There’s no guarantee that any AI agent you create will do exactly what you want it to. All depends on how well you develop reward functions in the process of its training. What that means? Well, basically - you need to be crystal clear with your instructions because AI can misinterpret and choose a wrong trajectory.

Automate, we can automate so much with AI. But there is a hidden risk of eliminating human touch completely from the areas that should never lose humanity. Even more - specialists that delegate too many tasks to AI assistants start noticing that their skills deteriorate and they lose their touch in areas they were previously experts in.

Your algorithm may be perfect and the model may be trained for different situations. And yet, users won’t be users if they don’t poke at the things they're not supposed to poke, just to see what would happen. Unprepared for this exact behaviour, AI agents start malfunctioning.

But isn’t AI supposed to be unbiased? Well, it depends on how it is programmed. More often than not AI picks up its creators’ implicit biases, or, if it can learn from interactions with users - from those too. Worse still, AI can not only repeat those biases but amplify them, bringing discrimination to a whole new level.

It’s actually quite easy to attack something that learns and transforms through interaction with people. Attackers use adversarial inputs or prompt injections to manipulate AI. And the consequences could be quite serious if you entrust important matters to your AI.

What you envision your AI agent doing and what people use it for are two completely different things. Since you release it into the world you should be ready for users to find some unconventional applications. Real problems come when it becomes utilised for real malicious purposes - like scam, fakes or malware generation.

It really can get out of control. Not a film-like out of control where evil AI conquers the world, but the kind where everyone shrugs and says they’ll fix it, and in the end everyone just gets used to it. But of course missing the perfect time for scaling is never an option.

Like all things that have an important place in our lives, AI shapes the world around us and changes social norms. Sometimes in inconvenient and even potentially harmful ways. The more it gets involved in our lives, the faster it happens. And we might fail to notice it before it’s too late.

Every good AI model needs high quality data to learn on, and a lot of it too. But it might be flawed or even purposefully changed to not allow AI training on it. The bigger the amount of defective data, the faster your AI’s behaviour deteriorates as it starts to see the faulty examples as a standard.

That is a real risk, just not in the way most people moan and groan on the Net about it. Unchecked processes are far more dangerous than many can imagine, especially when in the rush to get the most profit companies simply stop caring about individuals, putting them under more pressure.