Serbia Data Protection & Privacy Regulation

Algorithmic, Biometric & Surveillance Governance

Serbia currently has no dedicated statutory framework governing algorithmic or high-risk artificial intelligence systems within its data protection law. That gap is precisely what the draft 175-article replacement Law on Personal Data Protection, released on 30 July 2026, is reported to address. Among its additions, the draft would introduce a mandatory data protection impact assessment requirement for high-risk AI systems, paired with a prior-opinion mechanism requiring sign-off from the Commissioner for Information of Public Importance and Personal Data Protection before such systems could be deployed. Neither requirement exists under the current 2018 Law, and neither is yet in force; the draft remains in post-consultation status with its next legislative step unconfirmed.

The significance of this provision lies less in its detail than in its structural placement. Rather than creating a free-standing AI law, Serbia's draft approach would route AI governance through the existing data protection authority and the existing DPIA mechanism already used for other high-risk processing activities. This would hand the Commissioner a gatekeeping function specifically over deployment of high-risk AI systems, prior to launch, rather than a purely reactive enforcement role after harm has occurred. If enacted as reported, this would represent a meaningful expansion of the Commissioner's practical authority, extending its remit from traditional data protection supervision into pre-deployment technology review. </p>

Because this entire development rests on a draft that has not been enacted, and because the precise operative detail, such as how high-risk status would be determined, what categories of system would be captured, and what form the prior opinion would take procedurally, has not been independently confirmed, this finding is reported with appropriate hedging. Reports suggest the mechanism exists in the draft text; confirmation of its final form, or any form at all, depends on what happens to the broader legislative package.

Outlook

The principal uncertainty is whether the draft law advances at all, and if so, whether the AI-governance provisions survive the legislative process intact or are narrowed. Because this provision is embedded within the larger 175-article package rather than standing as separate legislation, its fate is tied to the overall bill's progress, meaning that confirmation of the National Assembly timetable would be the clearest forward indicator to watch. Any confirmation that the DPIA and prior-opinion requirement has been carried into an enacted text, with defined scope for what counts as a high-risk AI system, would mark a substantive first for Serbia's data protection regime and would be the natural point at which a more assertive characterisation of this development becomes appropriate.

Algorithmic, Biometric & Surveillance Governance

The draft Law on Personal Data Protection published for consultation on 30 July 2026 is reported to introduce the most detailed treatment of artificial-intelligence processing that Serbia's data protection framework has had to date. The draft is understood to distinguish three categories of system: high-risk systems, which would trigger a mandatory data protection impact assessment together with a prior opinion from the Commissioner before deployment; systems that interact directly with individuals; and minimal-risk systems, which presumably carry a lighter compliance burden. This tiered structure echoes the risk-based approach familiar from EU-level AI governance discussion, applied here specifically within a data protection instrument rather than a standalone AI law.

A further reported feature of the draft is a restriction on the legal basis available for training artificial-intelligence systems: the draft would permit training on a legitimate-interest basis only in exceptional cases, which would narrow the default pathway many AI developers currently rely on for training-data processing in less prescriptive regimes. If adopted as reported, this would place a meaningfully higher compliance burden on AI training activity touching Serbian personal data than currently exists, requiring either a different lawful basis or qualification for the exceptional-case carve-out before legitimate interest could be relied upon.

The draft is also reported to add a dedicated chapter addressing video surveillance, a form of processing that the current 2018 law does not appear to regulate with comparable specificity. The reported chapter covers surveillance of business premises, residential buildings, and public areas, and is understood to impose a general prohibition on monitoring employees in the workplace, subject only to narrow exceptions. Retention limits are a central feature of the reported chapter: footage itself would be capped at six months' retention, while records of access to that footage would be retained for one year. The combination of a general employee-monitoring prohibition with specific, relatively short retention ceilings suggests the drafters are treating video surveillance as a processing category warranting its own tailored proportionality rules, rather than leaving it to be governed by the law's general processing principles alone.

Taken together, the AI risk-tiering provisions and the video-surveillance chapter represent the two most substantial reported additions to Serbia's data protection framework under the draft law, and both sit within governance of automated and surveillance-adjacent processing specifically, rather than in the law's general provisions. This is consistent with a broader regional and EU-adjacent trend of treating algorithmic and surveillance processing as categories requiring dedicated statutory attention beyond general data protection principles, rather than folding them into existing lawful-basis and data-subject-rights provisions without modification.

It bears repeating that none of this is yet in force. The draft remains at the post-consultation stage, with consultation having closed on 10 September 2026 and no confirmation of further legislative progress. The reported provisions on AI risk tiers and video surveillance should therefore be read as the government's stated direction for reform, subject to revision before, if, and when the draft is formally submitted and adopted.

Outlook

The central question for this governance area is whether the AI risk-tiering and video-surveillance provisions survive from draft to enacted text in their currently reported form, and on what timeline. Given that these are the draft's most novel substantive additions relative to the 2018 law, they are also the provisions most likely to attract stakeholder comment during any further legislative process, which could affect their final shape. Confirmation of government or parliamentary submission, when it comes, will be the next concrete signal of how these provisions are likely to evolve before taking effect.

