a worker is a Centaur or a Reverse Centaur has a big impact on the health and safety implications of AI for them. The impact on workers is heavily influenced by the reasons for management introducing AI, and the voice that workers have in its introduction and implementation.

High road or low road?

The logic of capitalism demands that firms always try to undercut their competitors, by introducing new technology that either improves the product or lowers the production cost or both. If they do not, they will eventually go out of business.3 But which path do firms choose? A large survey across 4 nations found that managers in 26% of firms expected AI to cause job losses in their firm, while 11% expected job gains.4 Those who anticipate employment gains from AI are likely expecting, or at least hoping, that their firm’s market share will grow, presumably through some improvement in the product. They are on a ‘high road’ path. On the ‘low road’ are those who anticipate job losses, who are expecting to be able to get a lot more done with less labour. Maybe they are not expecting product improvements at all.

These data suggest that, overall, AI is more often about cutting jobs than improving products. In the study mentioned previously, the low road had more traffic than the high road, at least in the major Anglophone countries (US and UK) – though not in Germany and Australia, where the split between the 2 roads was fairly even.

Worker voice and the high road option

For those people running one of these high road firms that think they are going to augment or create new products, and grow employment, the relevant evidence from other studies is that, overall, firms do better when they involve their employees.5 There should be no surprises there.

Workers have a vested interest in finding the most productive systems for their organisations, because it makes their jobs more secure and gets them higher pay. So it is that, in German manufacturing, where employees have an institutionalised say in the introduction of technological change through ‘works councils’, it appears that ‘exposure’ to robots is associated with an increased probability that an employee will keep their job.6

Giving workers a voice, whether it is through unions or something else in non-union workplaces, is key to making this process work properly – a topic that I will come back to shortly.

Supressing worker voice and the low road option

It is when firms are on the ‘low road’, and there are lots of Reverse Centaurs clogging the traffic, that the case for worker voice takes on a different tone.

If firms introduce AI in order just to cut down the number of workers whose wages have to be paid, those workers left may well end up working harder. They will likely end up producing more. They will certainly be under more pressure. Some of that AI might impinge on workers’ privacy or rights. It may be used to enable tighter surveillance of workers, or to enable speedup, that is to intensify work.7 Or it may be used to prevent workers from cooperating or organising. An example is AI’s relevance within documented practices to discipline workers, punish or even sack troublemakers (activists) at an Amazon fulfilment centre in England’s West Midlands.8 If workers do not have a voice at work, all this can have very bad effects for workers – effects like stress, injury, sickness – all important workplace health and safety (WHS) outcomes.

Why regulate AI?

At the macro level, there is uncertainty about the effects of AI on the number of jobs in the labour market. To the extent that prices fall (from savings due to the adoption of AI) and savings are made by consumers, those consumers will create demand in new areas, so jobs will be created elsewhere. If there are good structural adjustment policies in place, if there are mechanisms for reallocating labour to more efficient uses that take account of what people actually want and can do (again, an argument for greater worker voice, this time in structural adjustment policy) then there might still be just as many people in jobs – they just might be different jobs.

In this context, it is important not to fall for the argument that job losses attributed to AI are actually due to AI. Firms might also use AI as a mask for why they are cutting jobs (’AI-washing’). Tech companies themselves have been recently engaged in some mass layoffs, allegedly due to AI.9 But AI is a good cover for bad management. It is in the interests of the big tech firms that claims about AI’s effects stay overblown. They need the financial markets to think AI is the way of the future. They want markets to think that AI is going to lead to massive job cuts, savings and profits.10

The question of net job losses is not the main rationale for AI regulation. It is at the micro level, in what happens within people’s individual jobs, where the core problem exists – particularly the major WHS implications.

What I said above, about firms’ objectives in introducing AI, all assumes that they actually know what they are doing. What if they do not? What if they drank the Kool-Aid? Often, management does not really know what AI can do, and they overestimate its capabilities. Companies might think AI means they can sack large numbers of workers. They end up with AI chatbots giving lousy service to customers. 

Companies frequently think that AI can do so much right, but actually it gets so much wrong. ‘Microslop’11 and hallucinations12 are just the more obvious forms of the problem. With almost anything genuinely complex, the output of generative AI in particular needs to be checked and sometimes corrected by humans. Workers spend a huge amount of time doing this. If workers are working with AI, they might not be just either a Centaur or a reverse centaur. They could be a Heracles, the Greek hero who had to clean all the horse excrement out of the Augean stables.

