Can there be pro-worker AI under a deregulated capitalism?

Second, profits constrain choices. Under real competition, a “pro-worker” tool that maintains higher labour costs will be defeated by an “automating” tool that reduces those costs. The “pro-worker” firm will be driven out of business by the firm that sets the lower regulating price.

There is no “choice” at the firm level here any more than the firm can “choose” to set its net prices/revenue below net costs. Capital will simply flow towards the higher profitable firms.

Third, at the capital augmentation process, the report confuses speed-up with productivity.

We have talked about this before, but "productivity" does not mean what people think it means. Doubly so in the neoclassical field. However, we do not even need to get into that, the report fails to understand the impact of AI at the management level.

Management uses AI/algorithmic management to reduce idle time (of managers and workers), not just to manage “better”.

This is speed-up (extracting what management can call getting more effort) and is not a true “productivity” gain, that is producing more with less total labour time or the same with less labour time. But, the impact of speed-up is increased profitability.

Without a union, there is little that can be done without regulating the floor for all workers. Such a regulatory approach is an active dynamic regulation, and cannot be achieved through some “privacy” acts or tax incentive.

Fourth, “new task creation” is simply believing in the process of natural employment creation. This is an expression of the assumption of “full employment” in the neoclassical economy. You can infer different ways to get to full employment through this belief, but it does not mean that it what is going to happen or in an equitable way.

Our critique of this is broad and is what the “Just Transition” program is about. We do not believe that “large”, single-point, displacement of jobs is exactly like the churn of the regular economy. It requires specific policy tools and state-level intervention to deal with. Large and broad displacement of jobs aligning with particular skills operate in a similar way. The reserve army of labour becomes large quickly from broad systemic layoffs, driving-down wages of the potential jobs that those workers can go into. This shift is a novel harm because of the speed/timelines involved.

There is no guarantee that jobs are created for the displaced workers, what is assumed to be created is an equivalent number of hours of work doing “something else”. What is also ignored is that the work hours created can be created producing the automation industrial machine in a different place/country than the laid off worker. Which is obviously of no help to the laid off worker or affected community.

This well known process is ignored as is the need for state intervention in ensuring investment strategies building work other than “retraining”.

Fifth, Simon Johnson says that AI is “distributed over the digital rail”. This is similar to our analysis that there is a full supply chain and physical infrastructure that is important to consider when talking about “AI” and (flawed) concepts of “digital sovereignty”.

Canada’s digital rail is almost entirely foreign-owned IP and infrastructure. And, Cohere is no different from Anthropic here except the flag on the name plate. It is entirely reliant on foreign physical infrastructure and its model is not open and thus is under the same market logic that makes other AI firms objectionable.

Canada does not currently own the underlying IP for the “Physical Layer” including semiconductors, technology found inside a Data Centre, and even the data because of the US CLOUD Act. That includes Cohere or any other Canadian “champion”.

Some ministries of government have been good at clearly defining “Sovereignty Digital” as sovereignty over the logic and data layers. That includes the software level, publicly owned, open-source LLM models, running on Canadian-governed encryption, that can run on any hardware. There is a model that exists, but it is not being identified under these policies.

Without a public option, you do not have the ability to impose any choice in the market on what models are tailored to do, are limited to do, or provide a price point that is competitive without dealing with cost competition that drives those “choices” to be nice firms out of business.

There are other issues with the report. Mostly because it is a neoclassical (New Keynesian) expression of the current problem resulting in many points of flawed policy. This reason is also the reason that Carney's folks love this report. Central bankers live in this ideology.

But, we wholly reject this ideology on the Left. There is no reason to think AI policy based on this same rejected ideology could result in “pro-worker” policies any more than central banks could make pro-worker interest rate decisions.

The basis of "AI" type systems is what we used to call Big Data. That is, the large, centralized data sets that we put into centralized databases and on the internet have allowed the training of AI LLMs/generative systems that power those chat bots and agents.

Those public datasets have now been exhausted. Like a mine that no longer produces, companies selling this technology look to new data collection. What has been added to this list is mined surveillance data.

This is not a new process. Surveillance for improving management of workers to attempt to increase efficiency goes back to-before the birth of capitalism and underlies the one of the most present tendencies in capitalism of the rising Organic Composition of Capital (OCC). The process of moving towards more fixed capital paid for by reducing the ratio of workers needed for output.

