Ethical aspects

Ethics refers to questions involving moral concepts such as right and wrong, or good and bad. 

Normally people intuitively know what is right and wrong, an ethical knowledge that is acquired during human socialisation. However, different concepts can come into conflict with each other, e.g. when the self-driving car has to decide whether to run over the child or put the driver in danger.

At this point, however, concepts such as Immanuel Kant's categorical imperative "Act only according to that maxim by which you can at the same time will that it should become a universal law" or John Stuart Mill's idea of utility, which goes hand in hand with the greatest happiness for the greatest number of people, should not and cannot be analysed. This is what philosophy should investigate.

Here we try to show some aspects that are more generally understandable.

Incorrect information

Even if the accuracy of the statements made by the widespread AI offerings quickly improves, it will probably remain with incorrect statements or images.

Probabilites

The AI tools are largely based on artificial neural networks. As we have seen, the statements are of a statistical nature. This means that there can be no 100% correct answers. Errors are inherent to the system.

Due to the amount of data it processes, it can sometimes draw connections or conclusions that are statistically valid but may not make sense in a real-world context.

Outdated training data

Training data may be out of date. Newer knowledge is therefore less likely to be represented in the learning data. the output is no longer correct.

Biases

Although an AI is trained on a mix of licensed data, data created by human trainers and vast amounts of text from the internet, this broad knowledge base can contain biases and inaccuracies in the data. This also leads to it providing answers that are not always accurate.

For example, one study found that the image generation tool Midjourney only returned images of younger men and women for non-specialised job titles. For specialised roles, both younger and older people were shown, but the older people were always men.

This implicitly reinforces a number of biases, including the assumption that older people do not (or cannot) work in non-specialised roles, that only older men are suited to specialised work, and that less specialised work is the domain of women. (Thomson & Thomas, 2023)

Actually, OpenAI tries to avoid such images in Dall-E 3. We have seen that it more often presents police officers with dark skin as the default. But of course such a role-based programmed reversal message is also a bias.

Energy consumption respectively environmental pollution

Why is the question of energy consumption categorised here as an ethical issue? 

If the concrete danger of climate change is not denied, then this question arises because the vast majority of the energy consumed still comes from the burning of fossil resources, which releases large quantities of CO2, which is responsible for the greenhouse effect. Even if it is argued that this consumption for AI may come from renewable energy (wind, water, solar), this clean energy could be used for other things.

Sources of Global Energy (Alley et. al., 2023)

In fact, it is very difficult to research current statistics on the energy consumption of ChatGPT. The information varies widely. Let's try anyway:

Overall consumption

When training a large language model, one computing unit out of thousands can consume over 400 watts during operation. As a standard rule, you also need to consume a similar amount of electricity for cooling and energy management. "Overall, this can lead to up to 10 gigawatt-hour (GWh) power consumption to train a single large language model like ChatGPT-3. This is on average roughly equivalent to the yearly electricity consumption of over 1,000 U.S. households."
"Today there are hundreds of millions of daily queries on ChatGPT, though that number may be declining. This many queries can cost around 1 GWh each day, which is the equivalent of the daily energy consumption for about 33,000 U.S. households." (McQuate, 2023)

Perhaps  that doesn't sound like it's too much. That's why it's interesting to look at each request in comparison with a search engine query.

Consumption per request

The following calculation is based on the statement by Sam Altman, CEO of OpenAI, the operator of ChatGPT:

"average is probably single-digits cents per chat [...]" (Altman, 2022)

The user KFilter calculates on ai.stockexchange.com, based on this statement, assuming that this is €0.09/request in the worst case [Actually, it should be 0.09 €/chat. (W.E.)]:

"I guess a least half the cost are energy at a cost of 0,15€/1kWh, a request would cost 0,09€/request*50%/0,15€/1kW=0,3kWh/request = 300Wh per request. 60 Smartphone charges of 5Wh per Charge ;)". (KFilter 2022)

This calculation is definitely suitable for technical or business subjects at vocational schools. The final comparison with charging a mobile phone is nice, because this is of great value to our students, even though the electricity bill is often still paid by their parents.

There was another more recent statement in 2023. This time from Alphabet (Google etc.) chairman John Hennessy: He "[...] told Reuters that having an exchange with AI known as a large language model likely cost 10 times more than a standard keyword search, though fine-tuning will help reduce the expense quickly." (Dustin et. al., 2023)

This comparision with the classic search engine would be therefore cheaper by a factor of 10 (according to Hennessy) or by a factor of 1000 (Altman), since Google, for example, as Google stated its energy consumption per query at 0.3 Wh (Google, 2009). We don't really know what's right. Anyway, the current figure below shows a 10 times higher energy consumption of ChatGPT in comparision to classic Google search (De Fries, 2023).

A possible learning objective could be in the classroom: From an energy-sustainable perspective, use a classic search engine rather than a large language model for simple queries, as the energy consumption for queries is many times higher.

Whether this will remain so simple in the future is questionable, however, as Microsoft's Bing search engine already uses the Copilot chat engine, which makes its statements with the help of queries in GPT-4 Turbo. The boundaries between the two options could become blurred in the future.

Alex de Fries' research predicts much higher energy consumption from AI-powered Google searches:

Estimated energy consumption per request for various AI-powered systems compared to a standard Google search  (De Fries, 2023).

Zeit Online quotes Heike Brugger, Head of Energy Policy at the Fraunhofer Institute for Systems and Innovation Research: "The use of artificial intelligence will double the demand for future data centre capacity in the coming years." (Neumann, 2024) 

Precarious working conditions for data workers

The training of large language models is initially unsupervised, namely when it comes to learning statistical word patterns. Later, however, during fine-tuning, so-called supervised training is added, which means that statements made by the chatbot must be checked and corrected by humans in order to avoid false statements or so-called hallucinations. Here, the big tech companies often rely on large quantities of data workers. The quality of their work has a direct impact on the performance of the AI models. Companies have been criticised for the precarious working conditions and inadequate salary of data workers.

Let's read what ChatGPT itself has to say about this:

Question

"What do you know about the precarious working conditions of the data workers who are or were used in the training of ChatGPT?"

ChatGPT