AI in HR
How do AI chatbots work, and what does that mean for HR?
Florian Rocktäschel · Last reviewed: November 2026 · 8 minute read
In short
An AI chatbot predicts, word by word, the text that most plausibly follows your question. It has no knowledge of your company unless you give it documents, and it sounds just as sure when it is wrong. For HR this means: check every result, keep personal data out of tools your company has not approved, and leave decisions about people with people.
How does an AI chatbot produce an answer?
Tools such as ChatGPT, Copilot or Mistral Vibe are built on large language models. Such a model has been trained on very large amounts of text with one task: predict the next word, or more precisely the next token, which is a word or part of a word. After billions of examples it has learned which words tend to follow which, in which style and in which order.
When you ask a question, the model writes the answer one token at a time. Each step picks a likely continuation of everything that came before. A second round of training teaches it to follow instructions and to answer politely and helpfully. That is why the result reads like a competent colleague wrote it.
The model does not look anything up while it writes, unless the tool adds a web search or your documents. What it “knows” are patterns from its training data, up to a cut-off date.
Why does it sound confident and still get things wrong?
Grammar and style follow regular patterns, so the model learns them well. Rare facts, such as the number of a specific article in a Romanian law or the date of a court ruling, cannot be derived from patterns. Where the model has no reliable pattern, it still produces the most plausible-sounding text. These invented details are called hallucinations.
OpenAI explained in September 2025 that standard training and testing reward guessing over admitting uncertainty. A model that says “I don’t know” scores worse than one that guesses and is sometimes right. The result is a tool that rarely hesitates. Microsoft’s own documentation for Copilot states that generated responses are not guaranteed to be factual.
For HR the risk is clear: a made-up legal article in a policy, a wrong notice period, or a summary that quietly leaves out an exception.
What is the context window?
The context window is the amount of text the model can take into account at once: your conversation, the instructions and any documents you attach. Everything outside it does not exist for the model. In a very long chat, early parts can drop out. Research also shows that models use information at the beginning and end of a long text better than in the middle. Long policies are therefore best handled section by section.
Why does the chatbot not know your company?
Your pay bands, your collective agreement and your internal rules were never in the training data. A general chatbot fills such gaps with what is typical elsewhere. It only works with your rules if you give them to it, by pasting text, uploading a document or using a tool connected to company systems. Microsoft 365 Copilot, for example, can read emails and documents through the permissions of the person using it. That is useful and also a reason to check who can see what.
What happens to the data you type in?
That depends on the tool and on the contract, not on the model.
| Tool and plan | Used for training by default? |
|---|---|
| ChatGPT Free, Plus, Pro | yes, can be switched off in the settings |
| ChatGPT Business, Enterprise, API | no, according to OpenAI |
| Microsoft 365 Copilot with a work account | no, according to Microsoft; EU traffic stays in the EU Data Boundary |
| Mistral Vibe, individual and team plans | yes according to Mistral’s help centre, with opt-out; Enterprise excluded |
| Open model on your own servers | no data goes to a model provider |
The table reflects the providers’ statements as of November 2026; settings and terms change. Check them before your team uses a tool for work, and keep in mind that training is only one question. Storage, access by the provider and the location of the servers matter as well.
What follows for HR work?
Three principles cover most situations:
- Check every result. AI suggests, a person checks, decides and can explain the decision.
- No personal or confidential data in tools the company has not approved. A private account is not an approved tool.
- People decide about people. AI can prepare a text, but it does not select, rank or judge anyone.
Which five uses work today?
These tasks have one thing in common: the result is a draft that someone with HR knowledge reviews, and no decision about a person depends on it.
- Drafting job ads. Give the tool the job description and ask for a neutral, clear ad. Then check for wording that excludes people, for example by age or gender, and add the pay range yourself.
- Structuring interview guides. AI turns a job profile into questions grouped by competence. The same questions and the same scoring sheet for every candidate make interviews fairer. The scoring stays with the interviewers.
- Summarising policies. A long works rule or procedure becomes a one-page summary for managers. Compare it with the original, section by section, because summaries tend to drop exceptions.
- Translating. For groups working in Romanian, English and German, AI produces good first versions of letters and policies. Legal or contractual texts still need a person who knows both languages and the subject.
