Your AI Summary is biased

Since the emergence of ChatGPT in 2020, hundreds if not thousands of applications have transformed our lives and boosted our everyday performance. One of these game-changers is text analysis and summarization.

With text summarization, we can skip the filler words and information that don't contribute much to our understanding of the text we want to read but don't have time for.

Here is where AI comes into play. It can read the text for us and create a nice summary, allowing us to focus on what's really important.

For example with a prompt like this (highly simplified):

`Please summarize this text in 200 words`

Or maybe even better to tell GPT to focus on some parts:

`Please summarize this text in 200 words and focus on xyz`

This might work, and you you get a good summary if the model has good language understanding. We can check these capabilities anytime in various benchmarks, e.g., in the Chatbot Arena.

This all can work, but there is a catch:

GPT, Claude, and all models are biased all the time. That is, they imitate a certain character (which isn't very clear to us) and based on that, they give us whatever information fits that character. I know this dives deep into the psychology of these models, but bear with me.

In order to simulate that, I conducted an experiment with MagicMarker.ai, a tool that lets the AI highlight the important parts of a text (from the subjective perspective of the LLM).

Fortunately, we can set the persona of most LLMs (not o1 as of today) using the 'act as if you' technique via the system prompt, for example by telling the LLM which personas to imitate during our conversation.

And this works incredibly well.

Let's take an example:

Here is our test scenario: an article and three different personas.

Original Article

City X has recently launched a comprehensive new public transportation system designed to address growing traffic congestion and environmental concerns. The initiative includes the introduction of electric buses, expansion of subway lines, and the addition of numerous bike-sharing stations throughout the city. Officials believe that the new system will not only alleviate traffic but also reduce greenhouse gas emissions, promote economic growth by improving access to businesses, and enhance the quality of life for residents by providing affordable and reliable transportation options. The project has been funded through a combination of municipal bonds and federal grants, reflecting the city's commitment to sustainable development. Additionally, community feedback played a significant role in planning the routes and services to ensure they meet the diverse needs of City X's population.

Now, if you give the AI this text to summarize, you might get something like this—the focus on certain aspects is random and will depend highly on the current unknown "personas" the LLM has and how many words it can use for the summary.

Now With personas (System Prompt)

We will use the same article to summarise, but now with three different personas and see how this affects the summary.

Descriptions of the Three Personas

Person A (Environmental Perspective):

Person A is deeply passionate about environmental conservation and sustainability. They prioritize initiatives that protect natural resources, reduce pollution, and promote renewable energy. Person A is attentive to the long-term ecological impact of policies and actions, advocating for measures that ensure a healthier planet for future generations. Their decisions are often guided by a commitment to mitigating climate change and preserving biodiversity.

Person B (Economic Perspective):

Person B is focused on economic growth and financial stability. They emphasize the importance of initiatives that boost the local economy, create jobs, and enhance business opportunities. Person B values efficiency, profitability, and the effective allocation of resources. They are keen on projects that improve infrastructure, attract investments, and support entrepreneurial ventures, believing that a strong economy leads to overall societal progress.

Person C (Social Perspective):

Person C is dedicated to improving the quality of life for individuals and communities. They prioritize initiatives that promote social equity, accessibility, and community well-being. Person C values inclusive policies that address the diverse needs of the population, ensuring that services are affordable and reliable for everyone. They are committed to fostering a connected and satisfied community by incorporating feedback and addressing social challenges such as education, healthcare, and public safety.

In order to see the focus areas we usually don't see (imagine like the brain areas that start to activate when you see something you have an emotional connection to), we can simply upload the document (in the image below on the left side) and copy-paste the character description above into the "act as if" input.

Then, enter the following text in the user prompt:

`Highlight only the important parts in this text that matter to me.`

Results

I did this for the first persona above, and below you can see the highlights followed by a summary based on the highlights.

Persona 1

Note the focus on environmental aspects in the highlights and summary.

Note the focus on environmental aspects in the highlights and summary.

I suspect that if we increase the maximum token size (length of the summary) more aspects will emerge, but the more focus and the shorter the article, the more we will see the focus areas.

Person B

Again, here for the second persona, you see the focus is on the economic and business side of the article.

And here for Persona C.

As you can see, setting the focus using the "act-as-if" technique can make a huge difference, especially for short summaries (the longer the summary, the less sense it makes to make it in the first place :)).

You can leverage this technique on any important summary or in general any text understanding tasks you do with LLMs, for contracts (taking the role of, say, the supplier, the employee, etc.).

Don't let the LLM choose based on random factors.

Happy summarizing! :)