OpenAI partners with NewsCorp.
What does this mean for users and democracy?
On May 22, 2024, OpenAI announced a partnership with NewsCorp. There are significant risks here to consider. We have also seen the launch of Google’s Gemini-powered search; how does this relate to the OpenAI news of the week? Let’s break down the implications into three aspects: Training AI, Representing Knowledge, and Power.
Resources:
1) Training AI
Generative AI models require vast amounts of data for training. Recently, there’s been considerable backlash from various industries regarding the use of proprietary data. For instance:
- Drake and The Weeknd had AI-generated fake content pulled from Spotify.
- Authors filed class actions against OpenAI and Meta.
- Scarlett Johansson making public complaints against OpenAI for using a voice similar to hers in ChatGPT4, reminiscent of the AI in Spike Jonze’s Her.
In response to these challenges, OpenAI’s deal with NewsCorp aims to secure data to train and enhance ChatGPT and related products. Unlike Meta, which may utilise data from its subsidiaries like Facebook and Instagram, OpenAI lacks such extensive in-house data resources, making this partnership a strategic move.
However, training a GenAI model on data heavily drawn from one source can embed specific values, ideologies, and biases into the AI. NewsCorp (i.e. Fox News, Sky News, New York Post, Foxtel, and many more) is known for its right-leaning perspective, raising concerns about the potential for biased outputs.
When you train a model, it picks up the embedded values, morals, worldviews, ideologies, and cultural traits that are prominent in the training data. Even after fine-tuning and placing guardrails around the outputs, subtle and covert bias will still frequently be present in outputs. If we think about the data that News Corp creates, which will be used to train newer versions of ChatGPT, what embedded values, political views, and cultural biases might be in that data?
We know that most News Corp subsidiaries are often right-leaning (to varying degrees), reflecting the right-leaning proclivities of the media empire’s owners. Numerous people have pointed this out, but if you want a quick dive, you can check out this Australian Parliamentary document on diversity in Australian media. This report also highlights the problem of media monopolies and potential for abuses of power.
Whether you are left or right-leaning, and how much, is beside the point here. What is crucial is how these partnerships can impact outcomes.
If you train a model on data that leans generally one way or another, the trained weights in the model will reflect bias in the generated outputs.
Consider a possible future scenario where we have OpenAI trained mostly on the right-leaning side of western ideologies and a competitor model trained on left-leaning views, you will end up with behemoth and skewed models that could lead to a strengthening of left versus right polarisation. This is not good for anyone and certainly not good for healthy democratic societies.
2) Representing knowledge.
So, we can see how we might end up with users of a right-trained model seeing more right-leaning views and reverse for left-trained models. Not a great situation in the isolation of using an individual user interacting with a GenAI model, but something bigger is at stake. We are facing two big shifts in the next 12–18 months. A change to search engines and (a bit more slowly probably) the development of personalised AI agents.
For this post, let’s focus onthe search engine changes. Google announced last week at its IO 2024 that their model Gemini will power their search engines going forward. At first, this will just be in the US, and they expect it to roll out more widely next year. We then have a new situation in that people looking for information are now not only directed by search engine rankings, SE optimisations, paid rankings, and potentially skewed resources like Wikipedia. With this significant change, we are heading to a situation where people might not even make it to the actual websites. They will just rely on the top-level Gemini AI-generated text the model provides you.
With the imminent changes in search engine technology, the embedded biases in the training data behind the models could significantly impact how users receive information.
If Microsoft, a major backer of OpenAI, leverages this partnership for a similar search engine strategy, we could see a Microsoft/OpenAI search engine influenced by being trained on NewsCorp’s right-leaning data. If we see models that may be more left-leaning, the scenario could create a polarised digital landscape, with AI-generated content reinforcing existing silos of left and right ideologies.
3) Power.
This partnership exemplifies the convergence of power and money, human society's age-old mover and shaker. Unfortunately, OpenAI and NewsCorp, two immensely influential entities, are joining forces, with Microsoft (the largest company in the world with a Market Cap of USD$2.89T) playing a significant supporting role.
Behind power and money are individuals. The Murdoch family’s political influence is well-documented, from advising Trump when he was in office to lobbying UK leaders like Thatcher, Blair, Cameron, and Johnson. How might a powerful media company — one that seeks to align nation-state politics with their profit interests — try to influence their AI partners?
Integrating such power into AI models could herald a new era of subtle persuasion, far surpassing the influence exerted by entities like Cambridge Analytica. The implications for democracy and societal balance are profound, necessitating critically examining these developments.
Why is this news important?
The OpenAI-NewsCorp partnership is more than just a business deal; it marks a pivotal moment in the intersection of AI, media, and power. As generative AI becomes increasingly integrated into our daily lives, the sources of its training data take on new significance.
Do you think News Corp might see this partnership as a new opportunity to influence voters?
This alliance could profoundly shape the information landscape, embedding subtle biases and potentially deepening societal divisions. It’s crucial for all stakeholders — developers, policymakers, and the public — to scrutinize these partnerships and their broader implications. As we navigate this new era, ensuring transparency and balance in AI development will be key to maintaining a healthy, democratic society.