- Preparing training material. Outlines, exercises, quiz questions and examples for onboarding or internal training. Facts and legal references in the material need checking like everything else.
Which three uses should HR leave alone?
- Deciding about or ranking candidates. Sorting CVs or scoring applicants is exactly what the AI Act lists as high-risk in Annex III no. 4 (a). As of November 2026, the AI Act provides strict duties from 2 December 2027 for companies that use such systems: human oversight, informing staff, keeping logs.
- Suggesting pay. A tool that proposes salaries can fall under Annex III no. 4 (b), which covers decisions on the terms of employment. Whether a specific tool falls under it is a question for a law firm. The Pay Transparency Directive expects pay to rest on objective criteria, and in our view a statistical guess is not such a reason.
- Judging performance. Monitoring and evaluating the performance and behaviour of employees is also high-risk under Annex III no. 4 (b). Emotion recognition at the workplace, such as analysing faces or voices in video calls, has been banned since 2 February 2025 under Article 5 (1) (f), with exceptions only for medical or safety reasons.
What does the AI Act require of employers now?
Article 4 on AI literacy has applied since 2 February 2025. The Digital Omnibus on AI, in force since 27 July 2026, kept the duty and changed its form. The article now requires companies that use AI to take measures to support the AI literacy of their staff. They are no longer required to guarantee that every person reaches a certain level. In our view, that means in practice rules for use, training adapted to the tasks, and a written record of what the company has done. A survey of 1,003 Romanian companies in July 2026 found that only 26 % had a policy for AI use. A training such as our AI in HR course can be one of these measures, but no single course fulfils the duty on its own.
Which tools, and when?
We name tools here so that you can classify them, not as a recommendation.
- ChatGPT. Yes, with a business plan for texts without personal data, or for research and wording. No, for applicant or employee data in a private or free account.
- Microsoft 365 Copilot. Yes, if your company already works in Microsoft 365 and has set permissions carefully, because Copilot sees what the user sees. No, if file permissions have grown unchecked for years.
- Mistral Vibe (formerly Le Chat). A European provider based in France. Yes, if data should stay with an EU company. No, before you have checked plan and training settings.
- Open models on your own servers or at a European hosting provider. Yes, for sensitive HR data and for groups that want full control. No, as a quick start, because operation and security need IT capacity.
Four rules for daily work
- Check: AI suggests, I check, decide and can explain it.
- Data: No personal or confidential data in tools the company has not approved.
- People: I do not hand decisions about people to AI.
- Doubt: When in doubt, I ask my manager, data protection or compliance, not the AI.
Frequently asked questions
In our view, only if your company has approved the tool for personal data and the contract excludes training on your inputs. With a free or private account, no. A CV is personal data under the GDPR.
The model does not change during your conversation. Whether the provider uses your inputs later to train new models depends on the plan and settings. Business plans usually exclude it, free plans often include it by default.
As of November 2026 it is not banned, but under the AI Act it counts as high-risk from 2 December 2027, with duties such as human oversight and informing applicants. Until then the GDPR already applies. We recommend that people make the selection.
Article 4 requires companies to take measures for the AI literacy of staff who use AI. It does not prescribe a specific course or certificate.
Sources
This article reflects our understanding as of November 2026. Laws, drafts and court practice can change at any time; we update the article when they do, but we cannot guarantee that it is complete or current. It is general information, not legal advice; whether a specific case is legally justified is assessed by a law firm.
- AI Act, Regulation (EU) 2024/1689, Art. 4, Art. 5 (1) (f), Annex III no. 4: EUR-Lex
- Digital Omnibus on AI in force, high-risk rules from 2 December 2027: White & Case
- Article 4 after the Omnibus: KI-Campus, July 2026
- AI use and policies in Romanian companies: undelucram.ro, July 2026
- Why language models hallucinate: OpenAI, September 2025
- Long contexts: Liu et al., Lost in the Middle, 2023
- ChatGPT and training on user data: OpenAI Help Center
- Microsoft 365 Copilot, data and privacy: Microsoft Learn
- Mistral and training on user data: Mistral Help Center
About the author
Florian Rocktäschel is an engineer (RWTH Aachen) and has been building data and AI systems since 2013.
Where does your company stand?
The self-test takes about ten minutes. It shows where your company stands on the main duties of the directive. Free of charge.

