What Happened
Since July 7, 2026, every new car, van, truck and bus registered in the European Union has to come with a system that watches where the driver is looking. The requirement is called Advanced Driver Distraction Warning, or ADDW, and it is the final stage of the EU's revised General Safety Regulation. New vehicle types have had to carry ADDW since July 2024. As of last week, the exemption for models already in production expired, and the rule now covers the whole market.
In practice this means a small camera — usually mounted on the steering column, the dashboard, or near the interior mirror — pointed at the driver's face for the entire drive. Software estimates head pose and gaze direction many times per second, decides whether the driver's eyes have drifted away from the road for too long, and escalates a warning if they have. It is the first time a machine-learning system that continuously observes a human being has been made mandatory equipment in a consumer product across an entire continent.
The regulation is unusually specific about what the system must and must not do. It has to work in darkness, in glare, and in poor weather. It has to keep false alarms low. And, notably, it is forbidden from using biometric identification — the camera is allowed to see that a driver is looking down, but not to work out who the driver is.
Why It Matters
The European Commission estimates that driver distraction contributes to somewhere between 10% and 30% of all road crashes — a wide range, and probably a conservative one, because after a collision it is very hard to reconstruct where someone was looking. The General Safety Regulation as a whole is projected to prevent more than 25,000 deaths and at least 140,000 serious injuries by 2038. ADDW is one of the pieces expected to do the heavy lifting.
What makes this interesting from an AI standpoint is not the ambition but the constraint. Most consumer AI is judged on average-case performance in benign conditions. ADDW has to work on a specific human, in a moving box, under whatever lighting the sky provides, at a false-positive rate low enough that people don't rip the fuse out. Earlier attempts to infer distraction indirectly — from steering micro-corrections or lane drift — turned out to be too weak. Those signals catch drowsiness reasonably well, but a driver texting at a steady 90 km/h in a straight lane looks, to the steering wheel, perfectly attentive. Only direct observation of the eyes closes that gap.
So the regulation has effectively forced a specific technical answer: cameras plus computer vision, running on the vehicle. That is a much stronger intervention than most AI policy, which tends to describe outcomes and leave implementation open. Here, the physics of the problem narrowed the options until only one remained.
The scale is substantial. Smart Eye, one of the main suppliers, puts the affected market at roughly 15 million vehicles per year. Within a decade, driver-facing perception will simply be part of what a car is in Europe, the way anti-lock brakes and seatbelt reminders already are.
How the System Decides You Are Distracted
The technical annex is worth reading, because it shows what happens when a fuzzy human concept gets turned into something a machine can be tested against. Gaze is measured from an “ocular reference point” fixed in the vehicle's design, and the space around the driver is divided into three zones: roughly, the places you should almost never look (the roof, extreme sides), the places you should be looking (windscreen, mirrors, windows), and the places you look down at (dashboard, centre console, gear selector).
Only the third zone triggers a mandatory alert, and the thresholds are speed-dependent. At 20 km/h or above, a continuous downward gaze must raise a warning within 6 seconds. At 50 km/h or above, that shrinks to 3.5 seconds — about the time it takes to read a short message, during which a car covers roughly 50 metres blind.
The warning itself must be visual, plus acoustic or haptic — a chime, or a vibration in the seat — and manufacturers are allowed to make it escalate until the driver looks up. Manufacturers may also choose to alert on the first zone, and may set activation at a lower speed than the regulation requires.
The privacy design is the part most often misread. ADDW must run as a closed loop: it processes the video on the vehicle, keeps only what the function needs, and is explicitly barred from facial recognition or any other biometric identification of the occupants. There is no footage streaming to a regulator, and no legal basis in the regulation for one. Whether every implementation honours the spirit of that as well as the letter is a fair question, and one that national data-protection authorities will be watching.
The Reaction
Suppliers, predictably, are pleased. Smart Eye — a Gothenburg company that has been building gaze-tracking software since 1999 — marked the date with a statement from CEO and founder Martin Krantz, who called it “a landmark day for road safety in Europe” and predicted the rule would set a precedent elsewhere. Seeing Machines, its main rival, has been publishing compliance guides for heavy-vehicle manufacturers for over a year. For a sector that spent two decades selling driver monitoring as a premium option, a legal mandate is the ending you write in the pitch deck.
Drivers are less uniformly delighted, and the complaint is rarely about surveillance in the abstract. It is about beeping. European cars have accumulated an impressive number of chimes — speed limit warnings, lane-keeping nudges, seatbelt reminders — and each new one erodes patience with the whole category. An attention warning that fires while you are legitimately checking a mirror or a dashboard readout does not just annoy; it teaches the driver to tune out the sound. That is why the regulation's demand for a minimised false-positive rate is not a footnote. A distraction warning that cries wolf is worse than no warning at all, because it trains the exact inattention it exists to prevent.
Heavy vehicles complicate this further. A truck cab is a different optical and acoustic environment from a hatchback: different ocular reference point, different zone geometry, a noisier cabin, and a driver who may be many hours into a shift. The same model that works in a small hatchback has to be retuned, revalidated, and re-approved for a coach.
