Superintelligence in HR: Bias, LGPD and Decisions to Keep Human

Superintelligence in HR: Bias, LGPD and Decisions to Keep Human

Superintelligence in HR, in the strict sense of a system far beyond human judgment in almost everything, does not exist today. What companies have is AI that is increasingly good at reading resumes, summarizing interviews, suggesting learning paths and analyzing internal sentiment, and it is that capability that calls for rules. The central risk is that a more fluent system looks fairer and more reliable than it is, and that decisions about people start being made without human review and without a clear legal basis.

Key takeaways

• More capable AI improves screening, summarizing and organizing, but it does not remove bias. It can reproduce the pattern in the company's own history with more conviction.
• In HR, almost everything involves personal data, and some of it may be sensitive. LGPD (Brazil's data protection law) requires purpose, necessity, transparency and care with automated decisions.
• Hiring, promoting, dismissing, issuing warnings and evaluating performance should remain human decisions, with AI as support and record-keeping.
• Culture is not measured by monitoring. Using AI to watch people destroys the trust you wanted to measure.
• The safe path is to map the process, limit data and actions, pilot one flow with logs, and only then decide whether to expand.

What does superintelligence in HR mean today?

The term comes from the debate about systems that would surpass human ability in practically every field. For someone making decisions in a company, it helps to separate that philosophical debate from what is actually running. For that distinction, read AGI, ASI and AI agents: what your company can actually run.

In HR, what exists are language models and agents that perform bounded tasks: sorting applications, drafting job descriptions, summarizing a conversation, suggesting interview questions, building a development plan from a profile. The more capable the model, the more convincing the text it produces. That is useful and dangerous at the same time, because a well-written justification for a candidate ranking does not prove the ranking is right.

So the useful question for a manager is not when superintelligence arrives. It is this: which decision about a person are we letting a system influence, with what data, and who is accountable for it?

Where more capable AI helps in recruiting, development and culture

There are real gains when AI handles preparation work and leaves judgment to people.

In recruiting, it can standardize job descriptions, organize resumes received by email or WhatsApp, summarize what each candidate stated against the written requirements, and prepare the same interview script for everyone. A standard script usually reduces improvisation, which is a known source of inconsistency.

In development, AI can match skills that employees describe themselves with the available learning paths, suggest readings and courses, and help managers write clearer feedback. The point of attention is who chooses and who sees the result: the plan belongs to the employee and the manager, not to a dashboard that ranks people.

In culture, it can summarize open-ended answers from an engagement survey, group themes and draft internal announcements. The risk here is re-identification: comments from small teams can expose who wrote them. If you want to measure work with AI without sliding into surveillance, there is a practical guide in how to measure AI productivity without surveillance on your team.

How does bias show up even with a better model?

A more capable model is not neutral just because it is more capable. Bias enters through at least four paths.

• History: if the system learns from or is calibrated on past hires, it tends to favor the profile the company has always hired.
• Proxies: postal code, school name, career gaps, surname or type of hobby can act as indirect stand-ins for social background, race, gender or age.
• Language: resumes written by people with less access to career coaching tend to score lower on style criteria, not on competence.
• Overconfidence: the human reviewer starts agreeing with the AI suggestion by default, and review becomes a rubber stamp.

The fourth point is the most underestimated. Human review only protects if the person has time, information and the authority to disagree. A flow where the recruiter sees only the candidates the AI approved has no real review, because nobody examines who was discarded.

In the United States, NIST publishes an AI risk management framework that helps organize this discussion around mapping, measuring and managing risk. In Europe, the EU AI Act treats systems used in employment and worker management as high risk, which shows where regulatory attention tends to point. Neither is Brazilian law, but both are useful as method references.

What does LGPD require for AI in HR?

LGPD applies to candidates, employees, interns and former employees. Here are some practical points, which do not replace the opinion of a data protection lawyer:

• Purpose and necessity: use only the data the decision requires. Photo, marital status and date of birth are rarely needed to screen an application.
• Legal basis: each processing activity needs an appropriate basis. Legitimate interest requires its own assessment, and consent from a candidate in a weaker position can be fragile.
• Sensitive data: health information, racial origin, religious belief, union membership and biometrics fall under a stricter regime. Keep them out of prompts and talent databases.
• Non-discrimination and transparency: candidates should be able to understand, in plain language, that AI is involved and for what purpose.
• Automated decisions: Article 20 gives people the right to request review of decisions made solely through automated processing that affect their interests, including those that define a professional profile. We cover this in LGPD automated decision-making: Article 20 and the right to review.
• Retention: define how long resumes and evaluations are kept and who has access, including in the tool's logs.
• Vendors and cross-border processing: if the AI tool processes data outside Brazil, the international transfer needs a legal foundation. See international data transfer under LGPD when using AI providers.

