Artificial intelligence needs ethics. And so do humans.

A journey on a broken-down bus is a good way to glimpse the ethical limits of generative Artificial Intelligence (AI). Any mechanic who has to repair the vehicle has the opportunity to familiarise themselves in advance with all the components and how they work, which makes it easier to identify the source of any fault. It is out of the question to prevent the use of buses or AI, but for some more or less unfathomable reason, many humans continue to expect to repair malfunctioning virtual assistants using the same logic with which motor vehicle breakdowns are resolved. And that is where the problem can take on global dimensions.

‘In a virtual assistant, the number of features and usage scenarios is always open-ended and much greater than that of a bus. That's why it's more difficult to guarantee the reliability of a chatbot than a bus,’ explains António Branco, professor at the Faculty of Sciences of the University of Lisbon (Ciências ULisboa) and coordinator of the NLX Natural Language and Speech Group, which has been developing the virtual assistant (or chatbot, or bot) Evaristo.ai.

Despite discrepancies in reliability, the number of chatbot users has continued to grow in recent times – and it is not surprising that, taking into account the tools available on mobile phones, it has already surpassed the total number of people who travel by bus. This raises another question: who can guarantee that the AI available in the world is not being used to generate malicious bots?

“This whole process started badly, when AI was allowed to learn all the good things and... also all the bad things that appear on social media and the Internet,” replies Luís Correia, professor of Science at ULisboa and researcher at the LASIGE centre.

After millennia as a human exclusive, ethical boundaries gave rise to the first large-scale application of machines that, ideally, should have the autonomy to discern right from wrong at any given moment. And with good reason. In transport, telecommunications, medicine, meteorology, energy and even agriculture, there is no shortage of examples of AI tools that process information, make predictions, identify patterns and perform actions.

While humans have become accustomed to religious or philosophical commandments that condemn those who kill, steal or simply eat forbidden foods, machines use algorithms in a logic that IT departments refer to as ‘if (...), then (...)’, which allows them to list several possible options whenever a particular scenario arises before the AI. It is with this logic that any automaton or AI platform gains the ability to decide whether to stop a runaway train to avoid a tragedy, choose a job candidate without racial or religious prejudice, or simply close the doors of a building if it detects an intrusion.

These are just the most predictable scenarios. Then there are all the other situations that no human can foresee, such as the recent news about the bot that recommended a user hire a professional assassin to kill her husband. Or the young man sentenced to nine years in prison after lengthy conversations with a bot and breaking into Windsor Castle with the intention of assassinating Queen Elizabeth II. Internet searches also reveal many more cases of suicide allegedly facilitated by chatbots than the big tech brands are willing to acknowledge. It is known that it is in families of ‘generalist’ AI systems that there is the greatest potential for errors, biases, manipulation, or ‘hallucinations’ that lead to unrealistic responses and descriptions, due to the fact that they deal with almost all topics addressed by humanity – but this does not mean that the safeguards applied to specialised systems that eventually deal with critical functions can be abandoned.

‘The application of ethical filters is desirable, but the most widely used systems today have been made available by major commercial brands, and there is still some opacity in what these systems can do,’ says Luís Correia. ‘It is not always clear how these virtual assistants work, but, on the other hand, we know that they can be manipulated,’ adds the professor from the Department of Informatics.

Even Elon Musk, the richest man in the world, who is known for defending the technological empire with fierce conviction, has come forward to admit that bots are manipulable. It happened in the summer of 2025, when a bot operating with Grok technology from social network X began publicly praising Adolf Hitler and the ideal criminals of the Nazi party.

Once again, it would be illusory to think that fixing a simple bot is enough to bring about a peaceful spring across the entire AI landscape. Perhaps because bots learn according to what they are given to process, Luís Correia and António Branco admit that, at first glance, the solution may lie in corrective measures taken by humans.

"The large language models that have been used by AI are trained with large repositories of text and data and follow reinforcement learning logic, which involves humans classifying the different responses to different questions or problems as appropriate or inappropriate. Once classified, these responses are introduced into the language used during interactions with other humans. This is how chatbots are conditioned from an ethical point of view," says António Branco.

In a perfect virtual world, nothing else would be needed besides human corrections. But in reality, things are different. ‘Since all these systems are based on probabilistic neural networks rather than deterministic processes, there is always the possibility that AI will do things that are not intended,’ adds the NLX coordinator.

Due to media coverage, it may be tempting to think that generative AI is the only methodology that exists in AI, but those who study the subject point out that there are other tools and concepts available in computer science laboratories. “Learning from data is a well-known area that has advanced AI, but there is still some difficulty in linking this type of learning with symbolic reasoning, which allows (an AI system) to make inferences and abstractions from the data it analyses,” emphasises Luís Correia.

Given the current state of technology, the expectation that AI will surpass humans is clearly exaggerated from the outset, when one understands that, without symbolic reasoning or another learning method, AI will continue to rely largely on mere probability logic to formulate responses based on the data and patterns it encounters. It has great potential to enable chatbots to imitate best practices—but it also opens the door to the imitation of undesirable actions. And that is why some people are already taking action.

“OpenAI hired experts to ‘detoxify’ the AI's responses. I believe there are more brands that have felt the need to do this work because they know that AI learns everything on its own without control and can end up inflating messages of racism, hate speech, misogyny, and many other things found on the Internet,” adds Luís Correia.

Although they can minimize the harm, “detoxification” processes are still time-consuming and expensive—and so they have begun to attract experts who test chatbots and check if they say things they shouldn't. It turns out that this same strategy can also be used to achieve the opposite results. Just like the tricks humans use when they want to influence, deceive, or cheat someone.

‘A chatbot can be programmed not to say or explain certain things that are considered dangerous for humans, but it is also known that humans have the ability to take certain conversations to certain “places” that can lead a chatbot to say what it shouldn't,’ Luís Correia points out.

Some theoretical approaches allow for the use of supervisory bots that monitor potential mistakes made by others, but Luís Correia points out that this solution will only be viable if the reliability of these supervisory tools is guaranteed. Which, in a way, also implies unpredictable results.

As is easy to imagine, the expectation of having AI that is better than humans is quickly limited once it is realised that current AI imitates humans. Not only because ethics and moral values can follow different logics, restrictions and objectives, depending on the geography or cultural history of end users. Result: it is possible to have mitigations and corrections, but it is reasonable to admit that any project that has the ambition to convert all bots to the “good” side is likely to have a degree of difficulty comparable to that of any project that aims to cleanse humanity of all evil. It will not be a lost “war”, but it tends to remain an open-ended challenge that requires ongoing monitoring and corrections.

While humans have behavioural and linguistic cues that help detect lies or manipulation, bots tend to give very convincing answers, regardless of whether they are correct or whether they are “hallucinating”, leading them to invent data and facts, recalls Luís Correia. In such cases, only a critical mind can save flesh-and-blood interlocutors from greater evils – whether by asking more questions and seeking alternative sources of information or by consulting the technical data sheet that indicates the ethical, technical and thematic limits of each virtual agent.

‘The more capabilities bots have, the greater the tendency for humans to treat them as people. When we hire someone, we conduct interviews, and then we put those people to the test, we see how they adapt to work, and through this whole process we come to know what we can ask them to do with confidence. When we ask AI something, we must take equivalent care,’ concludes Luís Correia.