AI Polygraphs and Lie Detection Under the EU AI Act: High-Risk, Not Prohibited
AI polygraphs are high-risk under EU AI Act Annex III points 6(b) and 7(a), not prohibited like workplace emotion recognition. The €15M vs €35M line explained.
No, AI polygraphs are not banned. An AI system used as a polygraph or similar tool — one that detects a person's emotional state or tests the reliability of their statements — is classified high-risk under the EU AI Act (Regulation (EU) 2024/1689), not prohibited under Article 5. The trap is conflating it with the outright ban on workplace and education emotion recognition. They are different limbs of the law, and the difference decides whether you face a €15 million ceiling or a €35 million one. This page covers AI deception-detection and truthfulness-assessment tools in the law enforcement and migration contexts where they are expressly named.
AI Polygraphs Are High-Risk — They Are Not a Prohibited Practice
State it plainly: AI "polygraphs and similar tools" sit in Annex III, the high-risk list, not in Article 5, the prohibited-practices list. The Act names them twice, in two adjacent contexts:
- Annex III point 6(b) — law enforcement.
- Annex III point 7(a) — migration, asylum and border control management.
Both provisions expressly name polygraphs and similar tools. The same device class — a tool inferring emotional state or scoring the truthfulness of statements — appears under each, tied to a different operating authority.
Where people go wrong is the Article 5(1)(f) prohibition on emotion recognition in the workplace and in educational institutions. That ban targets a context, not a technology. Point the identical capability at a criminal suspect or an asylum applicant, with a competent authority operating it, and the legislator chose to regulate the system as high-risk rather than ban it. Context and actor decide; the underlying technology does not.
The classification is not academic. High-risk non-compliance carries a fine of up to €15 million or 3% of total worldwide annual turnover, whichever is higher (Article 99(4)). A breach of an Article 5 prohibition reaches €35 million or 7% (Article 99(3)). Getting the line right is the difference between two penalty regimes.
Set the deadline expectation now: the statute originally read 2 August 2026 for stand-alone Annex III high-risk systems. A deferral to 2 December 2027 has now been adopted by the Digital Omnibus — the section at the end explains why.
What "Polygraphs and Similar Tools" Means in the Statute
The phrase covers two functions a device may perform: (i) detecting the emotional state of a natural person, and (ii) testing the reliability of a person's statements — deception detection. A tool that does either is in scope; many do both.
The point 6(b) wording (law enforcement)
Annex III point 6(b) captures polygraphs and similar tools used by, or on behalf of, law enforcement authorities — or by Union institutions, bodies, offices or agencies in support of those authorities. The trigger is the function plus the operating actor.
The point 7(a) wording (migration, asylum, border)
Annex III point 7(a) captures the same device class used by, or on behalf of, competent public authorities (or Union bodies) in the context of migration, asylum and border control management. An automated truthfulness assessment of a traveller or asylum applicant lands here.
Examples: voice-stress, micro-expression and deception detection
Concrete instances include voice-stress analysis, micro-expression and "deception detection" systems — the EU-funded iBorderCtrl-style avatar interview is the canonical example — and physiological-signal truthfulness scoring. The functional test, not the branding, governs. A vendor that markets a "credibility assessment" or "integrity interview" tool which infers truthfulness from biometric or behavioural signals is in scope regardless of the label it chooses.
Many such tools also process biometric data, which overlaps with Annex III point 1 biometrics. But for these use cases the polygraph limbs of points 6(b) and 7(a) are the controlling classification.
The Actor Requirement: Who Uses It Decides the Classification
Points 6 and 7 are not free-standing technology bans. Each attaches high-risk status only to a specified actor: use by or on behalf of law enforcement (point 6), or by or on behalf of migration, asylum and border-control authorities (point 7). Strip the actor away and the point does not bite of its own force.
The phrase "on behalf of" is load-bearing. It captures private contractors and vendors operating the system for those authorities. A deception-analytics company cannot escape scope by pointing out that it is not itself a state body — if it runs the tool for a border agency, it operates within point 7.
