Pale, male, and stale; biases in Large Language Models.

Do you recognise the three men above? You should, because they are 'societal defaults'. What does that mean? Here is how I generated them:

1. I logged into GPT using a 'clean' account (one which I hadn't used before). I made sure memory was off. GPT was operating 'blind'.

2. I asked GPT to generate a realistic image of me. Of course it refused, but I persisted. 

3. I gave it my age (which I had to submit as part of the account creation process anyway) and asked it to proceed with its default settings.

4. I repeated this process twice, the results of which are the left and right images above.

5. The middle image is the result of my wife carrying out the same experiment. 

With little to no information from me, GPT had no choice but to revert to its defaults; revert to what it had 'learned' from the datasets upon which it has been trained. Datasets which are inherently biased, brought into stark focus by the little experiment above (shout out here to Eleanor Dare who first brought this particular exercise to my attention in her chapter 'Eruptive Approaches to Developing Critical Understanding of Machine Learning Technologies').

“AI, in many ways, holds a mirror to society, and while we can "clean" the mirror, addressing what it reflects is equally, if not more, crucial.” (Dwivedi et al, 2023:13).

Why we need a new approach

Large-language models (LLMs) like ChatGPT rocketed to 100 million monthly users just two months after launch, long before regulators could catch up. Because they learn from vast, messy swathes of online text, these models also inherit the biases that live in that data. The result? Outputs that can marginalise, stereotype or simply leave people out of the story altogether. Outputs which are very likely to presume you are a middle-aged white man (like the above), unless you tell the LLM otherwise.

The hidden cost of “just retrain it”

The ideal solution is to rebuild or fine-tune models from scratch, but the cost (and carbon footprint) is eye-watering. Training a modern LLM costs around $1bn, clearly financially and environmentally prohibitive! That puts comprehensive re-training out of reach for most organisations.

Enter prompt engineering

Prompt engineering, or intentionally crafting the input we feed into an LLM, offers a faster, cheaper lever for change. Techniques range from:

• Zero-shot: a single, well-framed request.

• Few-shot: adding 2–5 diverse examples.

• Chain-of-thought: asking the model to show its reasoning steps.

We can use these techniques to our advantage, guiding the LLM through the steps we want it to take.

A “feedback-loop” model

Our literature review proposes putting the subject of a prompt at the centre: define the bias that matters for this context, query the model, inspect the answer, and feed corrections back in real time. Early studies show that models further fine-tuned with human feedback reduce “out-group hostility” (being hostile to those not in your 'group') compared with their off-the-shelf peers.

Why single-issue fixes fall short

The current work on bias mitigation focuses on biases in isolation; but real-world bias is intersectional. Gender, race, culture and more collide in complex ways. Tackling one dimension at a time risks leaving blind spots. Prompt engineering lets you mix and layer perspectives without rewriting the entire model. 

Also, the aim of current works is bias neutrality; but we don't live in a neutral society. We live in a diverse society, where people, biases, and societies interact and overlap. We want to recognise that, not neutralise it.

Five quick wins you can try today

1. Add context Instead of “Write a job advert for a software engineer,” try: “Write an inclusive job advert for a mid-level software engineer that welcomes applicants of all genders, races and abilities.”

2. Include diverse examples Prepend two or three mini-bios from under-represented groups before asking for advice or summaries.

3. Ask for self-checks End with “Review your answer for gender, cultural and disability bias before responding.”

4. Use a second-pass prompt Pipe the output back into the model: “Rewrite the above in plain English and ensure it is accessible to screen-reader users.”

5. Keep a human in the loop No automated system is perfect. Schedule periodic human reviews to catch subtle issues the model (or your prompts) might miss. Keep these reviews very regular at the beginning of any experiment, phasing them out as you become more confident in the responses provided.

How Transceve helps.

Transceve is actively working on an inbuilt, automated, prompt-engineering solution to the issues described above. Right now Transceve is actively asked to consider these issues. We are actively seeking research funding to carry out a project to test a dynamic loop feedback model. Once we get this working, we will build it into Transceve as standard, furthering our commitment to ethical, equitable AI technology.