Will firms who have downsized through AI bring back the workers they actually need? Sure, some have already reversed cutbacks they attributed to AI.13 But other companies appear to permanently try to run understaffed.14

Theory tells us that, in a less than perfectly competitive labour market, firms can maximise profits if they ’hire less labor and make do with vacancies.’15 They will lay off more staff than they ‘need’ to, and then the ‘survivors’, who have to clean up this mess, are left overworked, stressed, ill and injured. If workers do not have a say, the firm suffers and the workers suffer as well, including through health and safety effects.

What do we do about mitigating the negative impacts of AI at work?

Bear in mind, there is a lot happening at the international level in the relevant soft and hard regulation spaces.16 Bear in mind, also, that practices that get established in non-union workplaces will determine the fate of unionised workers if there is to be a ‘race to the bottom’. In addition, we cannot expect even well-organised workers to have the technical capacity to know the full impact of a particular piece of technology. Workers need to know the alternatives.

You don’t have voice if you’ve got Hobson’s choice. 

A new institutional structure to regulate AI

To start, Australia needs a strong national framework of principles that is consistent with other progressive national and cross-national frameworks. Australia almost had this in place with the proposed ‘mandatory guardrails’ that were drafted and circulated by the Federal Government.17 These were closer to the European framework than to the unfettered US approach. However, the guardrails were watered down and the mandatory character of them abandoned.18

The first step in adequate AI governance would be to reinvigorate the mandatory guardrails model.

A decent framework should provide rights for worker voice where voice can be achieved. But we also need to acknowledge that, in many workplaces, voice will not be achieved. Only a minority of workers are in unionised workplaces, and these constitute only a small minority of all workplaces. While voice mechanisms exist in some non-union workplaces, these presently exist only at the will of management.

In the absence of universal union coverage, a voice mechanism that relies on workers being able to negotiate on anything resembling an equal footing with management might be steamrolled in practice in most non-union workplaces (and some unionised workplaces). If the absence of voice leads to more reverse Centaurs and Heracles in those workplaces, then there would be a race to the bottom that would drag many unionised workplaces down with it.

A good regulatory system should include a mechanism to ensure that the high road framework is applied in all workplaces, both union and non-union.

There also needs to be a way of accessing the special expertise that genuinely understands the technical side of these issues. Even management has limited understanding of the technical issues, most workers have almost no understanding of them. 

Therefore, we need an institutional arrangement at the organisational or workplace level. One option is an organisation-level institution that is specifically focused on enhancing worker voice (perhaps commonly through unions) and thereby to redress the imbalance of power. I call these ‘Joint Digital Consultative Committees’ (JDCCs). They could be modelled on joint consultative committees that are established through many union collective agreements, or on Works Councils established under law in Germany and some other northern European countries. 

Another is a more technical/procedural organisation – or workplace-level institution that can ensure the national principles are applied at that level, for organisations above a threshold size. They could be called ‘Independent Digital Ethics Committees’ (IDECs) and they would include experts from outside the organisation. 

Perhaps both are required, or perhaps instead some organisational/workplace institution can combine the key characteristics of both these suggestions. 

We also need an intermediate institution, one that sits between the tech firms and their clients. The best proposal so far is for audits of AI software sold to firms, to check consistency with those principles. This is a proposal that originated with Deb Raji from the AI Now Institute.19 They could be termed ‘Independent Digital Audit Procedures’ (IDAPs). This is another way of bringing in the technical expertise that workers cannot be expected to have. This is also an extension of the idea of using (government) procurement procedures to vet and evaluate software’s effects, as discussed by another contributor to the symposium, Jake Goldenfein20, but takes the idea further by applying it to all AI procurement, not just procurement by government.

Due to the sheer scale of AI sales, audits can only cover a sample of software. So they could only supplement, not replace, adequate organisational or workplace institutions.

Finally, we need to have effective enforcement mechanisms. This is not only a reference to a specialised body that enforces the above features (though other bodies such as unions would also need to be able to take breaches to courts). It also requires the availability of high penalties that, for large firms, would be substantial enough to affect their behaviour.

In Australian industrial relations, rights are established under different levels (legislation, awards and enterprise agreements) and enforced by institutions with the capacity to impose penalties. The model proposed here, similarly, would operate through a range of levels and institutions.

This article is part of a suite of papers written following the symposium ‘Which Way Forward? AI and Decent Work’. Read more here.

