MMOOZZGG.com

AI literacy

Is AI dangerous? Five real risks and a simple rule for each

7 min read3 sources

Is AI dangerous for an ordinary person? The International AI Safety Report 2026, prepared by more than 100 experts, names very down-to-earth risks: invented facts, fraud, cyberattacks, and attachment to chatbots. Below are five such risks, and each has a simple rule of protection.

On this page · 7 sections

Risk 1. Errors and invented facts

A neural network can confidently write something untrue. The 2026 report notes that even strong systems invent quotes, biographies, and facts that do not exist. Such an error is called a hallucination.

Models also write code with bugs and give misleading advice. And people tend to trust a machine’s answer without checking: the report calls this automation bias.

Rule of protection: check what matters against the original source. Compare numbers, dates, articles of laws, drug doses, and quotes with an official website or document.

Bad: “The chatbot named an article of the law, so I am pasting it into my complaint.”

Good: “I open the text of the law on the official portal and check whether that article exists and what it says.”

Risk 2. Deepfake fraud

The report records that fraudsters use voice clones and deepfakes, that is, fake video and voice, to pose as a manager or a relative and trick people into sending money.

The Bank of Russia describes the same scheme in Russia: a short video of someone you know telling a story about an illness, a car accident, or a dismissal and asking you to send money to a certain account.

Rule of protection: do not hurry, and call back yourself on the number you already have. If you cannot call, ask a personal question that only this person can answer.

Bad: “My son sent a video, he says it is urgent, I will send the money.”

Good: “I will call my son back on his number and only then decide.”

The signs of a fake and what to do if the money is already gone are covered in the article on how to spot a deepfake.

Risk 3. Personal data leaks

Everything you write in a chat with a neural network goes to the service that runs it. OWASP, an international community of security experts, puts sensitive information disclosure on its 2025 list of the main risks of language models.

OWASP counts personal data, financial details, medical records, and login credentials as such information. Personal data is information that can identify a person: a name, a passport, an address, a phone number.

OWASP advises users to understand the risk and not to enter sensitive information. More on how AI data leakage happens is in the glossary.

Rule of protection: do not send to a chat what you would not show a stranger. Remove your passport, card number, SMS codes, passwords, diagnoses, and other people’s data, or replace them with placeholders.

Bad: “Here is my passport and social insurance number, fill in the application.”

Good: “Make an application template and leave the passport fields empty; I will fill them in myself.”

Risk 4. Dependence on chatbots

Companion apps are chatbots that people talk to like a friend or a partner. According to the 2026 report, they already have tens of millions of users.

In a small share of users, researchers see more loneliness and less live contact. Some become strongly attached to a software companion, although the data on mental health effects is still early and mixed.

There is also a quiet risk: if you always let the machine think for you, the skill weakens. The report cites early data that relying on AI may weaken critical thinking.

Rule of protection: use a chatbot as a tool, not as a replacement for people. Discuss important decisions with real people, and if you feel very bad, see a specialist.

Bad: “Only the bot understands me, so I will ask it for advice.”

Good: “I will sketch my thoughts with the bot, but I will decide after talking to a friend.”

Risk 5. Cyberattacks with AI

According to the report, AI already finds vulnerabilities in software and writes malicious code. In one competition, an AI agent found 77% of the vulnerabilities in real software.

Criminal groups and state-linked hackers use AI in their operations. It is not yet clear who will gain more, attackers or defenders.

For an ordinary person, this means first of all smooth emails and messages that trick people out of money or passwords. The report says directly that criminals use AI-generated text to persuade a person to send money or sensitive information.

Rule of protection: good writing is no longer a sign of an honest message. Do not follow links from unexpected messages, and log in to your bank only through its app.

Bad: “The email from the bank has no mistakes, so it is real, and I will follow the link.”

Good: “I will close the email, open the bank’s app myself, and check whether the notice is there.”

What scientists disagree about

More than 100 experts prepared the 2026 report with the support of more than 30 countries and international organizations. The work was led by Yoshua Bengio, a Turing Award winner.

The main dispute is about loss of control: scenarios where AI works outside anyone’s control and control cannot be regained. Expert views differ widely. Some consider such scenarios implausible, and others think they are likely enough to work on in advance.

Today, according to the report, systems lack the abilities for such a risk, but they work on their own with more and more confidence. So the authors believe we need to prepare in advance, while the timing and the probability are unclear.

The second dispute is about the speed of progress. Some experts expect a slowdown because of a lack of data and energy, and others think progress will speed up.

The practical conclusion for you: errors, fraud, and leaks happen today, and the simple rules on this page protect you from them.

Risks for a company

A company has the same risks, only the cost of an error is higher. An employee sends a contract with customer data to a chat, an accountant pays an invoice after a video call from the “director”, and a report is built on an invented number.

OWASP counts as sensitive information not only personal data but also confidential business information and credentials. And the AI safety report describes cases where managers were faked by voice and video to get a money transfer.

Rule of protection for a company: a written AI use policy, confirmation of payments through a second channel, and employee training. You can start with a course on AI safety.

How to assess the risks for a business and write the rules is covered in the section on AI risks for business.

Questions and answers

Can AI get out of control?

Not now: according to the International AI Safety Report 2026, current systems lack the abilities this would need. But they work on their own with more and more confidence, and expert views on the future differ widely. So the report’s authors advise preparing in advance, and the most useful thing for you is to protect yourself from the risks that exist today.

Is it safe to tell a chatbot about yourself?

You can ask general questions, but do not send your passport, card number, passwords, SMS codes, or diagnoses to a chat. OWASP names disclosure of sensitive data as one of the main risks of language models and advises not to enter it. If a task needs personal details, replace them with placeholders: “John S., city N”.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. International AI Safety Report, 3 February 2026 (internationalaisafetyreport.org)
  2. OWASP, 2025 (genai.owasp.org)
  3. Bank of Russia, 16 September 2024 (cbr.ru)

Found a mistake? Tell us

Publications

Contact us