NewsNotFound: An AI-Powered, Unbiased News Website

Table of Contents


The media landscape of today is one that is heavily polarised and sensationalised, with many news sources being accused of pushing their own agenda and misleading their readers.

News organisations are no longer viewed as reliable sources of information, and the general public’s trust in traditional media has been steadily declining for years.

The rise of the internet and social media has allowed fake news and misinformation to spread like wildfire, and the need for trustworthy and unbiased news sources has never been greater.

NewsNotFound is a fully transparent news platform.

All news articles are AI-generated based on real-time news data, including the cover image, content, and headline.

All source code is publicly available on GitHub so anyone can see exactly what happens behind the scenes of article creation.

By replacing the human middle-man with our AI-centric article creation process, we can reduce and control the potential for bias and ensure that all coverage remains impartial.

Our mission is to lead the way in AI journalism by providing neutral and unbiased news articles that can be governed by the public. We want to build the most unbiased news platform on the internet.

The main problem we are trying to solve with this project is the monopoly that some of the largest media companies have over the world’s information, with complete freedom (and often incentive) to polarise and sensationalise their news coverage to increase clicks and gain the most attention.

Trust In Traditional Media Is Declining

The decline of trust in traditional media is a worrying trend that has been observed across the world.

In the United Kingdom, only 34% of adults say they trust the news. In the United States, this figure is even lower at 26%.

This lack of trust is compounded by the increase in the number of people avoiding reading the news, with the percentage of people actively avoiding news in the UK increasing from 24% in 2017 to 46% in 2022.

One major issue with the news that has been highlighted in recent years is the presence of false or misleading information.

As of February 2022, in North America, 53% of adults reported seeing false or misleading information about COVID-19 in the news in the last week, while in Europe, this figure was 47%.

These statistics reveal the extent to which the public has been exposed to fake news, which can have serious consequences for public health, safety, and democracy.

Even traditionally trusted sources of news are not immune to public distrust. For example, from 2018 to 2022, the share of people who said they trusted the BBC dropped from 75% to 55%.

The pandemic has only exacerbated the problem of fake news, with 80% of US adults reported seeing fake news about the COVID-19 pandemic. Additionally, while 65% of Americans put their full trust in traditional media outlets, 27% believe mainstream media outlets report fake news.

One factor that contributes to the spread of fake news is the way it is disseminated.

For instance, Facebook posts from misinformation sources receive six times more engagement on average than reputable news sites.

This emphasises the importance of identifying and addressing the sources of fake news, as well as developing methods to promote the spread of accurate information.

How NewsNotFound Works

For a complete technical specification on how our platform works, please see our GitHub repo here.

At the heart of our platform is a fully automated process for generating unbiased news articles. In this section, we will provide a brief breakdown of how our site works, step by step, to give you an idea of what exactly goes into every article we publish.

Step 1: For each category, we begin by scraping the latest headlines from publicly available news sites.

Step 2: Using sentence embedding, we identify any headlines that are related and choose these headlines as the subject of the article. The idea here is that if a story has multiple related headlines across different sources, it is more likely to be of greater importance than a story with none or only one related headline. This helps us determine which story to focus on and write an article about.

Step 3: Once we have a list of related headlines, we then use GPT-3 to generate a list of facts, not opinions, that are present within each of those headline’s respective articles. This step ensures that our articles are based only on factual information, not speculation or opinion.

Step 4: With these facts, we use GPT-3 to generate an article brief that combines all of the available information on the story. We use the prompt below to do this:

“I have this python list of bullet points, where each list item is a different collection of bullet pointed facts about a news story that have been generated by GPT3 from multiple different scraped news articles on the story. Your job is to filter and combine these facts, and generate a very detailed and informative brief for a news reporter based on all of the available facts and information about the news story. Leave out no information. Your only output should be the brief so it can be stored as a string in python.\n\n {bullet_points}”

Step 5: We now ask GPT-4 to generate a whole, unbiased news article based on the brief using the following prompt:

“Based on the below brief, please generate me a whole news article for use on a news company’s website. The article must be completely unbiased and wrote from a neutral point of view. Please make the article as long and detailed as possible, with as much information as you can fit. Do not write any conclusion. Do not make any assumptions or add any suggestive language. Strictly stick to the facts given to you in the brief. Please write in the inverted pyramid format, and lay the content out so that each sentence is its own paragraph. Please also try to ensure no more than 25% of sentences contain more than 20 words:\n\n{brief}”

This prompt ensures that the article is as unbiased as possible and written from a neutral point of view. It also directs GPT-4 to use the inverted pyramid format, ensuring that the most important information is presented first.

