Is it possible to use LLMs ethically?

To answer whether you can use LLMs ethically, we need to consider environmental factors, data bias, social inequalities, ownership of training data, human-cost of fine-tuning, impact on jobs…? The list goes on.

I work with a lot of mission-driven companies and non-profits, and I get this question all the time. There’s so much to unpick, made more challenging by the different meanings of “ethical” from person to person, but I thought I’d make a start on some of the areas that can contribute to a decision.

Environmental factors of LLMs

We’re starting with the trickiest to pin down. And the reason is that large tech companies, the ones with the LLMs, are not particularly great at sharing information on their environmental impact.

Creators of LLMs often refer to them as being cloud-based, which might have you believe that they just sort of exist in the air, but this is far from the truth. LLMs require giant, physical data centres which take up large areas of land, decimate local water supplies and need significant amounts of power to run.

Some of these data centres are built in the countries that most use LLMs, but I’m sure it’s not a surprise to learn that these are also often built in developing countries. Often in communities with limited access to water already, who need live off the very land that is taken away. Data centres are also incredibly loud, destroying the peace of those nearby.

So before we even consider the carbon footprint, there’s an environmental impact for those who now live near data centres.

Of course the large tech companies are really transparent about the tech centres they’ve built, how much water and power they use and what the emissions are. Oh no, wait, they aren’t at all. What’s interesting is that OpenAI’s own training resources cover this topic, so you’d think we’d have a clearer idea for ChatGPT. Instead it mentions that figures are based on speculation and link to external research. OpenAI says it is working on minimising their carbon footprint, but what we really want is transparency so users of the technology can make their own decision on whether it’s worth the carbon cost.

We shouldn’t have to rely on guesswork because these things can be measured. Until this is enforced, we’ll not truly know the impact we’re having by using AI technologies. And for some people that means that using them is unethical.

Data bias

It’s well known, and accepted, that bias exists in everyone and can’t be eliminated.

We’ve all seen the social posts about “Generate an image of a doctor” with every one coming back as a white man. And whilst LLM owners are doing what they can to fix this there is no way to totally eliminate bias in the data – because everything has some element of bias.

What I think is a greater issue here than generating bias (because I don’t think that can be prevented), is the lack of ownership and responsibility. When a human says something that is biased and unacceptable we can be held accountable. But because these LLMs are trained on vast quantities of data written by many different people, and the technology is developed by large teams, it’s an almost impossible task to assign responsibility for what is generated. Is it the fault of the originator of the data? What about the researchers that selected the data? Or the engineers working on it? The tech giants making the LLMs certainly aren’t accepting the responsibility.

This is where I think companies that are using AI technologies in their products, services or processes need to have really clear conversations about who they consider responsible for the outputs. What is acceptable and what is not? How does the company detect when something unacceptable happens? And whose job is it to action that? What does actioning it look like?

A lot of this depends on where the LLM is being used; whether it’s customer-facing, internal only, and whether there’s always a human-in-the-loop. But regardless of the impact, there needs to be a clear understanding of where the responsibility lies.

Social inequalities

We understand that an LLM is trained on data, billions and billions of documents, and what it generates is based on its training data. So given that the majority of the data has come from the internet, and the internet is not an equally balanced representation of all languages and cultures, there is going to be some areas in which the LLM has more data to pull from.

This means people whose first language is not English may get poorer responses than those speaking in English. It may also respond with incorrect cultural explanations, based on how things are represented by English speaking countries, rather than how the people living in those cultures would explain – further exacerbating the issues of Eurocentrism.

And if non-English speakers are unable to interact with the technology because it’s not good enough, or simply because it cannot handle their language, then they may fall behind in areas where LLMs are powering advancements.

This is such a huge topic and I’m not suggesting this has gone any way to covering it – it goes WAY beyond LLMs, and is a problem for all AI technologies – but this is a small introduction for now.

Human-cost of fine tuning

Earlier I referenced how LLM-owning companies are working to address bias. The way that this is often done is contracting people from developing countries to evaluate LLM outputs and rate or categorise it. On the surface this may seem fine, but the reality is they are paid very little and have no guarantee of steady work, resulting in them working all hours to make money when they can.

On top of this, the data they are evaluating is often graphic in nature, abusive and traumatising. Reports from the likes of journalist Karen Hao, who covers this in her book Empire of AI*, speak of individuals becoming so disturbed from the work they damage relationships with their friends and family – and are offered no support from their employers. So their life is ruined.

The irony is that the people being most impacted by the development of this technology are the same people most affected by the social inequalities, or the environmental factors, discussed above.

Large tech companies often hide this human-cost, preferring to sell the LLMs as “magical” technology that just does what it does. When in reality the reason it does what it does is because of the low-cost human labour behind the scenes.

Impact on jobs

Five years ago if you wanted some copy for your website you had to write it yourself or hire a copywriter. Same thing if you wanted a cute new headshot – you had to enlist the help of another human (or a tripod and a timer). All that’s changed, now you can plug some words into to your favourite LLM and get an arguably okay result.

My opinion on the ethics of this has changed in the last year or so. I used to think that creative jobs will still very much exist (I still do!) and that the creme de la creme would raise to the top, but for those small, somewhat basic tasks, using an LLM would have no real impact on society. But I’ve changed my position. Whilst I still think that the jobs will exist and that those who are really good will still be very much in demand when you want some proper writing, or really professional pictures, by giving the entry-level tasks to an LLM we’re removing the training ground needed for those excellent creatives to hone their craft.

Decisions on whether using LLMs where a human would have otherwise been employed is an ethical decision companies need to make. It might be that your company is entirely profit-over-everything, and it makes no sense to hire someone when a machine can do it fine. For other companies this is very uncomfortable ground.

Ownership of training data

To follow on from LLMs taking jobs, creatives are being doubly impacted. Not only are the entry-level jobs in their industries at risk, but it’s their creative works that has trained the very LLMs they are being replaced with. This had led to multiple lawsuits from Disney, The New York Times and others against the various LLM owners.

The grey area here, though, is in the term “fair use”. In June 2025 a US judge ruled that training LLMs on books is considered fair use justifying it by claiming that Anthropic’s use was “exceedingly transformative”.

This is an un-proven and changing area, with lots of cases still outstanding and not every country is yet to set a legal precedent – I’m sure we’ll see some headlines on this topic in 2026.

So where does that leave us – can you use LLMs ethically?

I feel like I’ve not even touched the surface of the ethics of LLMs. We’ve not talked about data security, scams and deep fakes, what this means for our children growing up…and so much more. But hopefully I’ve done enough to introduce you to the topic.

Now, my purpose with this article is not to make the decision for you on whether anything written here makes LLMs unethical for your company, only you know what is within the boundaries for what is acceptable for you. So I hope you’ve read this and feel a little better informed to make that decision.

_{*Affiliate link – We make a small commission at no cost to you if you purchase via this link.}