Yes, AIs can lie; understand why this happens and see how to avoid it

Artificial intelligence chatbots, such as ChatGPT and Google Bard, have become common features in our daily interactions, but have recently been the subject of discussions about their limitations and propensity to provide wrong answers.

These glitches, called “hallucinations,” can not only confuse users, but also spread misinformation and even propagate prejudices and stereotypes.

Understand what AI ‘hallucinations’ are

The chatbots of AIin their operation, they respond using machine learning, creating sequences of words to satisfy user requests.

However, hallucinations occur when these tools do not have enough information for an adequate response, resulting in incorrect but convincingly presented answers.

These errors can be exacerbated by triggers such as insufficient, outdated, or low-quality databases, as well as a lack of understanding of ambiguous or complex commands.

The lack of information encourages AI hallucinations – image: Shutterstock/Reproduction

In addition to confusing users with inaccurate information, these flaws can fuel the spread of prejudiced content and fake news.

Experiments carried out by renowned institutions show that chatbots can generate false information more convincingly than humans and, in some cases, perpetuate racial and gender stereotypes.

While the responsibility for preventing hallucinations should lie with the companies behind these AIs, users can adopt some strategies to reduce the likelihood of these errors.

Precise and specific commands limit the possible results, reducing the risk of hallucinations. Requesting more direct answers, such as “yes” or “no,” and directing the AI ​​not to provide false information are useful practices.

Prevention strategies

1. Objective instructions: By giving specific commands, the user narrows the possible results, reducing the risk of hallucinations.
2. Specific function: assigning a specific role to AI, asking it to respond as an expert in a given field, thus limiting responses.
3. Exclusion criteria: Ask the tool to discard outdated data or fictitious information.
4. Reference models: generate data tables or reference models to avoid ambiguities in numerical or complex answers.

Although total prevention of these errors is responsibility For companies that develop AIs, understanding these limitations and adopting practices to minimize risks are essential steps in interacting with artificial intelligence chatbots.