Step 6: We then run the article through a “bias checker,” which uses sentiment analysis to rate the sentiment polarity of the generated article. A rating of 0 is extremely negative, 10 is extremely positive, and 5 is completely neutral.

Step 7: If the bias rating is not equal to 5, we pass each paragraph in the article to GPT-3 and ask it to remove any instances of bias and rewrite the paragraph so that it is neutral, using the following prompt:

“You must take on the role of an article reviewer and editor for an unbiased news company. Your job is to look at the below paragraph and ensure that it is 100% neutral and unbiased. If you find any instances of biased words/phrases, or positive/negative language, you must rewrite or reword them in a completely unbiased and neutral way. Aim to be factual, not opinionated. If a paragraph is already neutral and unbiased, just output the original paragraph.\n\n{para}”.

This process is repeated until the article receives a bias rating of 5, or reaches a recursion limit.

NOTE: Step 6 and 7 have been temporarily removed as of 19/04/23. This was implemented when we used GPT-3 for generating the article, however GPT-4 seems to be much more consistent at producing unbiased articles. Using the bias checker with GPT-4 was harming the quality of the article, so it is currently pending improvements.

Step 8: Once we have an unbiased article, we format it so that each sentence is on a new line, improving readability and preventing text from being clumped together in one large paragraph.

Step 9: We now ask GPT-3 to generate a headline for the article using the following prompt:

“You must take on the role of an article reviewer and editor for an unbiased news company. Your job is to look at the below article and generate a 100% neutral and unbiased headline. The headline should use completely neutral and unbiased language. Aim to be factual, not opinionated. Keep the headline as short as possible (ideally no more than 10 words). The headline does not need to be overly descriptive, it just needs to inform the reader quickly on what the subject of the article is. Just output the headline.\n\n{article}”

Step 10: All articles have two main sections: ‘At a glance’ and ‘The details’.

The entire article that has just been generated goes under the ‘The details’ section.

Here we ask GPT-3 to generate five bullet points that best summarise the article to go under the ‘At a glance’ section.

This step ensures that our articles are easily readable and provides readers with a clear understanding of the most important information of the news event without having to dig through paragraphs of text.

Step 11: Finally, we use Stable Diffusion to generate a cover image for the article using the following prompt structure:

“a modern, highly detailed image in the style of an oil painting that best visualises the following news headline: {headline}”.

The reason we specify ‘an oil painting’ is so that users don’t mistake AI generated images for real images, as an oil painting is much more likely to be seen as ‘an artist’s perception’ of the event and not a realistic image.

The above steps are all executed on a schedule for each category, so each day should see a new batch of unbiased articles added to our website. There are much more technical steps in-between the ones outlined above, but we’ve tried to provide a brief overview of how our platform works in a way for anyone to understand.


As of April 2023, all of our source code is available for viewing on GitHub.

This ensures maximum transparency with users of our platform, and allows anyone to contribute to the project and highlight any issues/improvements.

We believe this is the key to a trusted news company – allowing users to see and have control over our technical architecture.

If you have any suggestions for improvements, or if you’ve spotted any bugs, please open an issue on our GitHub repository.


Our platform aims to provide unbiased news to readers, a pressing need in today’s world where trust in the media is at an all-time low. Our system employs artificial intelligence techniques to gather the latest news from various sources, and generate completely neutral and unbiased articles.

By automating the process of news reporting, we eliminate the risk of human bias and error, and provide users with accurate and objective information.

We believe that our platform can help restore trust in the news and promote a more informed and unified society.


Last updated: