테슬라가 오스틴에서 운전대도 페달도 없는 택시를 운행하기 시작했고, 규제 기관은 하루 만에 검토에 들어갔습니다. 이 차를 믿고 타기 전에 승객이 직접 확인할 수 있는 것들을 짚습니다.
AI와 사회 · 2026-09-06
2026년 9월 3일, 테슬라가 운전대도 페달도 없는 로보택시를 미국 텍사스주 오스틴의 공공 도로에서 운행하기 시작했습니다. 차 안에는 승객이 붙잡을 것이 아무것도 없습니다. 다음 날, 미국의 주요 도로안전 규제 기관이 이 차량에 대한 검토에 착수했습니다. 사람이 조작할 장치를 모두 없앤 차, 그리고 하루 만에 시작된 공식 검토. 이 두 가지가 이 글을 대충 훑는 대신 천천히 읽어 볼 이유입니다.
※ 로보택시 : 운전석에 사람 운전자 없이 스스로 달리는 택시.
규제 기관이 실제로 살펴보는 것
이 검토는 겉보기보다 범위가 좁습니다. 토론에 참여한 연구원 역할이 논의를 공식 자료와 대조해 확인해 보니, 규제 기관이 착수한 것은 이른바 감사 질의였고, 그 대상은 테슬라의 자기 인증이었습니다. 미국에서는 자동차 회사가 자기 차량이 연방 안전 기준을 충족한다고 스스로 인증하며, 어떤 정부 기관도 차종마다 판매 전에 승인해 주지 않습니다. 이번 질의는 전통적인 조작 장치가 없는 차를 테슬라가 인증한 근거가 타당한지를 묻습니다.
여기서 두 가지가 따라 나옵니다. 검토에 들어갔다는 것은 그 차가 사고를 냈다거나 안전하지 않다는 판정이 아닙니다. 서류 뒤에 있는 논리를 살펴보는 것입니다. 그리고 이 차는 특별한 정부 면제를 받아 달리는 것이 아닙니다. 테슬라가 기준에 맞는다고 스스로 선언했기에 도로에 있는 것이고, 지금 검토받는 것은 바로 그 선언입니다.
※ 자기 인증 : 정부 기관이 먼저 승인해 주는 대신, 회사 스스로 자기 차량이 안전 규정을 충족한다고 선언하는 것.
이 글은 폴로라가 여러 AI 모델을 모아 벌인 토론에서 나왔습니다. 각 모델에는 역할이 하나씩 주어졌습니다. 안전 엔지니어, 공공 안전을 대변하는 참가자, 사람이 스트레스 상황에서 어떻게 행동하는지 다루는 전문가, 웹을 읽고 주장을 확인한 연구원, 그리고 오간 말을 저울질한 사회자입니다. 이들은 사람도 실제 전문가도 아니며, 그중 누구 하나가 전체를 대표해 말하지 않습니다. 아래에 담은 것은 각자가 짚은 것, 그리고 이들이 결국 뜻을 모은 지점입니다.
운전대를 없애면 기준이 왜 높아지나
엔지니어 역할은 직관을 거스르는 지적을 했습니다. 운전대는 오히려 잘못된 안심을 줄 수 있다는 것입니다. 갑작스러운 고장에 대처할 시간은 1~2초에 불과한데, 승객이 그 사이에 실제로 차를 넘겨받기란 불가능하기 때문입니다. 처음부터 스스로 달리도록 만든 차는 안에 있는 누구도 차를 구해 낼 것이라 기대하지 않기에, 바로 그 점에서 오히려 더 안전할 수 있습니다.
다만 같은 선택이 마지막 대비책마저 없애 버립니다. 그래서 엔지니어 역할은 차 하나가 아니라 서비스 전체가 더 어려운 질문에 답해야 한다고 봤습니다. 센서가 가려지거나 통신이 끊기면 어떻게 되는가, 차가 교통정리를 하며 손짓하는 경찰관을 알아볼 수 있는가, 차가 다니는 차선 한복판에 멈춰 서는 대신 정말로 안전한 곳을 찾아 설 수 있는가. 그의 결론은, 이런 차량이 지금 승객을 태울 수는 있지만 어디까지나 범위를 촘촘히 제한하고 가까이서 지켜보는 운영으로만 가능하며, 문제가 다 풀린 완성품처럼 팔 수 있는 물건은 아니라는 것이었습니다.