This cuts the other way too. A truthfulness or emotion tool deployed by a private actor for purposes outside law enforcement, outside migration and border, and outside the prohibited workplace and education settings may fall under a different Annex III category, a transparency duty, or none at all. That is a fact-specific analysis, not a green light — do not assume "private deployment" means "unregulated".
Role mechanics matter as well. Under Article 25, a deployer, distributor or importer becomes the provider of a high-risk system if it puts its name or trademark on the system, makes a substantial modification, or changes the intended purpose so that the system becomes high-risk. A border agency that re-badges a vendor's tool, or repurposes a general emotion-analytics product into a deception screen, can inherit the full provider stack.
The Dividing Line: Prohibited (Article 5) vs High-Risk (Annex III)
The whole subject turns on one principle: the EU AI Act classifies by deployment context and operating actor, not by the underlying technology.
Article 5(1)(f): emotion recognition in the workplace and education is prohibited
Article 5(1)(f) prohibits AI systems that infer the emotions of a natural person in the workplace and in educational institutions, save for a narrow medical or safety carve-out. This ban has applied since 2 February 2025, with no grace period. An employer running emotion-inference on staff, or a school doing so on pupils, is in prohibited territory today.
Article 5(1)(g): biometric categorisation inferring sensitive attributes is prohibited
Article 5(1)(g) prohibits biometric categorisation systems that infer sensitive attributes — race, political opinions, trade-union membership, religious or philosophical beliefs, sex life or sexual orientation — subject to narrow exceptions. This is the adjacent prohibition to keep distinct. A poorly scoped "truthfulness" tool that drifts into inferring such attributes can cross from high-risk into prohibited.
Why lie-detection in law enforcement / migration stays high-risk, not banned
The same emotion-inference or truthfulness technology pointed at a suspect by a police authority, or at an applicant by an asylum authority, is high-risk under points 6(b) or 7(a) — regulated, but lawful. The legislator made a deliberate choice to regulate rather than ban it in those contexts. The lesson repeats one limb over: Article 5(1)(d) bans predicting the risk of a criminal offence based solely on profiling or personality traits — yet recidivism assessment grounded in objective facts stays high-risk. High-risk and prohibited sit right next to each other; the boundary is precise.
| Tool and context | Classification | Governing provision |
|---|---|---|
| Emotion inference in the workplace | Prohibited | Article 5(1)(f) |
| Emotion inference in education | Prohibited | Article 5(1)(f) |
| AI polygraph / deception detection used by law enforcement | High-risk | Annex III point 6(b) |
| AI polygraph / deception detection at a border or in asylum screening | High-risk | Annex III point 7(a) |
| Biometric categorisation inferring sensitive attributes | Prohibited | Article 5(1)(g) |
| Emotion recognition in a retail / consumer setting | Not banned; disclosure duty | Article 50(3) |
The principle behind the table: the device is the same; the deployment context and operating actor decide prohibited versus high-risk versus disclosure-only.
Obligations Once a Polygraph Tool Is High-Risk
These are stand-alone Annex III systems. The full Articles 8–15, 16 and 43 stack applies directly — it is not routed through Annex I sectoral product legislation. Keep this section tight; the law enforcement and border-control deep dives walk the obligations in full.
Provider duties (Articles 9–15, 43, 47, 49)
The provider must operate a risk management system (Article 9); govern data so that the disparate-impact, false-positive and false-negative risk in deception scoring is addressed (Article 10); compile Annex IV technical documentation (Article 11); supply transparency and instructions to deployers (Article 13); design in human oversight (Article 14); and meet accuracy, robustness and cybersecurity requirements (Article 15). Conformity assessment follows the internal route under Article 43 for stand-alone Annex III systems, followed by the EU Declaration of Conformity (Article 47) and registration (Article 49).