The ANPD, Brazil's data protection authority, is the official source for guidance. For higher-risk processing, such as systematic evaluation of people, ask your data protection officer or legal team whether a data protection impact assessment is advisable before the pilot.

What should remain a human decision?

A simple rule: the greater the effect on a person's working life, the more human the decision. Use this list as a decision guide.

• Hiring or rejecting: AI can organize and summarize, but the decision and its justification belong to a named person.
• Promoting, adjusting pay or changing roles: criteria defined and recorded by managers and HR, with review by someone who did not take part in the nomination.
• Issuing warnings, suspending or dismissing: never delegate. AI should not even draft the final conclusion of the case.
• Evaluating performance: AI can consolidate evidence the manager already has, but should not assign a score or ranking without review.
• Interpreting health, emotion or behavior: do not use AI to infer emotional state, mental health or personality of candidates and employees.
• Handling complaints, harassment or conflict: led by trained people, with confidentiality and traceability.
• Defining criteria and weights: what a good candidate means is a company decision, written down before the tool is switched on.

What can be automated safely: scheduling interviews, sending confirmations, answering hiring process questions with approved text, summarizing documents the candidate has already provided, and organizing training material.

Illustrative scenario: screening operational roles

Imagine, as an illustrative scenario, a company that receives many resumes for operational roles by email and WhatsApp. Today HR reads everything in spreadsheets and ends up prioritizing whoever writes best.

A careful design would look like this: the AI extracts from each application only relevant experience and availability, in fixed fields, without name, photo, address or age at the reading stage. The recruiter sees all applications, not just the approved ones, with the summary next to the original. The AI flags open questions instead of scoring, for example: shift information missing. Every rejection is made by a person, with a reason chosen from a list tied to the job requirements. The logs record what was extracted and what the recruiter decided, for audit and to handle review requests.

The gain is reading time and consistency. What does not change is who decides.

How to apply it: map, limit, pilot and hand over

In the method we use in AI projects, order matters, and it fits HR well.

• Map: describe the current process end to end, who decides each step, what data moves through it and where it lives (spreadsheet, email, ATS, WhatsApp). Record a current quality measure, such as response time to candidates or rework in screening.
• Limit: define which data the AI can see, which actions it can take (summarize, schedule, draft) and which stay out (reject, rank, send a final decision). Remove sensitive data and unnecessary identifiers.
• Pilot: choose one small, real flow, such as interview scheduling or resume summaries for a single opening. Keep logs, compare samples of results against human evaluation across different profiles, and agree in advance which metric defines success.
• Hand over: give HR the documentation, the person responsible for operations and the procedure for review requests, and then decide whether to expand, adjust or stop.

The pilot should include a simple bias test: run equivalent applications that differ only in signals such as name or institution, and check whether the result changes. If it does, the flow is not ready, no matter how good the writing looks.

To evaluate rigorously, treat the set of test cases as part of the product: it needs an owner, periodic review and a record of failures. Switching models or vendors without repeating this test is a common way to bring old problems back.

Frequently asked questions

Is superintelligence in HR already a reality? Not in the sense of a system that surpasses human judgment across the board. What exists are models that are capable at specific tasks, and that is where the rules on bias, LGPD and human oversight need to be applied now.

Can I use AI to rank candidates? It may be technically possible, but it is the highest-risk practice, because it moves the flow toward an automated decision about a professional profile, subject to Article 20 of LGPD. If the company chooses this route, it needs documented criteria, bias testing, transparency to the candidate and effective human review. Otherwise, use AI to summarize and organize.

Do candidates need to know the company uses AI in hiring? Transparency is a principle of LGPD, so telling candidates clearly that AI is involved and for what purpose is the safest practice. The exact wording and the legal basis should be defined with your legal team or data protection officer.

Where do I start without taking on too much risk? Start with tasks that do not decide anything about people, such as interview scheduling, standard answers to hiring process questions and summaries of documents already submitted. Once you have logs, a bias test and a named owner, consider extending to steps that influence the decision, always with a person in charge.

Sources

• NIST: AI Risk Management Framework
• ANPD: Autoridade Nacional de Proteção de Dados
• EU Artificial Intelligence Act: visão geral