What is AI and what different models are there
Artificial intelligence, or AI for short, refers to the ability of machines to imitate human abilities such as logical thinking, learning, planning and creativity. AI models are trained with existing information and encouraged to provide solutions using algorithms. Artificial intelligence is able to learn lessons from previous actions and adapt courses of action accordingly. The complexity of the analyses running in the background increases with each data set that is fed, so that the AI becomes increasingly precise over time.
The first AI-supported algorithms have been around for more than 50 years. However, the maturity of the technology has increased rapidly in the last two decades in particular. More resources, better developed technologies, greater availability of data volumes and targeted research ultimately led to the breakthrough that made AI accessible even to private individuals.
What models are available?
The type of AI used differs depending on the area of application. In the medical field and in finance, for example, models are used that have been trained using supervised learning. Sample data whose answers are already known - such as symptoms and their associated clinical pictures - are fed into the AI, which recognizes conclusions and patterns in order to be able to derive more targeted diagnoses in the future. One particularly exciting and increasingly successful application is in the early detection of cancer.
Another model is unsupervised learning. In addition to verified data, unlabeled data is also processed and the model is allowed to search independently for patterns and structures within this data. This allows large volumes of data from business and technology to be analyzed and structured, for example.
A special model of these two is the deep neural network (deep learning), which recognizes the most complex patterns. Deep learning is one of the most resource-intensive models, but produces some spectacular results. The widely known large language models (LLMs) such as ChatGPT (OpenAI), Llama (Meta) and Gemini (Google) also operate within these neural networks.
In reinforcement learning, the AI model is reinforced or punished depending on the feedback in order to promote the best possible output, for example in robotics or chess computers.
There are also generative models that are trained to generate new content from known data. This can be text, images or even music and videos. The best-known areas of application here are deepfakes, for example, or well-known internet trends in which superhero figures are created from your own portrait. But this is only a small part of the various possibilities offered by generative AI.
How has AI use changed in recent years?
There is no doubt that the free accessibility of AI models such as Claude or ChatGPT has also greatly increased the volume of use. Integrated AI models in smartphones and operating systems such as Gemini or Copilot, which act as everyday assistants, have also given a major boost to the global figures for AI usage.
There are several hundred million queries per day - and the trend is rising. This is a result of easily accessible AI assistants and the undeniable advantages of AI query results. Search results via tools such as ChatGPT are summarized more compactly than search queries via Google and often require less personal research. This makes them much more popular. These AI models can also help to take over repetitive tasks through training in day-to-day work. However, blind trust in the results of AI searches can also be dangerous: according to a study , only a quarter of German AI users check the results that are displayed to them. The risk of spreading misinformation as a result is high.
What impact is this AI boom having on the environment?
It is now widely known that AI usage consumes more energy than standard search engine queries. This is mainly due to the fact that AI models are located in huge data centers and require large amounts of computing power. These data centers need to be cooled, powered and, in some cases, constantly expanded. Depending on its complexity, a single query can sometimes cause the processors to overheat by up to 80 degrees Celsius. For the average energy consumption for standard requests via text models (LLMs), it can be assumed that the energy requirement is around ten times higher than for a request via search engines such as Google or Bing. ChatGPT-4 itself speaks of around 0.0003 kilowatts per hour when using the cloud version. For the more resource-intensive image generation via tools such as Stable Diffusion, Dall-E or Midjourney, the values are estimated at 0.0029 kilowatt hours per image in the cloud applications.
In order to be able to imagine this in tangible dimensions, studies speak of comparative values such as water consumption per query, meters of driving per query or cell phone charges per query. For a simple text query via ChatGPT-4, this would correspond to a water consumption of the size of a shot glass or a cell phone charge of approx. 2 minutes or 0.12 g of CO2 emissions. The following table summarizes the values in compact form:
| Type of request | Energy Consumption | Water consumption (~95 L/kWh)1 | Smartphone charges (~0.012 kWh)2 | CO2 emission (~400 g/kWh)3 | Driving equivalent (~180 g CO₂/km) |
|---|---|---|---|---|---|
| Text (ChatGPT-4) | 0,0003 kWh | ≈ 0.03 l (shot glass) | ≈ 0.025 charges (~2 min) | ≈ 0,12 g | ≈ 0.7 m (70 cm) |
| Artificial image (average) | 0,0029 kWh | ≈ 0.28 l (glass of water) | ≈ 0.24 charges (~30 min) | ≈ 1,16 g | ≈ 6 m |
| Image with Stable Diffusion XL etc. | 0,012 kWh | ≈ 1.1 l (slightly more than a normal water bottle) | ≈ 1 full charge | ≈ 4,8 g | ≈ 25 m |
The individual figures are certainly still manageable. However, if you consider the daily mass of requests for AI models, it becomes clear that AI usage is a resource guzzler that should not be ignored. Assuming there are one million requests to ChatGPT-4 per day, then an average of 0.0003 kilowatt hours becomes 300 kilowatt hours. A one-person household in Germany consumes around 1,400 kilowatt hours per year and could therefore be supplied with electricity for almost a quarter of a year with a daily quantity of ChatGPT requests. The figure is even higher for image generation. For example, one of the most popular applications, Midjourney, states that around 30-40 million images were generated daily in 2023. Assuming a consumption of 0.029 kWh per image, this results in an energy consumption of 87,000 to 116,000 kWh per day, i.e. 0.087 to 0.116 GWh. This corresponds to the consumption of around 83 single households per year.
Positive influences of AI on environmental protection issues
The enormous consumption of resources when using AI cannot be denied and the number of active users of artificial intelligence is increasing rapidly almost daily. Providers of AI solutions have a duty to make their technologies more environmentally friendly and sustainable in order to keep pace with the demand for resources. AI can also help with this by using algorithms to calculate ways of making technologies more future-proof and sustainable. However, this not only applies to the narrow field of AI research, but also to automation, energy supply, business, transportation and industry. As resources become increasingly scarce, it is essential to find ways to use existing raw materials effectively, save water and make our daily lives more sustainable - with and without the support of AI.
Save resources with remanufactured IT hardware from IT-Market.com
At IT-Market.com you will find refurbished IT hardware from the areas of switches, servers, communication, security, routers or storage. In this way, you too can help to conserve resources and make your IT more sustainable. If you have IT hardware that you would like to offer for resale, we also have the right service for you. Do you have any questions about our IT remarketing offer? Our team will be happy to help and advise you.
1 spectrum.ieee.org/how-much-water-does-it-take-to-make-electricity
2 www.theverge.com/24066646/ai-electricity-energy-watts-generative-consumption
3 arxiv.org/html/2311.16863v3