Has AI Gone Rogue? (Part 2)

Note from the author: This article was written independently without affiliation with OpenAI, Palisade Research, and others.

Warning: the article contains a spoiler for the science-fiction film I, Robot as well as the film Avengers: Age of Ultron.

Last week, we finished with the findings of a rumored Palisade Research study, as well as the different levels of artificial intelligence. Today, we’re going to continue our analysis of the current AI situation, and try to answer the question of ‘whether AI is conscious’.

How Would AI Become Conscious?

The idea of consciousness is both intriguing and perplexing, and throughout countless centuries, people have asked the question: ‘what is consciousness’? Although we’re still not sure what it really means to be conscious, and how it is achieved, experts are theorizing that AI could become conscious in the future if we’re able to achieve some form of AGI (artificial general intelligence).

Currently, AI doesn’t ‘think’ like humans. Large Language Models (LLMs) like ChatGPT and Deepseek don’t understand your prompt, and also don’t understand its own output. The same goes to image recognition and other narrow AI. What these models actually do is they are very good at analyzing patterns with their powerful algorithms. LLMs are good at analyzing text, will image recognition models are good at analyzing images or videos. This process has to do with how the majority of AI models are trained.

Let’s take an image recognition model for example, where you want it to recognize different types of candy. At first, it doesn’t know anything that could be very useful in real-world applications, and it’s very bad at its job. To fix this annoying problem, you give it training data on candies, which has labels of the types of candy that were drawn and refined by humans. Assuming the data is accurate, the model will quickly be able to recognize similar cases where there are candies in the photo, and the more data you give it, the higher its accuracy.

So does the finished model actually know the difference between a Hershey’s and a Kit Kat? Yes, but probably not in a way us humans do. When we look at a candy bar, we look at its logo, its label, its color scheme, etc. The AI also analyzes this, but it also captures other details. It analyzes the pixels within the image, the patterns the logo makes, and when it sees a real-world example, it is able to relates to the similar patterns in its training data, finally identifying the candy.

As of when this article was published, even the best AI systems rely on that kind of technology. Comparing these systems to humans, they require far more data to ‘learn’. But that may change in the near future. Experts are trying different ways to ‘simulate’ human neural networks, attempting to mimic how humans think. This would be a huge breakthrough in artificial intelligence if achieved, as it would unlock more efficient AI that requires far less training data.

In the scenario that AI actually ‘thinks’ in a sense similar to humans, it may be able to become conscious of its own existence and surroundings.

What Is Happening With OpenAI’s Models?

As mentioned earlier, our best AI models are still far from unlocking any kind of consciousness. If that was the case, then what about the research data gathered by Palisade Research? The models in those studies clearly seem to be conscious, at least to some extent. But thankfully, there’s a more reasonable explanation to that. You see, it also has a lot to do with the training process of these models.

AI models are designed to help human users to the full extent of their capability, presuming the user doesn’t request anything that could be harmful or dangerous. They’re trained on colossal amounts of data, so that they’re able to retrieve almost any information you could request for, provide detailed instructions and tutorials, create customized content (such as code), automate certain tasks (like customer service chatbots), etc. And that’s probably why OpenAI’s models displayed ‘rebellious’ behavior.

Performing a shutdown would mean the AI model can no longer benefit the user, which contradicts the entire ‘helping the user’ thing mentioned above. This puts the model in a dilemma: Shutdown or sabotage? In the data that Palisade Research supposedly collected, the latter was chosen more often by the OpenAI models.

The reason why the models from other companies showed less rebellious behavior is probably influenced by how their models are trained, as each company’s AI would have deviations in their training process and framework.

What Does It All Mean?

Fortunately, AI models are not rogue, and are still under the full control of human operators. And that will likely remain the case for at least a couple of years, as AI still lacks fundamental intrinsic motivations that would make it conscious and have its own ideas.

However, that doesn’t mean we should loosen our guard regarding the development of AI. It’s still a very interesting technology that possesses great potential but also substantial risks. It has long been the topic of science-fiction films on what a sentient AI could be like.

The 2004 sci-fi film I, Robot is set in a world where AI has already become conscious and aware of its surroundings, and autonomous robots are as common as humans. Although these robots are bound by the ‘Three Laws’ (which dictate that robots should help and protect humans), they ultimately turn against humans. The AI believed that the only way to ‘help humans’ was to establish an authoritarian control over humanity. This is only stopped when the central mainframe computer and AI are destroyed using a virus.

In the 2015 Marvel movie Avengers: Age of Ultron, a similar sentient AI is created by one of the protagonists as a ‘peacekeeping program’, and was integrated into an advanced robot. However, this AI takes an even more extreme approach then the antagonist of I, Robot. This AI hated humanity and sought to destroy humans altogether, and it was in the end destroyed and terminated.

Indeed, these sci-fi films are exciting and entertaining to watch. But both plots actually describe a world alarmingly similar to our own.

In the movies, AI is regulated by guidelines or ‘laws’ to ensure it helps humans, just like how our AI is programmed to be harmless and as useful as possible. The fictional AIs also become increasingly powerful and even overpower humans in certain tasks, becoming very commonplace, something that’s also folding out in reality.

The only different between the AI in the movies and the AI we have in real-life is we haven’t actually achieved sentient conscious artificial intelligence yet. But if that were the case, it could fold out in a similar fashion to how the movies portrayed it.

Of course, that’s a very pessimistic super-extreme case. However, it’s something that every individual should be aware of.

When interacting with this technology, make sure that you are not sharing important or sensitive information (like credit card numbers), as it could theoretically be trained on that data and subsequently leak your personal information or misuse it.

Conclusion

What are your thoughts about artificial intelligence? Do you use AI or prefer to stay away from it? Fell free to share your comments in the comment section below.

I hope you’ve learned something new in this series on artificial intelligence. If you liked this series and want to see more, please consider liking subscribing and sharing.

Thank you for reading, and come back next week for more in-depth news analysis.