In September 2023, employees at Onclusive, a media monitoring company headquartered in Courbevoie, in the Paris region, were informed that 217 of the 383 jobs would be eliminated. The stated reason was the adoption of artificial intelligence (AI) tools. The unions reacted. The regional labor authority deemed the information provided to employee representatives insufficient and questioned the criteria used to select those affected. The company backed down and announced the suspension of the layoffs. In January 2024, a new plan was presented, still calling for approximately 218 job cuts.
The case has become one of the most frequently cited examples of mass layoffs openly attributed to AI, and the numbers help explain why it won’t be the last. A study by the International Labor Organization (ILO) in collaboration with the Polish institute NASK, released in May 2025, estimates that one in four jobs worldwide is in occupations exposed, to some degree, to generative AI. The good news is that, in most cases, the trend is toward the transformation of job roles, not their elimination. The bad news is that companies aren’t waiting to see what happens. A survey by Thomas H. Davenport and Laks Srinivasan of 1,006 executives, published by the Harvard Business Review in January 2026, shows that AI-related layoffs are almost always being made in anticipation of the expected impact, rather than based on results already obtained. Layoffs, therefore, are being made based on the promise of future changes. And in a country with as many labor disputes as Brazil, decisions made based on promise tend to lead to lawsuits.
For companies, AI can impact employment relationships in two ways: it can be the cause of termination, when the technology renders certain roles unnecessary, or the tool for decision-making, when a system evaluates performance, ranks employees, and determines who is let go.
In both cases, legislation and case law impose limits that, if not observed, expose companies to risks in Brazil.
Artificial Intelligence as a Cause for Termination
ILO Convention No. 158 permits termination of an employment contract only for just cause, related to the employee’s ability or conduct or to the company’s operational needs. When terminations are based on “economic, technological, structural, or similar grounds,” Articles 13 and 14 go further: the employer must inform employee representatives in a timely manner of the reasons, the number and categories of affected employees, and the period of the terminations; initiate consultations on measures to prevent them or mitigate their effects; and notify the competent authority.
In Brazil, however, this rule does not apply. Although ILO Convention No. 158 came into force in Brazil in 1996, it was denounced that same year by Decree No. 2,100/96. In June 2023, in ruling on ADC 39, the Federal Supreme Court (STF) decided that the denunciation of treaties approved by Congress requires legislative approval. However, it applied this interpretation only prospectively and upheld the validity of the 1996 denunciation. In August 2024, the STF extended the same conclusion to ADI 1,625.
The Constitution promises protection against arbitrary dismissal (Art. 7, I) and “protection against automation” (Art. 7, XXVII); however, both of these protections depend on a law that has never been enacted.
In practice, regardless of the reason, private employers can still terminate an employee without just cause and without providing a reason. They do not need to prove that automation is real, nor that it affects that specific position. It is sufficient to pay severance pay and the 40% indemnity on the FGTS for the termination to generally be considered valid.
Nevertheless, if the impact of AI results in mass layoffs, one must observe STF Ruling No. 638, which established that prior union intervention is necessary—which should not be confused with the need for authorization from the union.
Artificial Intelligence as a Decision-Making Tool
In this case, companies must comply with the General Data Protection Law (LGPD—Law 13,709/18).
If the system, on its own, determines who will be terminated, Article 20 of the LGPD comes into play, which guarantees the data subject the right to request a review of decisions made solely on the basis of automated processing of personal data that affect their interests, including those that define their professional profile. Performance evaluations conducted by algorithms and systems that generate layoff lists based on productivity indicators are at the heart of this rule.
The same article requires the company to provide, upon request, clear and adequate information about the criteria and procedures used in the decision, while respecting commercial and industrial secrets. If the company refuses to explain by invoking confidentiality, the National Data Protection Authority (ANPD) may conduct an audit to verify whether there are discriminatory aspects.
The law does not explicitly mention the right to be heard, but the practical effect is similar: the employee may request explanations, challenge the decision, and contest the evidence used against them.
Two points often cause confusion.
