By Robert Moss, April 2025

Like many of us, I have been looking at the potential of AI applications in assisting with work, online communication, and content creation. We have all been exposed to AI whether as consumers or creators, and whether we are aware of it or not. This article looks behind the convenience and hype at some of the larger environmental and societal impacts of AI on a rapidly changing World.

Environmental Impact (Yes … we will need a Planet B as well!)

The power costs of AI are horrible. The processing power required to mine the internet to generate content from a few user commands is immense, particularly now that societal engagement with AI is becoming more mainstream.

Even LinkedIn now offers the option to create content written by AI within the platform and without the need for using third-party applications such as SuperGrow AI.

At a global scale, the energy consumption required to power the data centres delivering AI application usage is very hard to estimate. This is because those organisations training and then delivering large language model AI applications do not reveal their associated energy consumption. Within a 2024 Verge article, “How much electricity does AI consume?”, Alex de Vries attempts to reverse engineer an estimate of this undisclosed energy consumption by looking at the volume of microprocessors being used within the AI industry:

“Nvidia accounts for roughly 95 percent of sales in the AI market. The company also releases energy specs for its hardware and sales projections.

By combining this data, de Vries calculates that by 2027 the AI sector could consume between 85 to 134 terawatt hours each year. That’s about the same as the annual energy demand of de Vries’ home country, the Netherlands.”

(Vincent, 2024)

Where there is reason for hope on the environmental front is through technology efficiency improvements. Much as the first steam engines in the early 18^th century were woefully inefficient at around 0.5% in the case of the Newcomen engine, it is likely that the energy efficiency of large language models and the microprocessors that power them will increase:

“Thomas Newcomen (1712) introduced a steam engine to pump water out of coal mines in England with an efficiency of just 0.5%. … Modern combined cycle power plants anchored by the gas turbine have efficiencies that exceed 60%.”

(Visualising Energy, 2023)

We see early signs of such progressive efficiency gains with the launch of the Chinese DeepSeek large language model in January 2025. DeepSeek requires much less processing power for delivering similar results to its western large language model equivalents … allegedly.

Advanced Micro Devices (AMD) has also been betting on future growth in the development of device orientated processing architecture for AI and gaming graphics. This development could remove or reduce the need for large data centres providing remotely accessed large language model processing, along with the associated huge energy consumption.

The Irish Perspective (Gunna need more Bit Barns!)

In Ireland, AI server growth has contributed toward the electricity energy usage by data centres having overtaken that of both rural and urban residential electricity consumption by 2023. The Republic of Ireland Central Statistics Office (CSO) reports that, “In 2023, urban households accounted for 18%, and rural households for 10% of total metered electricity consumption” (CSO, 2024). This compares to a percentage of total metered electricity consumption in Ireland being used by data centres having reached 21% in 2023.

According to the EPA, electricity demand in Ireland increased by 3% in 2023 from 2022. This said, over the period from 1990 to 2023 “emissions from electricity generation have decreased by 22.1% whereas total electricity consumption has increased by 164%” (EPA, 2024). For now, the improvement in emissions reduction by increased renewable electricity generation in Ireland is more than keeping pace with the demand for more electricity.

While the significant increase in electricity generation from renewable energy is a great success story for Ireland’s climate change mitigation efforts, further progress is limited by the current generation and transmission infrastructure:

“Achieving the renewable electricity target of up to 80% will entail investment of tens of billions of euro, including in the installation and maintenance of generation assets, and associated infrastructure and services, as well as in the development of supply chains and port infrastructure.”

(GOVERNMENT OF IRELAND, 2021, p96)

In the long term, the limitations of the current electricity transmission network to transport electricity to where it is needed (Dublin and eastern Ireland) from where it can most reliably be generated renewably (western Ireland) will be addressed by re-designing the transmission system, and the strategic planning for demand centres of electricity, such as data centres, to be located more closely to where the renewable electricity will be generated:

“Energy demand, including data centres, will be expected to operate within sectoral emissions ceilings and further signals will be required to locate demand where existing or future electricity grid is available and close to renewable energy generation.”

(GOVERNMENT OF IRELAND, 2021, p95)

Clearly, there is a finite limit to the supply of electricity in Ireland, not least because of the infrastructure required to transport and deliver it to end users such as data centres. There is therefore an economic as well as environmental cost involved in supporting an expanding AI industry in Ireland, and the cost of that expansion will eventually become unattractive.

Financial Impact (This is going to Hurt!)

