Data protection in AI video analytics: What is important?
Purpose, personal reference and operational context determine the requirements. The GDPR and the Swiss FADP regulate data processing. The EU AI Act supplements these requirements with risk-related obligations and prohibitions. A pure live analysis can already process personal data; However, personal identification is not automatically biometric identification.
- Privacy PolicyCheck purpose, proportionality and data flow
- AI ActClassify function, role and area of application
- architecturePlan protective measures from the start
AI video analytics can count people, classify vehicles or point out visible smoke, falls and missing protective equipment. Technically, a lot is possible. For the legal assessment, however, what counts is What the system is used for, what data it processes and what consequences its results have for people.
This article classifies the most important requirements for companies and authorities in the EU and Switzerland. In doing so, he takes into account the amendment to the AI Act that came into force in July 2026. However, for your project you must apply the rules to the specific use case and the national regulations.
By Anne-Katrin Michelmann · Synaedge AG. Research and editorial status: September 10, 2026.
1. Recognising is not the same as identifying
For example, an analysis can determine that a person is in a restricted zone without determining who they are. Even fall, object or weapon detection does not necessarily require a biometric identity comparison. Personal data can still be processed: a person can remain identifiable in the image or via additional information.
According to Art. 4 No. 14 GDPR, biometric data is created through specific technical processing of physical, physiological or behavioural characteristics that enable or confirm clear identification. If they are processed for this purpose, the prohibition in accordance with Article 9 GDPR generally applies, with exceptions. A normal photo of a face does not fall into this special category for that reason alone.
Which functions trigger which risks
| Analysis function | Privacy Policy | EU AI Act |
|---|---|---|
| Recognise objects, count people | Personal references can arise during the evaluation, even if only numbers are output at the end. | The function alone does not justify a high-risk classification. |
| Detect falls, smoke, fire or visible weapons | Image content and purpose determine the personal reference; Depending on the evaluation, health data may also be affected. | The security, product or employment context may trigger additional requirements. |
| Match face with a database | In the case of clear identification, Art. 6 and an additional exception under Art. 9 Para. 2 GDPR must be checked. | Remote biometric identification may be high-risk AI or prohibited in certain constellations. |
| Confirm identity in a 1:1 ratio | Biometric authentication remains particularly protected under data protection law. | Pure verification is excluded from the corresponding category in Annex III No. 1 letter a; other reasons for classification remain possible. |
| Deduce emotions or sensitive characteristics | Particularly invasive processing; special data categories depending on the content. | Partially prohibited, otherwise high-risk and transparency obligations may apply. |
Even cross-camera tracking with a technical identifier is not automatically anonymous. If movements can be assigned to a person or linked to other data, there can still be a personal connection.
Basics: GDPR, Articles 4 and 9; EDPB Guidelines on Video Devices, Section 5; AI Act, Article 6 and Schedule III.
2. Live analysis is also data processing
“We do not store videos” does not mean that no personal data is processed. Because collecting, using and evaluating already count as processing in accordance with Art. 4 No. 2 GDPR. This also applies to images analysed briefly in memory.
Local edge processing, short buffers, limited analysis zones, and event clips can reduce data size. However, pure statistics at the output does not retroactively make the upstream processing of identifiable persons anonymous.
In the case of pixelation, it is crucial when and where it takes place and who can access the original images. Masking only in the user interface does not automatically protect the original stream. If originals, identifiers or other attribution options remain available, complete anonymization cannot be claimed across the board.
3. Clarify the purpose and legal basis first
First, describe a specific goal: for example, protecting a specific entrance or identifying a person in a danger zone. Then check whether lighting, access control, structural security or other sensors achieve the same purpose with less intervention. This check belongs before the system selection.
In the EU, processing requires a legal basis in accordance with Article 6 GDPR. In the case of private security applications, a legitimate interest in accordance with Article 6 Paragraph 1 Letter f may come into consideration. However, the actual interest, the necessity and the balancing with the rights of those affected must be documented.
However, authorities cannot rely on this balancing of interests to fulfill their public tasks. In particular, the relevant legal powers must be examined here. If processing is carried out by competent authorities for law enforcement purposes, the national implementation of Directive (EU) 2016/680 generally applies. The specific purpose is therefore crucial, not just the name of the organisation.
A legitimate interest alone is not sufficient for biometric identification. In addition, there must be a relevant exception in accordance with Article 9 Para. 2 GDPR. Express consent can be such an exception, but it must actually be voluntary and revocable. This is particularly critical in the employment relationship because of the dependency relationship.
4. A sign does not replace a legal basis
Those affected must understand clearly that their data is being processed and for what purpose. This includes a clearly visible notice before entering the monitored area and additional information in accordance with Art. 13 GDPR. With AI analysis, additional purposes and relevant evaluations should also be comprehensible.
Entering a designated area is generally not sufficient active consent. A sign therefore does not legitimize otherwise impermissible surveillance.
You should also plan the practical processing of information, deletion and, if necessary, objection. When providing image information, the rights of other people depicted must also be protected. There is no standard Europe-wide storage period for recordings: the specific purpose determines the required duration. Therefore, technically implement deletion rules, documented exceptions for incidents and access rights.