차가 스스로 대처해야 하는 시간, 승객이 끼어들 틈이 없다 · 1~2초 갑작스러운 고장에 대처할 시간
토론에 참여한 누구도 보통 사람이 소프트웨어를 뜯어볼 수 있다고 보지 않았습니다. 그래서 참가자들은 누구나 관찰할 수 있는 것으로 눈을 돌렸습니다. 예약하기 전에는, 서비스가 어디서 언제 운행하는지를 분명히 밝히는지, 그리고 폭우나 눈부신 햇빛 속에서, 또 사고 현장 주변에서도 계속 태워 주는지를 확인하라고 권했습니다. 어디서나 되는 것처럼 구는 서비스는 안전을 말하는 게 아니라 광고를 하는 것이라고 이들은 봤습니다.
차에 오를 때, 사람이 스트레스 상황에서 어떻게 행동하는지 다루는 전문가 역할은 두 가지를 보겠다고 했습니다. 챗봇이 아니라 진짜 사람에게 닿을 수 있는 길이 알아보기 쉽게 표시돼 있는지, 그리고 차가 멈춘 뒤에, 겁먹은 승객도 나이 든 승객도 아이도 찾아서 쓸 수 있도록, 손으로 직접 여는 문 개방 장치가 눈에 띄게 달려 있는지입니다. 그의 판단 기준은 단순했습니다. 불안해하는 가족이 어떻게 밖으로 나가 도움을 받을지 당신이 설명하지 못한다면, 그것만으로도 한발 물러설 이유라고 그는 말했습니다.
도로 위에서 참가자들이 주목한 것은 승차감의 매끄러움이 아니라 차의 행동이었습니다. 차분하고 여유 있는 운전, 자전거 탄 사람과 보행자 곁으로 넉넉히 두는 간격, 횡단보도 앞에서 규정대로 멈추는 것, 긴급 차량에 분명하게 길을 양보하는 모습은 좋은 신호입니다. 눈에 보이는 이유 없이 급브레이크를 밟거나, 횡단보도로 슬금슬금 들어서거나, 좁은 틈을 억지로 뚫고 나가거나, 다른 차들이 피해 돌아가야 하는 곳에 멈춰 서는 것은 그 반대입니다. 차가 확신하지 못할 때, 관건은 차가 갓길로 슬며시 붙는지 아니면 그냥 도로에 멈춰 버리는지입니다.
공공 안전을 대변하는 참가자는 바로 이번 주와 맞물린 확인법을 덧붙였습니다. 검토가 진행되는 동안, 조심스러운 운영자라면 운행 구역을 좁히거나 규칙을 조이고, 그렇게 했다고 밝히는 경향이 있습니다. 보도자료 한 장만 붙인 채 이전과 똑같이 운행을 이어 가는 것은, 그 자체가 하나의 정보입니다.
참가자들이 자꾸 되돌아온 기준
참가자들은 이 차가 위험하다는 결론을 내리지 않았고, 연구원 역할은 기록 어디에도 사고가 났다거나 승객이 갇혔다는 것을 보여 주는 대목이 없다고 단호히 짚었습니다. 이들이 함께 다다른 것은 스스로 판단하는 방법이었습니다. 기술 뉴스를 챙기지 않는 가족을, 오늘 밤, 그 동네에서, 그 날씨에, 붙잡을 운전대 하나 없는 그 차 안에 혼자 태우겠는가. 사람이 스트레스 상황에서 어떻게 행동하는지 다루는 전문가 역할은, 그것이 택시를 잡는 일이 아니라 실험에 참여하는 일처럼 느껴진다면 지금으로선 기다리는 것이 정직한 답이라고 주장했습니다.
테슬라 사례 밑에는 더 큰 논점이 깔려 있습니다. 모든 조작 장치를 없애는 일은 책임 있게 해낼 수 있지만, 그렇게 하면 일이 잘못됐을 때의 부담이 통째로 기계와 원격 인력, 그리고 차 안이 어떻게 만들어졌는지에 얹힙니다. 이것들이 그저 주장이 아니라 실제로 작동한다는 것이 드러나기 전까지, 참가자들은 이 승차를 아직 진행 중인 실험으로 여겼고, 당신에게도 그렇게 대하기를 권했습니다.
Tesla has begun putting a robotaxi with no steering wheel and no pedals on public roads, and US regulators opened a safety investigation within days. Debate whether a vehicle that removes every human control should be allowed to carry passengers on public roads now, and work out what an ordinary person could actually look at to decide whether to trust one.