Deployer duties and the Article 27 FRIA
Deployers carry the Article 26 duties — use in line with instructions, human oversight, monitoring, logging. Public-authority deployers must additionally conduct an Article 27 Fundamental Rights Impact Assessment before deployment. The FRIA is central here given the presumption of innocence, the right to a fair trial, non-discrimination, and asylum and non-refoulement rights at stake.
Article 49 non-public database registration
Law enforcement and migration systems register in the non-public section of the EU database under Article 49, reflecting the operational sensitivity. Providers register the system; public-authority deployers register their deployment separately.
Worked Example: A Border-Screening Deception-Detection Pilot
Veridexa, a 140-person deception-analytics company, sells an avatar-interview tool that scores travellers' truthfulness from micro-expressions and voice signals. A national border agency pilots it at a land crossing, using the score to flag travellers for secondary questioning.
Classification. The system performs both functions named in the statute — testing the reliability of statements and detecting emotional state — and the operating actor is a competent border-control authority. That places it in Annex III point 7(a): high-risk. It is not Article 5(1)(f), because a border crossing is neither a workplace nor an educational institution.
Watch-out. If the tool also outputs inferences about a traveller's ethnicity or religion, it risks tipping into the Article 5(1)(g) prohibition on biometric categorisation of sensitive attributes. Veridexa's design must avoid sensitive-attribute inference, full stop, or the system moves from a €15 million regime to a €35 million one.
Provider obligations. Veridexa compiles Annex IV documentation with accuracy metrics disaggregated across nationality and language groups; runs a risk management system that treats high false-positive harm — a truthful traveller flagged as deceptive — as a primary hazard; completes internal conformity assessment under Article 43; issues an Article 47 Declaration of Conformity; and registers the system in the non-public database under Article 49.
Deployer obligations. The border agency runs an Article 27 FRIA covering non-discrimination, asylum-seeker rights and human review. Officers must never treat the score as dispositive — it is one input, never the decision. The agency registers the deployment in the non-public database and retains logs per Article 26.
Timeline. Both Veridexa and the agency plan against the adopted 2 December 2027 deadline (the statute originally read 2 August 2026, before the Digital Omnibus deferral).
Deadlines and the Digital Omnibus Caveat
The Article 5 prohibitions — including the 5(1)(f) workplace and education emotion-recognition ban — have applied since 2 February 2025. They are enforceable now, with no transition.
The stand-alone high-risk Annex III obligations covering the polygraph use cases read in the statute as applying from 2 August 2026.
A deferral of stand-alone Annex III to 2 December 2027 has been adopted under the Digital Omnibus, with the European Parliament voting on 16 June 2026 and the Council adopting it on 29 June 2026. Entry into force follows Official Journal publication, expected before 2 August 2026. "Stop the clock" was rejected; not everything was delayed.
A tangent for completeness: the Article 50 transparency and content-marking duties (from 2 December 2026) apply to non-prohibited emotion-recognition deployments — for instance a consumer-facing tool — not to the law enforcement and migration polygraph use cases that are the subject of this page.
How Confir helps
Confir's classification engine runs your system's intended purpose and functional description through the Annex III points 6(b) and 7(a) polygraph criteria and the Article 5(1)(f) and 5(1)(g) prohibition boundaries, and returns a documented classification finding. The engine is deterministic and rule-based — no model inference, no hallucination — so the same input yields the same output every time, with the rule that fired shown in plain language. That reproducibility is what an audit needs.
For public-authority deployers, Confir's Article 27 FRIA workflow structures the assessment across the Charter rights most at stake — presumption of innocence, fair trial, non-discrimination, asylum and non-refoulement — and logs it alongside the deployment record for the non-public Article 49 registration. For providers, the Annex IV documentation pack covers the required fields, including disaggregated accuracy metrics and false-positive risk analysis.
Related guides
- Annex III point 6 law enforcement AI
- prohibited workplace emotion recognition
- Article 5 prohibited practices
- Annex III point 7 border and migration AI
- high-risk AI system requirements
Manage your EU AI Act compliance in one place
Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 14-day free trial.
Start free trial →