As of the time of writing this article in March 2025, the Stock market capitalisation of the 58 largest AI companies stands at a “total market cap: $14.996 T” (Companiesmarketcap.com, 2025). How will AI deliver $15 trillion in profit to the stockmarket in the near future? It won’t! The investment craze for AI is a speculative bubble and these always burst. There are plenty of caveats about this however. Many of the largest AI companies such as Microsoft and Apple are not solely invested in AI, although they are strategically investing in it. Also, some of these companies will succeed and deliver shareholder profit. This said there are limits to the commodification of AI technology, and I suspect these lie significantly below $15 trillion for the foreseeable future. The uses of AI technology and applications are diverse. They span across research, administration, content creation, surveillance, knowledge transfer, communication, and disinformation. However, the ability of AI to commoditise this assistance is questionable … beyond assisting with $15 trillion worth of job cuts. In terms of automating manual work this can likely be done with only the most rote of tasks, and robotics have largely delivered this in ports and factories already.

Societal Impact (It’s hard to read … thanks for that Social Media!)

Copying creative ideas and styles, even passing your own work off as someone else’s is nothing new. However, the scale of large language model appropriation of information, both real and fictitious, is unprecedented. Content can now be produced and disseminated to millions of people in seconds. This has societal consequences more discreet than the impacts already discussed, but possibly of greater significance in the long run.

There is little time to consider, reflect, and learn in an online society where information is shared and promoted at frenetic pace, and on an unfathomable scale. It does seem that most people are broadcasting with few people listening. A doom spiral of shouting ever louder for less attention.

Large language models not only accelerate the speed and volume of content creation, but they also remove the guard rails of experience and knowledge. Anyone can be an expert on anything … kind of. All you have to do is type in a few commands, and maybe direct the large language model to online text or video, and then the application will do all the work. The fact that the content creator initiating this instruction learns nothing about the subject at hand may seem to relieve a burden for them. What is most problematic is that the content creator may not have had and personal knowledge to impart in the first place. Anybody with little or no knowledge, or even interest in a subject, is unlikely to invest much time in reviewing, editing, and correcting the large language model output they have created. What would be the point if they are largely ignorant of the relevant facts and nuances anyway?

So we now have content on internet media platforms that is unedited, and lots of it. Clunky, corny, full of superlatives and repetitions, as well as counter-factual hallucinations. An example is that of an aberrant video switching from the subject of Parisian Architecture to video clips of French Polynesia for no reason connected to the script. The algorithm of its large language model creator seemed to have not detected such nuances as French Polynesia being an overseas territory of France rather than a department of metropolitan France. Such glitches compound. Over half of text content on the internet is now artificially created:

“roughly 57% of all web-based text has been AI generated or translated through an AI algorithm, according to a separate study from a team of Amazon Web Services researchers” (Constantino, 2024).

Every time a large language model trawls through the internet for online knowledge and content it returns from its ramblings with the increasingly strange fruit of previous large language model creations.

The real danger is not the use of large language model applications themselves, but the plasticity of the human mind and by extension human society. As large language model usage continues to churn out facsimiles of facsimiles of facsimiles, then we are in danger of growing comfortable to its strange mangling of reality.

Critical thinking? Attention to detail and empirical research? Yawn! … where’s the factish-based infomercial conspiracy video content?

---

References:

Companiesmarketcap.com (2025) Largest AI companies by market capitalization. Available at: https://companiesmarketcap.com/artificial-intelligence/largest-ai-companies-by-marketcap/ (Accessed: March 8, 2025).

Constantino, T – Forbes Australia (2024) Is AI quietly killing itself – and the Internet? Available at : https://www.forbes.com.au/news/innovation/is-ai-quietly-killing-itself-and-the-internet/ (Accessed: March 8, 2025).

CSO (2024) Data Centres Metered Electricity Consumption 2023. Available at: https://www.cso.ie/en/releasesandpublications/ep/p-dcmec/datacentresmeteredelectricityconsumption2023/keyfindings/ (Accessed: March 7, 2025).

Environmental Protection Agency (2024) Energy. Available at: https://www.epa.ie/our-services/monitoring--assessment/climate-change/ghg/energy-/#  (Accessed: March 7, 2025).

GOVERNMENT OF IRELAND (2021) Climate Action Plan 2021: Securing our Future. Dublin. Available at: https://www.gov.ie/en/publication/6223e-climate-action-plan-2021/

Vincent, J – The Verge (2024) How much electricity does AI consume? Available at: https://www.theverge.com/24066646/ai-electricity-energy-watts-generative-consumption (Accessed: March 7, 2025).

Visualising Energy - Boston University Institute for Global Sustainability (IGS) (2023) Maximum efficiencies of engines and turbines, 1700-2000. Available at: https://visualizingenergy.org/maximum-efficiencies-of-engines-and-turbines-1700-2000/#:~:text=Thomas%20Newcomen%20(1712)%20introduced%20a,better%20than%20all%20previous%20versions. (Accessed: March 7, 2025).

Banner photo: courtesy of Robert Moss