This transformation is only possible through surveillance because industrial machines that automate workers must be imbued with the organic knowledge from the workers who used to do the work. Information must be collected by those who know how the process is carried out for it to be automated.

AI is the scaling-up of this surveillance of organic knowledge to drive automation.

"Surveillance" can happen in different ways. For software coders, the vacuuming-up of the code that runs software is similar to the surveillance of workers on the shop floor, or the surveillance of truckers/cabbies/consumers to train automated driving systems. It is the collection of the process of production that the system is driven by.

The AI conversation is also confused by the separation between consumer-level AI toys and industrial-scale applied AI tools and machines.

The lack of understanding of the complexity of modern production processes leads to people to think that they understand AI because they have used a chat bot or played with "agents" in preforming some task.

People think that a chat bot is generic enough to just point at any task and automate it. However, this is not how things work. The information about how to generally do a task might be included in some large dataset that trained/is the AI system, but that data needs to be added to by local proprietary and organic knowledge.

That local knowledge then needs to be coordinated in usually unique-to-firm ways of producing an output of some kind.

In AI, it is known by fancy terms including the word "augmentation". This information, however, is usually not written down anywhere—even in very large and complex firms. It definitely is not written down in small firms.

Consumer AI is not exactly what we mean when we are talking about the application of AI at the industrial level.

The process of transforming organic knowledge into industrial machines is not easy or cheap. It usually is more expensive up-front that simply hiring workers to do the work.

The cost of business transformation is one of the reasons that the new AI strategy released by the government of Canada misses the point. The strategy "gets it" in that the focus is on profit subsidies because the Liberals think that money is the barrier for adoption. But, the government does not get that profit subsidies alone will not make firms adopt AI any faster than without that subsidy. It is the central banker problem of pushing on a string.

Adoption of automation processes is not a straight line, has real risks that cannot be subsidized away, and the end result usually reduces quality of the product while also not always guaranteeing a profit for the firm.

Like with all investments and risks that a firm is thinking about taking on, the firm will not go ahead with that investment unless there is sustained profit all but guaranteed on the other side of that investment.

The subsidies are likely to be wasted in many places that it is tried by supporting companies that had no reason to try to apply AI to their systems. Where that application of AI is successful, it is unlikely that the subsidy was the determining factor in the success.

Profit subsidies, in this way, are going to turn out to be a rather large waste of money.

And, are going to promote the increased demand for "tokens" beyond what was necessary. This is the other side of the profit subsidy regime outlined by the government. And it is why the biggest proponent of the AI strategy is the industries around AI and not the users.

The government says it is a worker-focused and neutral program on AI. Nothing could be further from the truth. The entire system is a profit subsidy regime for the AI industry, an industry that is neither wanting for money nor deserves our handouts.

What does this mean for the left and AI?

For consumer-level chat bots, surveillance, and "safety", there are clear issues with the way that unregulated corporations are abusing all of us. The answer, however, is the same answer that that the materialist left has for all things. Building understanding, interrogating the issue with debate and study, and democratic processes dealing with the extremely complex impacts of technology on our lives.

Society is not something to be left to companies to decide, obviously.

For industrial AI, the process of responding is similar but, it includes the addition of workers. The industrial process is also what is freaking many people out when it comes to jobs, wages, layoffs, defence, interacting with government, and the economy generally.

The applied analysis must include both of these parts of the AI system and seek to understand how these different complex systems interact if we are going to provide analysis that makes sense.

To start with, we need to frame things along the following:

• Ownership and worker voice.
• A Public AI Option that is de-commodified and compliments worker and the public's economic and social interest.
• Data localization that actually keeps IP and its application under Canadian law.
• Failure management through establishing clear workplace protocols for AI "hallucinations", intermittent connectivity loss, and the psychological effects of constant machine-worker interaction.
• Human-in-the-Loop that mandates human circuit breakers and prohibiting any termination or discipline based only on automated AI outputs.
• Collective bargaining and social policy to ensure AI is used for augmentation, not as a tool for labour intensification through speed-up.
• Policy that clearly regulates data harvesting for AI training and establishes Information Rights for workers over algorithmic logic and management tools.
• Sovereign software and encryption mandates where Canadian public sector data is protected by encryption that is not subject to foreign "back-doors" such as through the CLOUD Act.