Only a few years ago, artificial intelligence felt like a technology of the future, to be debated at conferences. Today, it is shaping people’s daily lives: composing emails, optimising energy grids, forecasting floods and, occasionally, hallucinating confidently on our screens. The shift has been rapid, and for many, disorientating. Yet for all the visible progress, we are still in the early stages of what may prove to be the most foundational technology platform of our time.

As with other major technology transitions of the past, this one is likely to have profound consequences. Past experience suggests the benefits to humanity will ultimately outweigh the harms, but predicting exactly how that balance will shake out is impossible. It is critical that safeguards be built in now, while the technology is still in its early stages.

At Generation, our Global Equity and Growth Equity teams have been studying the potential implications of this technology since 2011. In this special report, we will outline some of our thinking about those implications, the potential use cases, the possible harms — and the improvements in governance needed to protect society, and the planet we all live on, from those harms.

Perhaps the most critical immediate issue is the exploding power demand associated with running the latest AI models, and how to meet it without driving up harmful emissions. The largest AI models can demand immense amounts of electricity. Training OpenAI’s GPT-4 alone consumed an estimated 50 gigawatt-hours of electricity,^a roughly equivalent to the annual usage of a small city. Today, AI-related data centres already account for 1 to 2 percent of global electricity-related emissions, and their power demand is projected to more than double between 2024 and 2030, according to the International Energy Agency. Left unchecked, this growth could strain energy grids and conflict with climate targets — particularly in geographies already facing energy insecurity.

The United States is a focus of particular concern, given that 65 percent of the world’s data centre capacity is forecast to be built there.^b Power demand had flatlined in that country for two decades as better technologies, like improved lighting, contributed to greater efficiency. But the trend could now reverse in part because of power demand from AI. In 2024, fossil-fuel generation in the United States rose, primarily due to the use of fossil gas. Attempts are being made across the country to fast-track new gas power plants to keep up with rising demand. The inventory of planned gas generation projects has nearly tripled, from six gigawatts in late 2023 to 17.5 gigawatts in 2025.^c Even before the huge demand spike associated with AI, the American grid was antiquated and under strain from decades of inadequate investment, unable to move electricity around the country to the degree required. The AI demand spike is rapidly worsening that problem. Efforts are beginning to supply cleaner power to newly built data centres, so that extra computing loads do not drive up emissions, but it is as yet unclear if these new sources of clean power will be able to keep up with fast-rising demand.

In addition to the energy challenge, AI is projected to consume 1.1 to 1.7 trillion gallons of freshwater annually by 2027.^d Unlike fossil fuels, water can be re-used, but there are still meaningful concerns around diversion from local communities.

While the potential for increased emissions from AI has gained a great deal of attention, the opposite prospect has barely been discussed — and could ultimately be more important. AI could be an enormous help in cutting emissions, if it is used to make systems like power, transportation and food more efficient. A recent study found that AI applications across those systems could reduce global emissions by 6 to 10 percent annually by 2035^e — more than offsetting the projected emissions from AI’s energy use. Earlier research from Microsoft and the consultancy PwC estimated that applying AI across the energy, agriculture, water and transport sectors could cut emissions by something like 4 percent by 2030 while adding $5 trillion to global economic output. These efficiency gains could also support large-scale job creation — with Microsoft and PwC’s model projecting 38 million net new roles across these four sectors globally by the end of the decade.^f

Whether forecasts like those will actually come to fruition is one of the most significant questions. Yet we are already beginning to see early examples of AI’s positive impact on the planet. For example, a company called WEKA, in which Generation has invested, helped its customers avoid more than 358,000 tonnes of emissions in 2024, through a platform that helps advanced computing workloads reduce their hardware requirements by up to 90 percent.^g Similarly, a 2023 study of Google’s AI-based aviation prediction models found that using them could reduce jet contrails, which are responsible for over a third of aviation-related global warming, by 54 percent.^h In the non-profit space, Climate TRACE’s AI-powered emissions tracking platform has created the first comprehensive, near-real-time inventory of global greenhouse gas emissions at the facility level. Combined with Altana’s supply chain intelligence tools, it enables companies to quantify emissions in their supply chains — something they could not do very well until these new tools became available.ⁱ

While AI’s net impact on the planet is still up for debate, so too is its impact on society. Evidence of social harm is stacking up: AI tools are making it easier to create ‘deepfake’ videos and photos of politicians, and they are being used for darker purposes like creating child pornography. In some teenagers, interacting with AI chatbots has worsened mental health problems and in one case possibly even led to death by suicide.^j Spammers and fraud artists are eagerly exploring what AI might be able to do for them. Governments are already using AI as a tool of mass surveillance. These early problems with the technology have demonstrated the fundamental inadequacy of the guardrails built into some models. While their creators are rushing to patch over particular problems as they crop up, we have no guarantee that AI will not result in large-scale social harm.

