It's startling to map what kinds of AI Big Tech pursues, for profit. What if public good were prioritised? What would we see then?
Map of investors in AI, post Chat GPT. Download PDF
One of the phenomena we’ve wanted to break down, over these years of Alternative Global, is communities’ often passive relationship to Big Tech in their lives. Of course, we know that there is collective power and opportunity for voice in these networks - we wouldn’t be communicating with you right now, if networks didn’t have this intrinsic social power.
But the emotional manipulations, the egregious surveillance and the heedless extractions they also enable, makes perennially relevant the question of control over the actions of major technology companies in our lives.
So it’s useful to find this paper, from the Common Wealth think tank, which addresses the nature of Big Tech’s dominance - and possibilities to democratise it - in a sustained way. Statist, though - and we’ll come to that at the end.
Titled “Dynamics of Corporate Governance Beyond Ownership in AI”, here’s the executive summary:
As hype around Artificial Intelligence (AI) reaches fever pitch, Big Tech has only cemented its control over the development and use of these new technologies. This is not simply exercised through ownership — acquiring AI start-ups, for example — but increasingly involves complex alternative ways of exerting control.
As this report shows, the mechanisms at play include:
Using corporate venture capital to get preferential access to capabilities, knowledge and information and steer the direction of start-ups’ R&D.
Entrenching a dominant position through providing cloud services, leaving other organisations no alternative outside paying Big Tech. This includes offering cloud credits which direct start-ups and academics to “spend” this investment in particular ways such as building applications on top of AI models owned by Big Tech.
Capturing the majority of AI talent, either directly or indirectly by paying academics to work part-time for Big Tech even as they retain an academic affiliation.
Influencing the direction of research by setting the priorities at conferences and shaping perceptions of what is seen as the frontier of AI.
While open source is sometimes positioned as a source of countervailing power for Big Tech, this has also been co-opted — Big Tech companies open source certain pieces of their software or models, as when Meta made its Large Language Model (LLM), Llama, open source.
This allows them to both benefit from a community of developers willing to work for free on the code and make possible improvements, but also makes it easier for developers to build applications on their own models and software, cementing their dominant position.
What is needed is genuine alternatives to for-profit AI, ones oriented towards people and planet. This needs to be underpinned by opening up datasets — including Big Tech datasets — where this will support collective good, as well as a public research institution to facilitate large-scale collaboration.
We found the following section particularly interesting, as a vista of alternative
Against [a backdrop where] the most likely scenario is that AI will remain controlled by a few US giants, there is an urgent need to offer public alternatives to the for-profit development of AI. The following sections offer three areas ways to explore developing AI for public good.
AI is not simply the cutting-edge technology of our time. It is a general-purpose technology, which means that it can be applied widely to the most diverse uses. And because it is a new method of invention, it has direct impacts on how scientific and technological investigations take place.
AI for addressing major challenges
AI is particularly useful for processing and synthesising large amounts of information, with multiple possible applications for addressing global challenges.
For example, AI could be deployed in planning the green transformation, both strategically and for solving specific chokepoints. Renewable energy management is made more efficient by centralising data about supply and demand and processing it with AI.
Cloud giants are already offering computing services for this purpose while gathering energy data. Instead, the state could be a steward of energy data sources and could use AI for public smart grids aimed at providing a more efficient, cheaper and socially-just distribution of energy.
Predictive maintenance of forests and more generally natural parks, using computer vision and other AI models could improve wildfire predictions and inform on the best species of trees to be planted according to the type of soil and forest.
Applying AI for these and other social and environmentally beneficial purposes should not be confined to market solutions, among others, because of a lack of private incentives for adoption coupled with the urgency of tackling the ecological breakdown.
Many applications of AI for the environment do not require personal data. They could rely on, for instance, weather data that are or could be accessed for research purposes. Currently, the same companies with the largest proprietary personal datasets are developing models with European public data.
In November 2023, Google published a paper in the journal Science with the results of GraphCast, an AI model trained with past weather data from the European Centre for Medium-Range Weather Forecasts (ECMWF). Even if the model is put in open source, Big Tech expands their knowledge on AI as they train models with public datasets. The resulting model can be adjusted or form the basis of other models then sold as a cloud service.
Today, computing services for weather forecasts are already available on Big Tech clouds and public ECMWF data and GraphCast could be used to improve those services.
Public data should not be shared with companies that harvest data from the internet and then keep it secret. Reciprocity should prevail. If tech companies want to access public data, they need to be reciprocal and share their datasets with the wider community for public research purposes.
Another application for AI is in healthcare. Today, healthcare digitalisation is driven by self-control and self-discipline, with wearables nudging people and making them individually responsible for their health and wellbeing. Instead of seeing healthcare as a social responsibility, these devices further individualise it.
But there are other alternative uses of AI for healthcare that could bring larger social benefits. Research has shown that the global health and biomedical sciences agenda is dominated by investigations on cancer and cardiovascular diseases approached from the field of molecular biology.
Investigations on neglected diseases, pathogenic viruses, bacteria or other microorganisms, and biological vectors were marginal at least until 2020, the period covered by this research.
While some of the latter changed with the pandemic, something that has not changed is that research on prevention, social determinants of health and assessment of socio-environmental factors influencing disease onset or progression remain overlooked.
The approach that dominates, both in terms of privileged areas receiving private and public funding for research, is the therapeutic and specifically pharmacological intervention with drugs.
There is, then, a lot of space for changing the approach towards a more holistic perspective. From focusing on treatments and looking at molecular degeneration, one could envision research that holistically synthesises multiple data sources to consider the social determinants of health and assesses socio-environmental factors influencing disease onset or progression.
