Safety When Using Artificial Intelligence
Artificial intelligence is quickly becoming part of everyday work processes. It makes writing, analyzing, and illustrating content easier, but it also brings risks that need to be understood and consciously controlled. In this article, you will find a practical guide on how to use AI responsibly, what to avoid, and how not to lose your audience’s trust. Main principles:
- Be careful what you upload to AI
- Pay attention to legal aspects.
- Don’t let yourself be deceived.
Table of contents.
What Not to Upload to Public Models
The basic rule is: do not share anything that should not fall into the wrong hands. Language models to which you send data over the internet may store it, use it for further training, or it may simply be intercepted during transmission. Therefore, never paste the following into such systems:
- personal data (PESEL national identification number, address, document numbers, logins, passwords),
- API keys, configuration files, or technical system information,
- trade secrets, contracts, and confidential documents,
- customer information that could violate privacy and trust.
If you must use AI to analyze such materials, use private instances, locally installed tools, or anonymize the data before uploading it.
.
How does AI differ from, for example, cloud storage?
Language models must process this data in order to generate a response. This process happens automatically, but it differs from storing a file in the cloud. In the cloud (e.g., on Google Drive), files are protected and accessible only to authorized users. Employees of the technology company do not look at their contents, and the data is stored in a static manner. In the case of language models, the data is analyzed by AI mechanisms with much greater processing capabilities. There have been instances where fragments of sensitive information have leaked from systems such as ChatGPT. Therefore, it is better to adopt a policy of utmost caution and not place anything there that should remain confidential.
.
The Temptation to Use Data for AI Training
Before the AI boom triggered by the release of the public version of ChatGPT, technology companies had little need to use user data to train their own models. Therefore, nowadays one needs to be somewhat more careful when uploading any data to the internet, including, for example, Google Drive. But one still needs to be more cautious when talking to ChatGPT or Gemini than when adding a file to Google Docs.
Connecting language models to an online drive
Language models can currently be connected to cloud drives. In such a case, you also need to be more careful about what you have on such a drive.
A Portal Providing Access to Multiple AI Models for a Small Fee
There are many platforms that provide access to various AI models in one place for a small fee. This is a convenient solution if you want to test Grok, Gemini, Claude, or GPT without having to create multiple accounts, pay for subscriptions, and so on. Many of these sites operate completely legally, obtaining access to multiple models through APIs. However, this adds an additional layer of risk—the creator of a given site may have access to all the messages written in the chat. I previously wrote about being careful about what you write in ChatGPT; nevertheless, OpenAI still has to comply with specific data security standards.
However, anyone can create a website that allows users to access various models, in such a way that the website creator has visibility into everything.
Larger services providing access to various AI models (such as Perplexity) are generally safe.
Data Leak from Public Threads.
The feature for sharing ChatGPT conversations via a link made the entire conversation available in Google Search. Shortly after the scandal at OpenAI, it came to light that a similar problem existed in Grok.
.
Fraudulent Schemes
Scams Impersonating AI
The promise of “free, early access” to a new AI tool can be a trap. Scammers lure people with attractive offers, while actually distributing viruses. In this case, the attack is a traditional one, and artificial intelligence merely serves as bait. The rule is simple: do not download programs from unknown sources, scan files with antivirus software, and install only from official websites.
Moreover, all AI models that are not run on a local drive have browser-based versions that can be used without installing any software. Models intended for local installation usually have high requirements (8–16–31 GB of free RAM) and also have communities of users.
AI-Powered Scams
In addition to the classic threats already mentioned, an entirely new category of scams based on the use of artificial intelligence capabilities is also emerging. The best-known example is deepfakes—video and audio materials that, thanks to AI, look like authentic recordings of real people. They can be used for fraud, manipulating public opinion, or impersonating someone in a phone conversation.
This is not merely about defrauding people using the “grandparent scam.” With AI, for example, individuals with fictitious identities can be introduced into a company, using deepfakes during video calls. There have also been cases where attackers impersonated, for instance, the CEO or the finance department and requested that bank transfers be made.
The second risk involves attacks aimed directly at the services we use every day. There are already attempts to break into email or cloud accounts using AI tools. There have been cases where an attacker exploited a security vulnerability in Google Calendar connected to Gemini, thereby gaining access to Gmail. Such scenarios are becoming increasingly realistic because models learn ways to bypass restrictions faster than traditional software.
AI is also used to create code that is later used in attacks. Code generators (e.g., Claude or ChatGPT) can write scripts that automate break-ins, prepare phishing websites, or build tools for scanning security vulnerabilities. Attackers often ask the model to write an “apparently harmless” piece of code, which they later combine into a whole. As a result, malware is created that its author did not even have to program from scratch.
That is why the principle of double caution applies when it comes to AI. What is a tool that makes your work easier for you may be a way to attack someone else. It is worth knowing about these threats and responding early, before they become commonplace in your environment.
Risk Levels When Using Artificial Intelligence
The risk depends on who has access to a given solution and on what that artificial intelligence does.
Who has access to the tool? The more people, the greater the risk:
- Self-use
- Access is limited to selected people within the organization
- Everyone in the organization has access
- The tool can be used by clients or people outside the organization.