What Comes Next
The near-term work is unglamorous: every model still being built for the European market needs a compliant system fitted and type-approved, which pushes driver-facing cameras down into the cheapest trim levels of the cheapest cars. That, more than any research result, is what turns a technology into infrastructure.
The more interesting question is what the hardware gets used for once it is universal. A camera that can tell whether you are looking at the road can also tell whether you are ready to take back control from a partially automated system — which is precisely the problem that hands-off driving assistants have never solved convincingly. Regulators have been careful to keep ADDW's scope narrow, but the sensor is now in the car, and the temptation to reuse it for driver-readiness checks, occupant detection, or comfort features will be strong. Each of those is a separate conversation about consent, and the closed-loop guarantee that makes ADDW palatable does not automatically extend to whatever gets built on top of it.
Krantz's prediction that other markets will follow is plausible but not automatic. The United States has taken a different route, leaning on voluntary safety ratings and rulemaking rather than a single continental mandate, and its appetite for in-cabin cameras is untested.
Closing Thoughts
There is a particular kind of AI deployment that arrives without a launch event, a demo, or a benchmark score, and turns out to matter more than most of the ones that do. Nobody is going to write a paper about ADDW. It will not top a leaderboard. But from last week, tens of millions of people a year will drive cars that quietly model their attention and interrupt them when it lapses — and the system is legally required to be right often enough that they don't learn to hate it.
That is a harder standard than most AI products are held to. A chatbot that is wrong 5% of the time is a mild irritation. A distraction warning that is wrong 5% of the time is a chime that goes off every few minutes, and a driver who reaches for the volume knob. The measure of success here is not accuracy in a lab; it is whether, in five years, anyone still notices the camera is there. If ADDW works, it will disappear the way the seatbelt reminder did — from an imposition into a background hum of the machine, doing its small, dull, life-saving job.
한글 요약
2026년 7월 7일부터 유럽연합에서 새로 등록되는 모든 승용차·밴·트럭·버스에 고급 운전자 주의분산 경고(ADDW) 장착이 의무화됐습니다. EU 일반안전규정(GSR)의 마지막 단계로, 2024년 7월 신규 형식승인 차량에 먼저 적용됐던 규정이 이제 시장 전체로 확대된 것입니다. 스티어링 칼럼이나 대시보드에 달린 카메라가 운전자의 머리 자세와 시선을 실시간으로 추정하고, 시선이 도로에서 너무 오래 벗어나면 경고를 울립니다. 유럽위원회는 주의분산이 전체 교통사고의 10~30%에 관여한다고 보며, GSR 전체로는 2038년까지 사망자 2만 5천 명 이상, 중상 14만 건 이상을 줄일 것으로 기대합니다.
규정은 놀라울 만큼 구체적입니다. 시선은 차량 설계상의 '안구 기준점'을 기준으로 세 구역으로 나뉘며, 대시보드·센터콘솔 등을 향하는 세 번째 구역을 계속 응시할 경우 시속 20km 이상에서는 6초, 시속 50km 이상에서는 3.5초 안에 경고가 시작돼야 합니다. 시속 50km면 3.5초 동안 약 50미터를 시야 없이 달리는 셈입니다. 경고는 시각 경고에 청각 또는 촉각(시트 진동) 경고를 더해야 합니다. 개인정보 측면에서는 얼굴인식 등 생체인식 사용이 금지되고, 영상은 차량 내부에서만 처리·보관되는 폐쇄형 구조여야 합니다. 조향 패턴만으로는 시속 90km로 곧게 달리며 문자를 보내는 운전자를 잡아낼 수 없기 때문에, 규정은 사실상 '카메라 + 온보드 컴퓨터 비전'이라는 단일한 기술적 답을 강제한 셈입니다.
업계는 반깁니다. 예테보리의 Smart Eye는 이 규정이 연간 약 1,500만 대 시장을 열었다고 보고, 창업자 마르틴 크란츠 대표는 이날을 유럽 도로 안전의 이정표라고 평했습니다. 반면 운전자들의 불만은 감시 자체보다 '경고음'에 쏠려 있습니다. 거울이나 계기판을 정상적으로 확인하는데도 경고가 울리면 운전자는 그 소리를 무시하는 법을 배우게 되고, 이는 규정이 막으려던 부주의를 오히려 훈련시키는 결과가 됩니다. 규정이 오경보율 최소화를 명시한 이유도 여기에 있습니다. 트럭·버스는 캐빈 구조와 소음, 장시간 운전이라는 변수까지 더해져 재조정이 필요합니다. 결국 이 기술의 성공 기준은 벤치마크 점수가 아니라, 몇 년 뒤 아무도 카메라의 존재를 의식하지 않게 되는지 여부일 것입니다.
참고: Smart Eye 발표(7월 7일) · EUR-Lex ADDW 기술 요건 · Seeing Machines ADDW 해설 · European Commission GSR