Tricking AI - especially if it has been built on elaborate personalisation and response algorithms to behave like an organic being is difficult because one might need the flaws in the models themselves. The general line of thinking is that horny AI can be easily tricked if you manage to circumvent the mechanisms through which these AIs parse and act upon whatever it is a user sends their way.
A method by which AI could be outsmarted is through the tampering of data that come in. For instance, we know that AI systems can be easily hacked by changing inputs just a little bit. According to a study by the Massachusetts Institute of Technology, for example, image recognition systems not infrequently erred in diagnosing objects just because several pixels had been changed. Likewise, chatbots - both personal and general object-led bots - can be prone to trickery with the application of phrases or cadence speech they have not been trained on how to properly interpret.
Calculating the extent to which an AI implements can be tricked is more of a problem, as this goes with just how advanced experiment for any kind of certain application. Basic models can be deceived in as many as 60% of cases with the use of intentionally designed inputs, while more sophisticated systems that leverage deep learning and natural language processing (NLP) may only be lead astray between 10-20% of the time [7]. The fact that AI architecture plays a role in determining whether an attack is successful underlines how one can also be tricked.]
Fooling AI, strictly in the context of trying to trick machines by inserting specific information, is a practice considering how different organizations have been working towards explaining AI and developing better systems. An incident occurred in 2017, Google Translate had a clear series of nonsensical nonsense outputs when inputs from certain languages this was solved almost immediately and the issue exploited to increased system accuracy...
This idea that AI can be duped gets into some pretty philosophical ideas about machine comprehension and learning as well. What Alan Turing, one of the fathers of computer science claimed is as follows; - We can only see a little way ahead, but we can see plenty there that needs to be done. This quote emphasized the never-ending evolution of AI capacities and reliability.
How successful attempts to deceive horny AI models are largely depends on the training data the model has experienced, as well as how robust a learning algorithm is. Current AI Systems are built on vast datasets and learning mechanisms targeted to recognize (and adjust with) a large set of inputs As a result most of them carry out in-depth reverse-engineering to be able to cheat these systems and not just anyone with technical skill can fool an AI as you must understand rightly how it was designed.
The question of tricking AI further complicates ethical discussions. While it shows the system failing, there are also potential negative side effects that can be seen across various sensitive applications. So then again the question remains to be, even if it is possible that one can fool AIs under specific circumstances, should we rather prioritize deceiving them over making them better performing and ethically deployable?
To learn more, you can checkout the concept of horny AI and how these interactions are crafted and tested. The fine line between improving AI and acknowledging its weaknesses remains as crucial a debate in the discourse around technology and society.