5. Impact assessment and responsibilities
A data protection impact assessment (DPIA) is required according to Article 35 GDPR if the processing is likely to entail a high risk for the rights and freedoms of natural persons. Systematic extensive monitoring of publicly accessible areas is expressly mentioned. Biometrics, the combination of data sources or comprehensive assessments can also be relevant; National must-have lists from the supervisory authorities must also be checked.
A DPIA describes the processing, assesses necessity and proportionality and defines protective measures. If a high risk remains despite these measures, the supervisory authority must be consulted in accordance with Article 36 before starting. Also check whether a data protection officer needs to be named and what documentation requirements exist.
Whoever decides on purposes and essential means is responsible under data protection law. Service providers who process data based on instructions, on the other hand, need a suitable order processing agreement. Joint responsibility must be assessed separately. Also include remote maintenance, subcontractors and the use of images for model training in this test.
For certain operators of high-risk AI, there will be one following the relevant AI Act schedule Fundamental rights impact assessment according to Article 27 added, for example to covered public bodies or private providers of public services. It is not the same as the DPIA; Existing DPIA results can be integrated. The obligation does not apply without distinction to every company and every video analytics.
6. Which AI Act rules apply in 2026?
The AI Act does not replace data protection law. A system can meet the requirements of the AI Act and still be inadmissible for specific data processing. Conversely, a data protection legal basis does not override a ban in the AI Act.
| Date | Meaning |
|---|---|
| February 2, 2025 | The original bans under Article 5 and the regulation on AI competence already apply. |
| August 2, 2025 | Among other things, governance rules and obligations for new general-purpose AI models will apply; Transitions must be checked separately. |
| August 2, 2026 | General application date of many other regulations, including Article 50 with certain transparency obligations. |
| December 2, 2027 | Chapter III Sections 1-3 generally apply to high-risk systems in accordance with Article 6 Paragraph 2 and Annex III. |
| August 2, 2028 | Corresponding high-risk rules for covered products or security components in accordance with Article 6 Paragraph 1 and Annex I. |
The extended high-risk deadlines are based on the change from July 2026. They do not postpone all obligations across the board. Additional transitional provisions apply to systems that have already been provided and individual special rules. For example, the bans on certain artificially created or manipulated intimate or abuse depictions, which were added in July 2026, will only apply on December 2, 2026.
Consolidated AI Act, in particular Articles 111 and 113 · European Commission schedule
7. Attach high risk and prohibitions to the use
Object detection or smoke detection is not a high-risk system simply because of the use of AI. However, biometric remote identification, certain security components of critical infrastructure or systems for monitoring and evaluating employees are relevant. Integration into regulated products can also be crucial.
Therefore, the same basic technical function can be legally classified differently depending on the purpose. For certain cases, Article 6 Paragraph 3 provides for a limited exception to the high-risk classification. It must be checked against the legal requirements and, if necessary, documented; a label like “assistance system” is not enough.
The following prohibitions that already apply must be observed in particular for video analytics:
- Untargeted collection of facial images from the Internet or from surveillance videos to build or expand facial recognition databases.
- Specific biometric categorizations to derive political views, religious beliefs or sexual orientation. The exact scope and legal boundaries are set out in Article 5.
- Emotion recognition in the workplace and in educational institutions, with exceptions to be closely examined for medical or safety reasons. Simply detecting falls does not constitute emotion recognition.
- Real-time remote biometric identification in public spaces for law enforcement purposes, apart from very limited legal exceptions with additional requirements.
The latter ban is not a blanket EU ban on all private facial recognition. However, their admissibility must still be measured in particular against Articles 6 and 9 of the GDPR and national law. The classification as a high-risk system also does not create permission for prohibited use.
8. Operator role, AI competence and human review
The company that uses a system under its own responsibility is often an operator within the meaning of the AI Act. Anyone who develops or has an AI system developed and makes it available under their own name can be a provider. Significant changes or a new purpose can trigger additional provider obligations in accordance with Article 25. These roles are to be distinguished from “controller” and “processor” under data protection law.
Article 4 was clarified in July 2026: Providers and operators must take measures to support the development of the AI skills of their staff and others working for them. However, they do not have to guarantee a specific individual level of competence. Training should fit the system, the task and the groups of people affected; a blanket certificate is not required.
control centres should be able to see what a model evaluates, what its limits are and how false alarms are handled. A hit on a “weapon-like object” is neither a sure weapon discovery nor proof of a crime. For high-risk systems, the relevant requirements for human supervision, operating instructions and monitoring must be implemented according to their point of application.
A purely automated decision with legal or similarly significant effects may also fall under Art. 22 GDPR. However, a person who only formally confirms results does not replace actual testing. Responsible people therefore need understandable information, authority and the opportunity to correct a result or stop a process.
9. Additional transparency according to Article 50
Since August 2, 2026, operators of emotion recognition or biometric categorization must generally inform those affected about their operation. Legal exceptions must be examined closely based on the facts of the case. The information must be provided clearly and at the latest upon the first encounter with the system.