Algorithmic, Biometric & Surveillance Governance

The draft Law on Personal Data Protection, published for consultation on 30 July 2026, is understood to introduce a tiered classification of AI-driven personal-data processing that has no equivalent in the 2018 Act. High-risk systems would trigger a mandatory data protection impact assessment and require a prior opinion from the Commissioner for Information of Public Importance and Personal Data Protection; a second tier covers systems that interact directly with individuals; a third covers minimal-risk systems. The draft is also reported to require labelling of synthetically generated or materially modified audio and video content, and to confine processing of personal data for AI training on a legitimate-interest legal basis to exceptional cases only. None of this is yet in force, and the final enacted text and its effective date are pending.

The draft is further reported to add a dedicated chapter on video surveillance, an area commentary describes as likely to require organisations' first immediate review of existing practice once the law takes effect. The chapter would cap footage retention at a maximum of six months, generally prohibit monitoring of employees' work performance subject to narrow exceptions, and ban remote, mass and indiscriminate biometric identification carried out through video-surveillance systems. These provisions are absent from the Law on Personal Data Protection currently in force.

Separately, the draft's definition of profiling is understood to drop the word "automated" that appears in the current formulation, redefining profiling instead as any form of processing used to assess certain characteristics of a natural person. This is a broader definition than the one in the 2018 law, potentially bringing manual or hybrid evaluative processing within the profiling concept where it previously fell outside it.

Outlook

Each of these three strands -- AI risk-tiering, video-surveillance governance and the broadened profiling definition -- depends on the draft Law's progression past consultation. Organisations operating video-surveillance systems or AI-driven evaluative processing in Serbia face a materially different compliance architecture if the draft is enacted substantially as published, but confirmation of enactment, timing and final wording is still pending.

Enforcement & Redress

Servia's current enforcement ceiling for data protection breaches is settled and well below the levels associated with the GDPR. Under the 2018 Law on Personal Data Protection, the maximum fine for legal-entity controllers and processors is approximately EUR 17,000, equivalent to RSD 2,000,000, a figure that has held since the law took effect and stands in stark contrast to the GDPR's maxima of EUR 20 million or four percent of global annual turnover. This gap has been a defining feature of the Serbian regime's practical deterrent profile since implementation.

The draft 175-article replacement law, released on 30 July 2026 and now past its 10 September 2026 consultation close, would adjust this picture, but only partially. Reports suggest the draft raises the statutory minimum fine from RSD 50,000 to RSD 200,000, a fourfold increase at the floor, and introduces a proportionality multiplier of up to twenty times the damage caused by a given violation. That multiplier, however, is capped overall at approximately five times the highest prescribed fine, which commentary places at roughly EUR 85,000. Critically, the fixed maximum fine of RSD 2,000,000 for legal entities is reported to be retained rather than replaced with a turnover-based test of the kind the GDPR uses. This means that even under the proposed reform, enforcement exposure in Serbia would remain capped at a fixed ceiling rather than scaling with the size or revenue of the entity involved, a structural choice that preserves a material gap with EU-level enforcement exposure even as the floor and proportionality mechanics tighten.

This combination, a materially higher minimum alongside a still-modest, fixed maximum, suggests a deliberate design intent to raise the penalty for minor and moderate violations without exposing large multinational controllers to GDPR-scale liability. Whether that reading holds will depend on how the provision is ultimately enacted, since the draft has not passed into law and its final form is not confirmed. The underlying penalty figures for the current law are reported with high confidence; the draft's proposed figures are reported as probable, reflecting reliance on secondary commentary describing a text that has not itself been independently verified against the Official Gazette.

Outlook

The enforcement trajectory to watch is whether the proportionality multiplier and raised minimum survive into an enacted text, and whether any future amendment moves further toward a turnover-based maximum given the gap that even the current draft would leave against GDPR levels. Absent such a shift, Serbia's penalty architecture will continue to diverge structurally from the EU model even as the rest of its substantive data protection provisions are reported to expand toward GDPR-equivalent coverage in other areas such as AI and surveillance governance. That asymmetry, broader substantive obligations paired with comparatively modest fixed-ceiling enforcement exposure, is itself the structural fact most worth monitoring going into any confirmed legislative next step.

Enforcement & Redress

The Commissioner for Information of Public Importance and Personal Data Protection holds the power to enforce its orders by threatening a company with a fine of up to ten percent of its annual income in Serbia for non-compliance. This is understood to be a significant enforcement lever on paper, though it is reported that this particular option has not yet been exercised in practice, leaving its real-world deterrent effect untested.

Separately, the current Law on Personal Data Protection sets out a distinct sanctions structure: fines of up to approximately EUR 17,000 for a legal entity and approximately EUR 1,275 for a responsible person within a legal entity, with the Commissioner also able to directly fine controllers or processors approximately EUR 850 in certain situations. These figures sit well below the scale implied by the ten-percent-of-income enforcement threat, and the coexistence of a large discretionary enforcement power alongside comparatively modest fixed statutory fines is a structural feature of the current regime.

The Commissioner's most recent annual publication, covering cases resolved in 2025, is understood to show a shift in enforcement approach: rather than framing decisions around abstract compliance discussion, the Commissioner is moving toward a fact-specific approach that weighs accountability, proportionality and necessity in individual cases. This suggests a maturing enforcement practice, though it is a qualified rather than confirmed characterisation, and it is not yet clear whether it corresponds to a rise in the number or severity of enforcement actions.

Outlook

Whether the Commissioner's fact-specific enforcement approach translates into a higher volume of published decisions, and whether the ten-percent-of-income fine-threat power is ever actually exercised, are the two enforcement-trajectory questions to watch. Both would be significant markers of a genuinely tightening enforcement posture in Serbia, distinct from the separate legislative-reform signal tracked under Regulator & Framework.