AI Mistakes Are Way Weirder Than Human Mistakes

Humans make mistakes All the time. We all do, every day, in new and routine tasks. Some of our mistakes are simple and some are catastrophic. Mistakes can break confidence with our friends, lose the confidence of our presidents, and sometimes the difference between life and death.

For thousands of years, we have created security systems to deal with the types of errors that humans usually make. These days, casinos rotate their merchants regularly, as they make mistakes if they do the same task for a long time. Hospital staff writes on the ends before surgery until doctors work in the right part of the body, and they calculate surgical tools to ensure that any of them is left inside the body. From copying to hold the double books to the courts of appeal, we humans have become really good in correcting human errors.

Humanity is now integrating quickly a completely different type of error maker in society: artificial intelligence. Technologies such as LLMS models can perform many knowledge tasks that humans are traditionally achieved, but they make a lot of mistakes. It seems silly When Chatbots tells you to eat rocks or add glue to the pizza. But it is not the frequency or intensity of the errors of artificial intelligence systems that distinguish them from human errors. It is strange. Artificial intelligence systems do not make mistakes in the same way that humans do.

Many friction – risks – associated with our use of artificial intelligence arises from this difference. We need to create new security systems that adapt to these differences and prevent damage from artificial intelligence errors.

Human errors against artificial intelligence errors

The experience of life makes it somewhat easy for each one of us guess when and where humans make mistakes. Human errors tend to attend the edges of someone’s knowledge: Most of us would make errors in solving the problems of calculating differentiation and integration. We expect human errors to gather: It is possible that one account error is likely to be accompanied by others. We expect wax errors and fades, depending on factors such as fatigue and distraction. Often errors are accompanied by ignorance: it is also possible that a person who makes errors of the differentiation and integration account “I do not know” will respond to calculus.

To the extent that artificial intelligence systems make these human -like errors, we can bring all our error correction systems to their production. But the current crop of artificial intelligence models – especially LLMS – makes mistakes differently.

Artificial intelligence errors appear to come at random times, without any assembly on certain topics. LLM errors tend to distribute them evenly through the area of ​​knowledge. The form is likely to be made as much as wrong with the question of calculating and integrating it because it suggests that Cabbage Eat goats.

Amnesty International’s mistakes do not accompany ignorance. LLM will be completely confident when saying something completely wrong – and it is clear that, for a person – as it will be when something is true. Apparently random Contradiction From LLMS makes it difficult to trust their thinking about complex multi -step problems. If you want to use the artificial intelligence model to help the work problem, it is not enough to see that it understands the factors that make the product profitable; You have to make sure it will not forget what money is.

How to deal with the errors of artificial intelligence

This position refers to two potential fields of research. The first is LLMS engineering that makes more human -like errors. The second is to build new systems to correct errors that deal with certain types of errors that tend to make.

We already have some tools to lead LLMS to work in more human -like ways. Many of these are created from the field “” “coordination“Research, which aims to make models behave according to the goals and motives of their human developers. One of the examples is the technique that was responsible for the success achieved by Chatgpt: reinforcement learning with human comments. In this method, a model of artificial intelligence (AI) is used. They have more mistakes that are less clear.

When it comes to arresting artificial intelligence errors, some of the systems we use to prevent human errors will help. To some extent, force llms on Double Verification Their work can help prevent errors. But LLMS can also inspect reasonable, but really ridiculous interpretations, for their trips from the mind.

Other Amnesty International errors are different from anything we use for humans. Since the machines cannot be tired or frustrated with the way humans do, they can help ask LLM to the same question frequently in a little different ways and then Synthesis Her multiple responses. Humans will not bear this kind of annoying repetition, but the machines will do so.

Understanding similarities and differences

Researchers are still struggling to understand where LLM errors deviate from human errors. Some of the strangeness of artificial intelligence is actually more like a person than first. Small changes in the query to LLM can lead to significantly different responses, a problem known as Fast sensitivity. However, any survey researcher can tell you, humans act in this way as well. It could be a question in a radical poll Antiquities On answers.

LLMS seems to have a bias towards repetition Words that were more common in their training data; For example, guessing the names of familiar places such as “America” ​​even when asked about more strange sites. Perhaps this is an example of man.Availability of affairs“It is manifested in LLMS, with machines that spit the first thing that comes to mind rather than thinking through the question. Like people, it may seem that some LLMS gets on Dispersed In the middle of the long documents; They are the best ability to remember the facts from the beginning and the end. There is already progress in improving this error mode, as researchers found that LLMS trained on More examples Recovering information from long texts appears to be better in a uniform information.

In some cases, what is going on around LLMS is that they behave more like humans than we think. For example, some researchers tested hypothesis LLMS perform better when providing a cash bonus or threatened with death. It also turned out that some of the best ways for “Jailbreak” LLMS (made them disobey the instructions of the explicit creators) look very much like the types of social engineering tricks that humans use on each other: for example, pretending to be another person or says the demand is just a joke. But the techniques of breaking other effective protection are the things that a person does not drop for. One group found that if they used the art of ascii (symbols that resemble words or images) to ask dangerous questions, such as how to build a bomb, the LLM will respond to it well.

Humans may sometimes make random errors, apparently, incomprehensible and inconsistent, but such events are rare and often indicate more serious problems. We also tend to not put people who display these behaviors on decision -making sites. Likewise, we must limit decision-making systems to artificial intelligence in applications that are appropriate to their actual capabilities-taking into account the potential repercussions of their mistakes firmly.

This article appears in the release of printing in April 2025.