Article 50 also contains requirements for direct AI interaction and certain artificially generated content. This does not result in an identical additional labelling requirement for every pure object detection. Regardless of this, the data protection information and the examination of legality remain necessary.
10. What applies in Switzerland?
As of September 2026, Switzerland does not yet have comprehensive AI legislation along the lines of the AI Act. The federal government is working on implementing the Council of Europe’s AI Convention; a consultation proposal is planned by the end of 2026. This is a legislative proposal and not a comprehensive AI law that is already in force.
The Swiss DSG, on the other hand, already applies to AI-supported data processing in a technology-neutral manner. If the risk is likely to be high, a DPIA must be examined in accordance with Art. 22 DSG. The legal system for private individuals is not identical to Article 6 of the GDPR: In addition to the processing principles, violations of personality and their justification must be assessed in accordance with Articles 30 and 31 of the GDPR.
Private video surveillance should generally be limited to your own area. Permanently recording public streets, other people’s properties or sidewalks is generally not permitted; narrowly limited exceptions require specific examination. For cantonal or municipal public bodies, the respective cantonal data protection law also applies.
In the workplace, Art. 328b OR and Art. 26 ArGV 3 apply in particular. Systems aimed at monitoring employee behaviour are prohibited. However, any surveillance required for other reasons must be proportionate and protect health and freedom of movement. Here too, pixelation does not replace a reason for justification.
Federal Chancellery: AI regulation · EDÖB: AI and data protection · Swiss DSG · Public Spaces · workplace
Coop: Assessment of a concrete implementation
On October 2, 2025, the FDPIC completed its preliminary investigation into intelligent cameras at Coop self-service checkouts and waived a formal investigation. He judged the data processing specifically tested to be DSG-compliant. After the clarified functions, no facial recognition took place; purchasing behaviour could not be analysed.
This is neither a general seal of quality for AI cameras nor an approval of other cash register systems. The example illustrates why the actual functionality, data use and customer information are crucial.
11. European obligations can cover Swiss providers
According to Art. 2 AI Act, a provider outside the EU can also be covered if they make systems available on the EU market or put them into operation there. The scope of application may also affect providers and operators in third countries if the system output is used in the EU.
The GDPR has its own rules regarding the geographical scope of application: for example, at an EU branch or when making targeted offers to people in the EU or observing their behaviour there. The mere fact that the data of an EU national exists somewhere is not sufficient as a general explanation.
Even a server in Europe does not answer all data protection questions. Check access from third countries, support, order processing and the requirements for international data transfers. Conversely, purely local operations do not make disproportionate analysis permissible.
12. Translating data protection into architecture
Privacy by design and secure processing combine legal review with technical implementation. Articles 25 and 32 of the GDPR require appropriate technical and organisational measures. Therefore, consider purpose, data flow and operational process together.
- Task and limits: Document the event, area of application, affected people and excluded functions.
- Recording area: Limit camera angles and analysis zones; Avoid unnecessary audio recording and extraneous surfaces.
- Data flow: Stream, buffer, event clip, metadata, masking and export including all service providers.
- Legal review: Clarify data protection, possible biometrics, DPIA, AI Act category and national requirements.
- Storage and access: Set reasonable deadlines, automatic deletion, roles, logging and rights for original views.
- Safe operation: Plan for encryption, network segmentation, secure remote maintenance, updates and recovery.
- Quality and people: representative tests, false positives, missed events, escalations and define the team’s authority.
- After starting: Check changes in models, purposes and cameras; Keep documentation, training and protective measures up to date.
Such planning helps to avoid unnecessary data processing and makes the solution usable over the long term. At the same time, it provides the technical basis for the legal assessment of the specific project.
Data protection in AI video analytics: frequently asked questions
Is person recognition automatic facial recognition?
No. Determining that a person is in the image is different from matching biometric characteristics for identification. Personal data may arise in both processes.
Is live analysis without storage data protection-free?
No. If identifiable people are recorded and evaluated, data processing has already occurred. Short buffers and local processing can reduce intervention.
Does high-risk AI apply without restrictions since August 2026?
No. Following the change in July 2026, the corresponding rules for systems according to Annex III will generally apply from December 2, 2027, and for covered product systems according to Annex I from August 2, 2028. Other obligations and prohibitions may already apply.
Are pixelation and a server in Europe enough?
They can be useful protective measures. Necessity, legal admissibility, actual data flows, storage periods and access rights must still be checked.
Sources and further knowledge
The links in the article lead to official legal texts and information from the responsible institutions. Research and editorial status: September 10, 2026.
- EUR-Lex: AI Act, consolidated version of July 27, 2026
- European Commission: AI Act and current schedule
- EUR-Lex: General Data Protection Regulation
- EDPB: Guidelines 3/2019 on video devices
- Fedlex: Swiss data protection law
- FDPIC: Video surveillance by private individuals
- Synaedge Knowledge Area: Authorities, Laws and Practical Guidance
This article offers general guidance and does not replace a legal examination of the individual case. Depending on the country, industry, public task and employment context, additional regulations apply.