Privacy and disinformation concerns are mounting too. As AI systems become increasingly embedded in our lives, users often engage with them in deeply human-like ways, unlike any previous technology. One study found that 70 percent of queries across one million ChatGPT conversations included personal identifiable information, while another found that 76 percent of AI chatbot users did not understand basic privacy risks.^k Many of us are already confiding in AI systems more freely than we might confide in other people. This creates new risks, especially when there’s little transparency into how data is exploited or fed back into model training.

Without deliberate intervention, AI could also deepen structural inequalities. Access to high-quality models, computing resources and relevant datasets remains heavily concentrated in the global north. In 2024, just 1.4 percent of global AI investment went to companies based in Africa, Latin America and Southeast Asia.^l And as AI begins to reshape job roles, the divide between workers who can make use of it and those who cannot will only grow, with potentially significant implications for inequalities in both wealth and well-being.

New technologies always trigger shifts in the type of work humans are required to do. While we expect AI to augment and displace various tasks and roles performed by knowledge workers, new fields will also be created in areas such as data science, AI development and the energy sector. As AI is embedded in the physical world over the coming decades, robotics systems will lead to further displacement in the skilled and frontline workforces — though this is happening more slowly. A 2025 paper by Larry Summers and David Deming of Harvard University suggests AI is already transforming labour markets: they noticed a barbell-shaped pattern in wage growth, as highest earning and lowest earning jobs are growing faster while middle-paid jobs are not, leading to growing wage inequality.^m

The idea that AI will make work easier or more tolerable is a long way from proven. A recent study from the Massachusetts Institute of Technology found that using AI for knowledge work reduces human memory recall.ⁿ The early indicators suggest, however, that the potential for improvements in both productivity and work quality are real. A recent study at Procter & Gamble, the multinational consumer goods company, found that teams with access to AI assistance delivered noticeably higher quality work than those without, even when both sets of workers started with the same data.^o

AI also holds potential to combat social and economic inequality, especially in personalising education. For instance, platforms like Speak and Masterschool are developing AI tutors that adapt in real time to individual learning styles — including for neurodivergent students like those with autism or other conditions, who are often poorly served by standardised instruction.

The ‘how’ of AI is just as important as the ‘what’ when it comes to accessibility. This includes promoting conversational AI through voice, translating for non-dominant languages and designing with equity at the core. For example, ElevenLabs AI voice-cloning technology has made computer-aided speech for disabled people sound exactly like their real voices.^p The ‘how’ also means broadening the circle of those who participate in AI design, testing and feedback; reviewing data used to train AI for bias; and ensuring the right data privacy and cybersecurity practices.

AI is becoming embedded in the infrastructure of essential systems like healthcare, finance and education. It is driving innovation in the physical, climate and medical sciences. The critical question is not whether AI will shape our systems, but how we choose to shape AI. That choice will determine whether these technologies reinforce fragility and inequality, or help build resilience, equity and human flourishing.

Momentum is building as the positive effects of AI are felt across industries. In healthcare, companies are accelerating drug discovery and equipping millions of doctors with AI note-taking devices, known as scribes. In finance, digital platforms are lowering barriers to adoption. In climate resilience, tools like real-time flood forecasting are saving lives. These examples are early signals of what is possible when AI is deployed with purpose. But they also remind us that the window for shaping norms, standards and safeguards is narrow. Governments are beginning to regulate, companies are beginning to adapt and investors are beginning to allocate significant amounts of capital. Yet the frameworks for sustainability remain incomplete.

So what do we mean by sustainable AI? At its core, it is AI that advances human and planetary well-being while safeguarding environmental and social systems. It is AI that reduces carbon intensity rather than inflating it, that expands access rather than excluding people, and that strengthens resilience rather than creating new vulnerabilities. This is not a distant aspiration: it is a design choice available today. The genie may already be out of the bottle, but we still get to decide what kind of future it creates.

In 10 years, in a world where AI may very well design, direct and influence large parts of our economy, what will we wish we had done differently today?