No more is artificial intelligence (AI) a far-off dream of science fiction, it is the current reality which has transformed the manner of shopping and file diagnosis in hospitals. However there is an ugly reality that surrounds the hype and hurry of using AI it is not as error free, impartial and reliable as we tend to believe. As a matter of fact, AI systems should never be trusted blindly, causing major mistakes, ethical dilemmas, as well as unintended outcomes. This is why even the most advanced AI must be treated by certain amount of skepticism.

AI Mirrors the Biases of Its Creators

One of the simplest explanations as to why you can’t always trust AI is that it mirrors the data — and biases — it was trained on. AI systems learn to identify patterns in huge sets of data, but if the data contains old inequalities, skewed demographics, or cultural prejudice, the AI will naturally learn and replicate them. This issue has appeared in mundane situations with alarming repercussions.

Consider facial recognition software, for example. Third-party studies by MIT Media Lab and others found that several of the most popular facial recognition systems correctly identified dark-skinned women with a 34.7% error rate, and correctly identified white men with less than 1%. Why? Because the training sets were predominantly lighter-skinned individuals. The outcome? Misidentifications that have led to false arrests and discriminatory surveillance practices.

This isn’t a limitation of technology — it’s social, embedded in code. And it reveals an ugly truth: AI does not operate in isolation. It inherits the flaws of humans and multiplies by them.

AI is Lacking in Human Context and Common Sense

Although capable of processing and analyzing huge amounts of data more quickly than a human, AI lacks one essential quality: context. AI systems, however advanced, do not “know” the world like humans do. They recognize patterns based on statistical relationships, not significance. The danger of this issue is especially high when AI is applied to decision making in such sensitive fields as medicine, law or education.

Transparency is Not Usually Present

Another big issue with AI is its “black box” character. Much of AI, especially deep learning AI, is so complex that even the creators don’t know how it comes to certain conclusions. That lack of transparency is a big concern in use cases involving accountability.

Suppose an AI is working within a hospital to recommend treatment plans. When a doctor takes an AI recommendation that leads to a bad result, who is responsible? The AI? The doctor? The designer? When even the designers of the system can’t demonstrate the reasoning process, audit or question decisions become virtually impossible.

This opacity has also been leveraged in domains such as finance. AI-based credit scoring is now widespread, but rejected applicants have no means of knowing or challenging the basis of the decision. What ensues is a highly asymmetrical power structure, with institutions using black box tools to make life-changing decisions with little accountability.

AI is Easily Exploitable

AI can be manipulated too. So-called “adversarial attacks” involve input data being subtly altered in ways a human wouldn’t even detect — but which can completely mislead an AI. One such example comes from computer vision science, where researchers proved they could manipulate a few pixels in a photograph so that an AI would be tricked into identifying a photo of a turtle as a rifle.

These vulnerabilities extend beyond the academic realm. In computer security, adversarial methods can be used as a means to defeat spam filters or faces detection systems. Bots, based on AI, are already deployed in social media to promote misinformation, influence the visibility algorithm, and affect the opinion of the general public. AI, when utilized nefariously, is a good deception tool.

Conclusion

One of the most revolutionary technologies of the modern world is artificial intelligence and the effects are obvious and beneficial. Not only is it unwise to believe in AI blindly, but it is also hazardous. To make the best out of it, we have to know the limits, create transparency, and never remove the figurative hands off of the steering wheel.