  1. Cory Doctorow, Reverse centaurs are the answer to the AI paradox, Medium, 12 September 2025, https://doctorow.medium.com/https-pluralistic-net-2025-09-11-vulgar-thatcherism-there-is-an-alternative-f1428b42a8fd ↩︎
  2. ibid. ↩︎
  3. e.g. Karl Marx, Capital: A Critique of Political Economy – Volume 1. Chicago: C.H. Kerr & Company 1906, esp Ch XV ↩︎
  4. Ivan Yotzov et al. Firm Data on AI, National Bureau of Economic Research, NBER Working Paper 34836, Cambridge MA, February 2026. https://www.nber.org/papers/w34836 ↩︎
  5. e.g. George Strauss, ‘Workers’ Participation in Management’. In Employment Relations: The Psychology of Influence and Control at Work, ed. Jean F. Hartley and Geoffrey M. Stevenson (Cambridge, MA: Blackwell, 1992), 291-311; Department of Employment and Industrial Relations, Industrial Democracy and Employee Participation: A Policy Discussion Paper (Canberra: Working Environment Branch, DEIR and AGPS, 1986); John R. Cable and Felix R. Fitzroy, ‘Co-operation and Productivity: Some Evidence from West German Experience’. Journal of Economic Analysis and Workers Management 14, no. 2 (1980): 163-80. ↩︎
  6. Wolfgang Dauth, Sebastian Findeisen, Jens Südekum, and Nicole Woessner, The Rise of Robots in the German Labour Market. Vox, 19 September 2017, https://cepr.org/voxeu/columns/rise-robots-german-labour-market ↩︎
  7. e.g. Annette Kamp, Sidsel Lond Grosen & Agnete Meldgaard Hansen, AI and Data-intensive Surveillance in Professional Work: Transforming Discretion and Accountability, in Tereza Østbø Kuldova, Inger Marie Hagen & Anthony Lloyd, Digital Technology, Algorithmic Governance and Workplace Democracy, Springer, 2025. ↩︎
  8. James Muldoon, Mark Graham and Callum Cant, Feeding the Machine: The Hidden Human Labor Powering AI, Bloomsbury, London. ↩︎
  9. Tim Sandle, New types of downsizing: AI layoffs in the tech industry surge past 39,000, Digital Journal, 27 April 2026, https://www.digitaljournal.com/business/new-types-of-downsizing-ai-layoffs-in-the-tech-industry-surge-past-39000/article ↩︎
  10. Caroline Castrillon, The Dirty Secret Behind AI Layoffs, According to Forrester, Forbes, 27 April 2026, https://www.forbes.com/sites/carolinecastrillon/2026/04/27/the-dirty-secret-behind-ai-layoffs-according-to-forrester/ ↩︎
  11. Michael Crider, Microsoft says stop calling it Microslop, or you’re banned, PC World, 2 March 2026, https://www.pcworld.com/article/3075135/microsoft-says-stop-calling-it-microslop-or-youre-banned.html ↩︎
  12. Gyana Swain, OpenAI admits AI hallucinations are mathematically inevitable, not just engineering flaws, Computerworld, 18 September 2025, https://www.computerworld.com/article/4059383/openai-admits-ai-hallucinations-are-mathematically-inevitable-not-just-engineering-flaws.html ↩︎
  13. David Marin-Guzman and James Eyers, CBA U-turns on AI job cuts, calls back humans, Australian Financial Review, 21 April 2025. ↩︎
  14. Cory Doctorow, Understaffing as a form of enshittification, Pluralistic, 23 March 2026, https://pluralistic.net/2026/03/22/nobodys-home/ ↩︎
  15. Alan B Krueger, The Rigged Labor Market, Milken Institute Review, 28 April 2017, https://www.milkenreview.org/articles/the-rigged-labor-market?IssueID=23 ↩︎
  16. European Union, Regulation (EU) 2016/679 of the European Parliament and of the Council
    of 27 April 2016, on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data Protection Regulation)  https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=CELEX:32016R0679 ↩︎
  17. Department of Industry, Science and Resources, Safe and responsible AI in Australia, Proposals paper for introducing mandatory guardrails for AI in high-risk settings, Canberra, September 2024, https://consult.industry.gov.au/ai-mandatory-guardrails ↩︎
  18. Jake Evans, Artificial intelligence to be managed through existing laws under National AI Plan, ABC News, 2 December 2025, https://www.abc.net.au/news/2025-12-02/national-artificial-intelligence-plan-growth-existing-laws/106086474 ↩︎
  19. Katherine Miller, Radical proposal: Third-party auditor access for AI accountability, Stanford University HAI (Human-Centred Artificial Intelligence), 20 October 2021, https://hai.stanford.edu/news/radical-proposal-third-party-auditor-access-ai-accountability ↩︎
  20. Jake Goldenfein, Evaluating AI at work: how would you know if productivity increases and cost cutting claims are true?, Which Way Forward? AI and Decent Work, symposium, Centre for Future Work, Melbourne, 9 April 2026. ↩︎