This would require an interdisciplinary team, using complementary qualitative methodologies to make sense of AI-powered results and access to healthcare data.
Data solidarity
AI requires data and this is one of the advantages of Big Tech; they have been amassing freely harvested data from citizens and organisations wide and large for decades. Opposing this free data harvesting with data privacy could, to some degree and if the companies comply with the regulation, put limits to their spoilage.
However, it will not build alternatives to use and develop technologies that are not mainly driven by profits but, above all, good for the people, other living beings and nature.
Public databases, such as healthcare datasets, should be built on the principle of data solidarity. The notion of data solidarity refers to the decision to share data and information between actors and countries.
As Kickbusch and Prainsack explain, in the Covid-19 pandemic, a false dichotomy was installed between respect for the privacy and freedom of individuals and the protection of health. This dichotomy is false because individual freedoms require collective goods, commons that enable the realisation of those individual freedoms, and vice versa.
Anonymised health data offer an example. Under proper regulation and governance, its access could not only help to address pandemics and other global crises, but also to identify courses of action to improve the living conditions of the population. Healthcare data processed with AI could also provide evidence to improve prevention, diagnosis, treatment and care delivery.
In short, it is desirable to promote data solidarity, especially to address critical situations such as ecological and health crises. The social costs, among others in terms of lives, and environmental costs of limiting access to data may be too great.
Data governance for these purposes could be placed in the custody of the World Health Organisation. The use of extracts from health databases should be guaranteed and facilitated for research purposes while preserving anonymity. The WHO role could be to ensure that data are shared only when their analysis is associated with investigations that contribute to increasing the public good and whose results remain in the public sphere.
A related promising project could be to centralise Electronic Healthcare Records internationally or inside Europe. The project itself requires experts in computing science as well as public health policy experts, researchers from healthcare and biomedical sciences and social scientists.
Instead of moving towards this direction, the UK government has increasingly provided access to NHS data to large tech companies, providing them a fundamental resource for advancing in the commodification of healthcare.
Amazon had a deal with the NHS to provide its Alexa devices to hospitals, potentially harvesting their healthcare data. By the end of 2023, the NHS granted the tech surveillance company Palantir with a £480 million contract to run its data platform.
No wonder Google has included the NHS on its recently created strategic partners’ list for healthcare in Europe and is developing an LLM tailored for healthcare while trying to strike a deal with the NHS to apply it. Microsoft is researching on similar applications judging by its scientific publication entitled “Benefits, Limits, and Risks of GPT-4 as an AI Chatbot for Medicine”.
Beyond healthcare and other cases in which citizens could be offered the chance to share their data with states when the latter if providing them a service, data solidarity shall also prevail in public procurement contracts.
States should require companies providing public services to share with municipalities and councils the data they harvest as a side-product, which of course must be gathered with the informed consent of those producing it.
These data could contribute to improving those services. For instance, data on the routes of rented bikes could provide information about where to prioritise the installation of bike lanes and where to place bike stations.
In all these and other potentially beneficial uses, data solidarity can only work if citizens are adequately informed about what data will be harvested and what are authorised uses and type of users of that data.
This requires short and explicit statements that inform citizens. The NHS app, for instance, could include a pop up in which UK citizens are asked if they want to share their data clearly specifying what data they will be sharing, their uses, who will get access and offering them the chance to get access to the research findings.
These pop-ups must provide easy ways to opt out, forbidding convoluted communications such as those frequently used to discourage us from disabling cookies. Personal data or data that can result in forms of surveillance and control should never be gathered (such as data about the individuals who are renting public bikes).
Large collaborations for AI for the common good
Currently, it is impossible to conceive of the production of foundation AI models without Big Tech. No single existing institution or organisation can afford to produce LLMs or other advanced AI models without relying on technologies controlled by Amazon, Microsoft or Google which could be extended to include AI GPUs from Nvidia, though the three cloud giants are also designing their own AI semiconductors.
Relying on other tech giants would not make much difference in terms of technological subordination, with the exception that other tech giants will probably not deliver frontier technologies, which are in the hands of those US Big Tech. States must build independent alternatives in the aim is to develop AI for the common good.
AI requires a lot of people working together. The OpenAI’s report introducing GPT4 was authored by 276 employees. The report includes a special acknowledgement to Microsoft, because some of its employees also contributed to developing GPT4, and to adversarial testers and red teamers (OpenAI, 2023). If over 300 people are needed to develop an LLM, it is unreasonable to expect a start-up or university doing it by itself.
The development of alternative foundational AI models will not be funded by venture capital, which is controlled by Amazon, Microsoft and Google and funnelled into their own research priorities. Only public funding can support transformative AI that puts social and environmental challenges first.
A new public research institution to support the development of foundation AI models should ideally be an international collaboration. It should develop models for the public good, taking into consideration the environmental, social and cultural impacts of AI, freed from the imperative to make a profit.
This institution will need to bring talent back from Big Tech. This would require the five “Ps”:
A clear purpose or mission.
Sufficient processing power, that could also be used widely for other public research projects.
Access to public datasets, achieved through data solidarity.
Adequate pay (which may involve an increase in academic salaries across the board).
Peers because foundational AI cannot be produced in isolation or in small groups with a principal investigator working with younger scholars. It requires hundreds of experienced people, peers, working together.
In general with this excellent paper, we would however note the emphasis on state power, and state-and-interstate regulations, to effect many of the countervailing powers to Big Tech.
In that spirit, keep an eye out for our blog on Open Civics at the end of the week, which sees the possibility of open, planet-friendly socio-technical networks arising from strong, self-determined communities - at least leading in advance where states might follow, as partners and supporters.
Again, full paper here.