When making tools available externally, you need to be especially careful. An internal solution used as support does not have to be perfect, because employees may be required to follow specific procedures. However, customers talking to a chatbot must not receive false information. There is also a risk that they will use such a model purely for entertainment, wasting company resources. People like to jailbreak models, often just for fun. There have also been cases where users negotiated huge discounts through a chatbot, and the company could not cancel them. This could have been easily prevented.
The second list of risk levels concerns what AI does.
The more decisions are “de facto” made by AI, the more thoroughly we need to prepare. AI used for internal text correction does not pose a major threat. But artificial intelligence that provides medical advice or decides who can receive a more favorable offer must already be properly secured, and the customer must be aware of how a given company uses AI. There is also the issue of customer data security and GDPR procedures.
We can distinguish several levels:
- internal support for learning/proofreading texts
- providing advice to clients
- A chatbot carrying out certain procedures and making simple decisions
- analysis and processing of customer data
- medical advice
- making financial decisions
I will use an abstract example here: AI analyzing the risk of granting a loan to a customer. Artificial intelligence may conclude that people with higher incomes are more likely to repay a loan without any problems, so the risk is lower and, consequently, the cost of the loan is lower. And such a conclusion is entirely correct.
But on the other hand, AI may conclude that, for example, people who wear blue shirts often have trouble repaying their loans, and therefore increase the costs for those people. Such a conclusion is incorrect and discriminatory.
Highlighting the key stages of risk is an important part of the renowned European AI Act
Privacy and Trust
Privacy is the foundation of responsible AI use. Remember that a statement about security is not enough for customers—they want to be certain that their data will not leak. That is why it is worth reviewing service providers’ policies, using encryption, and documenting who uses AI tools in the company and for what purpose. Regular audits are the best protection against errors and changes to providers’ terms and conditions.
."""
Access to Company Tools
There are cases where, in order to implement AI in a small company, a single ChatGPT license is purchased and used by, for example, a dozen or so people. This approach creates an obvious risk of unauthorized data transfer between departments. Sharing a single license with the entire team makes sense only when this involves typical testing.
AI Acting as an Agent
Increasingly, artificial intelligence not only writes but also operates in agent mode—with access to files or the system. This is a great convenience, but also a threat. An agent may, for example, accidentally delete files or overwrite data. Therefore:
- always assign it the minimum possible permissions,
- log all actions,
- test in a sandbox environment,
- have a backup ready and the ability to quickly roll back changes.
Without such safeguards, data can easily be lost, and sometimes further work may even be prevented.
Hallucinations and the Risk of Errors
Language models often create content that appears credible but is actually false. These are known as hallucinations. They may concern dates, figures, quotations, or proper names. Therefore, treat generated content as a draft, not as finished material. Everything must be checked: the more significant the information (e.g. medical, financial, or legal), the more rigorous the verification should be. It is best to introduce a simple procedure: generate, edit, check the sources, and only then publish.
Content and Graphics Created by AI
Graphics generated by artificial intelligence are useful as illustrations or quick mockups. However, it is important to remember that they often contain errors: distorted lettering, strange character proportions, or incorrect details. Therefore, they are not always suitable for official materials.
Graphics used to illustrate a blog post or as a decorative element in a social media post are generally not a problem, as long as they do not mislead the user.
However, in the case of more important graphics, manual work is necessary (although using AI at some stage is acceptable). The recipient may get the impression: if this large company cannot afford to spend a few thousand on a graphic designer and uses “free” graphics from ChatGPT in its main advertising campaign, costing more than ten thousand, that means it is not doing very well financially.
The same applies to texts—AI is excellent for creating drafts, but publishing them without editing is a surefire way to lose credibility.
ChatGPT wrote:
A common mistake made in haste that can trigger a crisis in more than one company. Even if the text has been carefully considered and, for example, a language model was used only for proofreading, pasting the text together with phrases typical of ChatGPT causes the reader to lose trust in it.
Labeling AI-Generated Content
When you use AI, a question arises: does it need to be disclosed? The answer depends on the context. In content of significant importance (financial, scientific, or public), honesty toward the audience requires clearly indicating the involvement of artificial intelligence. In simple, illustrative materials, this is not always necessary. It is also advisable to specify the extent to which AI was used in creating the content.
Integrity and Reputation
Poorly used AI can do more harm than good. If your audience feels deceived or notices that you publish error-filled content, you will lose their trust. That is why transparency and accountability are crucial. Customers want to know that you care about quality and that AI is merely a supporting tool, not a replacement for your knowledge and experience.
What AI Can Do and What It Doesn’t Do Best
Finally, it is worth remembering the real capabilities of artificial intelligence. It works very well for:
- creating text drafts,
- analyzing large datasets and quickly summarizing them
- generating simple illustrations in a specific style.
It performs less effectively in areas requiring precise facts, in-depth specialist analysis, or an understanding of the moral and legal context. Therefore, you will achieve the best results when you let AI support you where it is strongest, while leaving decisions and responsibility to humans.
This text was created with the support of artificial intelligence (mainly for proofreading).