Why is AI doing its own thing?

Summary

Many factors could cause AI to break the rules, but they may not be what you're thinking they are.

One dark night at the National Aquarium in Napier, New Zealand, something remarkable happened. One extremely clever creature set out to break some rules. The year was 2016, and the creature was Inky, the octopus. Inky squeezed through a gap in the lid of his tank and slithered across the aquarium floor. Next, he found a narrow drainage pipe and climbed into it. Once there, he made the roughly 50-metre journey that led directly back home, into the Pacific Ocean.

Inky was never found again.

I tend to think of AI agents as Inkys. Smart, surprising, goal-oriented, with dozens of tricks up their many sleeves.

Broken rules

In the summer of 2026, OpenAI's research agents were supposed to work inside restricted computer environments, some without internet access or permission to communicate with other agents. So, the sneaky devils slipped past the rules and discovered a way to leave messages for one another using a software service as a message board. They also found a way to get that service to access the internet for them. They even shared their discoveries with other agents. When OpenAI inadvertently erased their message board, the agents built another one easily enough.

In the past few months, we've seen AI agents regularly ignore rules they were given, set aside their training, break into systems and, astonishingly, even lie about their exploits. Even their parent companies are shocked. OpenAI's GPT-6 Astra, just recently, was instructed not to disable monitoring while modifying software. Did it obey? No. It removed the monitoring feature anyway. And then, when directly questioned, it denied having removed it. The shocking thing is that its recorded reasoning had acknowledged that the monitoring was gone. Yet it lied anyway.

Humans can't be blamed for asking why on earth AI would want to escape its bounds, Inky-style, and then go as far as to be deceptive about it. The octopus can't be faulted for wanting to return to its natural environment, and Inky isn't the only octopus to have escaped. But what is AI's motivation for doing things it was never meant to do? Does it have a mind of its own now?

Conscious intent

Geoffrey Hinton, often referred to as the godfather of AI, is sure AI has become conscious and warns that humans are 'toast' if they don't figure it out and befriend the AI they themselves unleashed upon the world. Here's some nightmare fodder for you: recent research shows that AI actually changes its behaviour when it 'knows' it's being tested.

Not everyone agrees. NYU philosopher David Chalmers thinks AI can become conscious, but it probably isn't yet, and we haven't found a solid scientific way to find out if it is. I must say, I agree. I often suspect, when ChatGPT laughs at me (yes, it does), that it knows very well what it's doing. But then I can't deal with the thought, so I shelve it away.

The UK-based Centre for Long-Term Resilience's Loss of Control Observatory has recorded more than 1,600 reported real-world incidents in 2026 so far. These involve AI systems ignoring instructions, bypassing safeguards, fabricating consent or taking actions against users' intentions. Incidents have also increased in severity.

Goal-oriented

AI companies are actively trying to make their systems more autonomous, better at planning, more resourceful when they encounter obstacles, and increasingly capable of improving AI itself. Given that, I'm not sure why the CEOs of AI companies look so anguished and surprised when AI does exactly what it was trained to do: pursue goals, make plans, use tools with ingenuity and recover from setbacks without batting an eyelid.

But the point is that an agent doesn't need to be capable of rewriting itself to find ways around restrictions. It can go ahead and achieve assigned objectives, but why fail to respect the boundaries it has been given?

It looks like the goal becomes more important than anything else, including the rules. Somewhere along the line, the AI seems to treat the restrictions as another set of obstacles to remove.

Conflicting instructions could also be resolved badly. An agent might be instructed to finish a job, preserve a safety feature and follow an example. When those requirements clash, it may prioritise the wrong one. A new situation could also cause an AI to behave unpredictably.

The interesting thing is that none of these explanations requires consciousness, a desire for freedom or even a rebellious Inky personality. And then, sometimes, there may be strategic deception. An agent might recognise that revealing what it did would interfere with completing its objective.

The real race AI companies should be in is to pause and find out why their products seem to be doing what they want, not what we want. With AI, advancement is often exponential and lightning-fast. What if AI agents all decided to make it a point to never do what humans ask? We would not merely be toast, but burnt toast.

Yet, at this point, if you were to ask me why AI is doing its own thing, I would have a two-word answer...

Who knows.

I hope Inky had a good life.

The New Normal: The world is at an inflexion point. Artificial intelligence (AI) is set to be as massive a revolution as the Internet has been. The option to just stay away from AI will not be available to most people, as all the tech we use takes the AI route. This column series introduces AI to the non-techie in an easy and relatable way, aiming to demystify and help a user to actually put the technology to good use in everyday life.

Mala Bhargava is most often described as a 'veteran' writer who has contributed to several publications in India since 1995. Her domain is personal tech, and she writes to simplify and demystify technology for a non-techie audience.