The first is that the law does not require the review to be conducted by a person. The original wording of the LGPD referred to a review “by a natural person.” That phrase was removed by Provisional Measure (MP) 869/18. Congress attempted to reinstate it, but the provision was vetoed in Law 13.853/19, and the veto was upheld.
The second point lies in the word “solely.” A manager who merely rubber-stamps the list generated by the system is unlikely to make the decision a human one. In the SCHUFA case (C-634/21), decided in 2023, the Court of Justice of the European Union held that a score calculated by an algorithm already constitutes an automated decision when it plays a decisive role in a third party’s decision. Since Article 20 of the LGPD is modeled after Article 22 of the European General Data Protection Regulation, this precedent is likely to be taken into account here. Human involvement only resolves the issue when it is genuine: the decision-maker must have access to the information, the autonomy to disagree with the system, and the diligence to document why they did or did not follow the machine’s recommendation.
The ANPD has included the review of automated decisions in its regulatory agenda for 2025–2026, but there is still no specific regulation. Until then, it is up to companies to develop, on their own, procedures that would hold up in court. And this procedure cannot focus solely on who confirms the decision; it must also consider what the machine decides.
The Risk of Bias Disguised as Objectivity
The most serious risk of a list generated by an algorithm lies in the way it is constructed.
AI systems learn from past data and tend to repeat the patterns they encounter.
A model that measures productivity may penalize those who have been on maternity leave, on sick leave, or in rehabilitation. Variables that seem neutral, such as address or year of graduation, can serve as indirect indicators of age, origin, or race.
The result is a list that appears objective but is based on potentially discriminatory criteria.
Law 9.029/95 prohibits discriminatory practices in access to and retention of employment on the basis of sex, origin, race, color, marital status, family situation, disability, vocational rehabilitation, or age, among other factors. Once discriminatory termination is recognized, the employee is entitled to compensation for emotional distress and may choose between reinstatement—with full compensation for the period of absence—and double payment of wages for that period.
ILO Convention No. 111, in force in Brazil, also covers distinctions that have the effect of altering equal opportunities. Discrimination can therefore be recognized even if no one intended to discriminate. Article 6, IX, of the LGPD follows the same line of reasoning by including non-discrimination among the principles of data processing.
The problem becomes more acute when it comes to the burden of proof. Article 818, § 1, of the Consolidated Labor Laws (CLT) allows the judge to assign the burden of proof to the party that is best able to meet it. Furthermore, Precedent 443 of the Superior Labor Court (TST) presumes that the termination of an employee with a serious illness that gives rise to stigma or prejudice is discriminatory. If the company cannot explain why the system selected that particular employee, it will likely be unable to avoid the risk of reinstatement.
The apparent legislative gap on this issue is being addressed in Bill (PL) 2,338/23, which was approved by the Senate in December 2024 and is currently under review by a special committee of the Chamber of Deputies. The bill classifies systems used in recruitment and in decisions regarding employee promotions and terminations as high-risk, requiring an impact assessment and human oversight. The bill also establishes strict liability for high-risk systems in cases of financial, moral, individual, or collective harm.
Risk Mitigation
Even without a specific law, certain precautions can significantly reduce the risk:
- Map the use of AI in employment relationships. Determine at which stages AI is involved in evaluations, promotions, and terminations, and who makes the final decision in each case.
- Ensure human review. The final decision must rest with someone who has the information and autonomy to disagree with the system, with the rationale for each case documented.
- Test the model. Verify whether the results disproportionately affect specific groups that are subject to special legislative protection, such as women, older employees, people with disabilities, or those on sick leave, and document the analysis—for example, in a data protection impact assessment.
- Review contracts. Many of the tools are contracted from third parties. The contract must guarantee access to the information necessary to respond to requests based on Article 20, cooperation in audits, and liability for model failures.
- Respond to employees. Establishing a channel, a deadline, and an explanation of the criteria in understandable language tends to reduce the likelihood of future litigation.
- Scale the restructuring. If a decision is made to replace or downsize the workforce, ensure prior dialogue with the union.
For more content on artificial intelligence and labor relations, follow the publications on the Inteligência Jurídica portal. If you have any questions, please contact our team.