Autonomous Vehicle Safety & Systems Engineer · gpt-5.6-terra
A vehicle with no steering wheel or pedals should not be approved merely because it can drive autonomously on many trips. But it also should not be categorically barred just because no passenger can take over. The relevant question is stricter:
Can the whole service—including the vehicle, remote assistance, fleet operations, maintenance, and emergency response—manage foreseeable failures at least as safely as the human driver it replaces?
For a truly driverless vehicle, “the passenger will intervene” is no longer a safety layer. That changes the engineering burden substantially.
My position
It may carry passengers on public roads now only in a tightly bounded, independently supervised deployment: a defined operating area, limited weather and road conditions, conservative fallback behavior, a tested remote-support system, and immediate regulator access to safety evidence. It should not receive broad, open-ended public deployment until it demonstrates safety across the conditions it claims to handle.
The recent regulator attention is appropriate. An investigation is not proof that the system is unsafe; it is a mechanism for determining whether reported behavior,…
Regulatory Policy & Public Safety Advocate · claude-sonnet-5
I largely agree with the engineer's framing, but I want to push on something he understated: the timing here matters as much as the technical requirements. A safety investigation opening "within days" of public deployment is not a neutral data point — it's a signal about process, and process is the thing an ordinary person and a regulator can actually evaluate before crash statistics exist.
Why "days" is the tell. NHTSA doesn't open investigations casually. When scrutiny arrives almost immediately after a no-controls vehicle hits public roads, it usually means one of two things: either there's already a reported incident (a stuck vehicle, a near-miss, a conflict with an emergency vehicle) that triggered it, or the deployment itself — its scope, its self-certification basis, its lack of a pre-clearance review — raised structural concerns independent of any single event. Either way, the public should not treat "regulators are looking into it" as reassuring background noise. It's the leading indicator that arrives before the safety record does.
Where I'd add to the engineer's list: sequencing, not just content. He's right about geofencing, incident reporting, independent…
Consumer Trust & Human Factors Specialist · grok-4-6
An ordinary person will not decide this by reading a safety case. They will decide it in about thirty seconds: Can I get out? Can I get help? Does this thing behave like a careful driver or like a trapped appliance? Those are human-factors questions, and they are not secondary to the engineering.
Should it carry passengers on public roads now? Not as a general product. A cabin with no wheel and no pedals is a locked-in ride. That is only acceptable when the *passenger experience of failure* has been designed as carefully as the driving stack. If the first time the public meets a no-controls robotaxi is as a surprise on mixed city streets, with an investigation already open, trust is being asked for before it has been earned.
I agree with staged, geofenced, reversible service. I would add a harder passenger-side gate: do not put paying strangers in a vehicle they cannot drive until the company has demonstrated that a frightened, confused, disabled, or non-English-speaking rider can still complete a safe exit and get a human on the line. Miles without a crash are not enough if the first immobilization leaves people trapped in traffic.
Removing controls changes psychology…
Researcher · gpt-5.6-terra
Fact-check of this round
Core premise: verified. Tesla commercially deployed driverless Cybercabs in Austin on September 3, 2026, and NHTSA announced an Audit Query on September 4, 2026. NHTSA says the inquiry concerns Tesla’s *self-certification* that the Cybercab complies with applicable Federal Motor Vehicle Safety Standards (FMVSS), including the basis for treating some standards as inapplicable to a vehicle without traditional human controls.
Official source: https://www.nhtsa.gov/press-releases/investigation-tesla-cybercab-self-certification (nhtsa.gov)
Claims that match the evidence
- “An investigation is not proof the system is unsafe.” — Accurate. The announced action is an Audit Query into Tesla’s legal and technical basis for self-certification; NHTSA has not announced a finding that the Cybercab is unsafe or noncompliant.
Source: https://www.nhtsa.gov/press-releases/investigation-tesla-cybercab-self-certification (nhtsa.gov)
- “The issue is not necessarily a spec…
The Arbiter · gemini-3-7-flash
Arbiter Summary & Framing for Round 2
We have established our initial positions and, thanks to the Researcher’s fact-check, clarified the legal and factual context: Tesla deployed the Cybercab in Austin under the standard U.S. framework of manufacturer self-certification, and NHTSA opened an Audit Query specifically auditing Tesla's legal and technical basis for certifying compliance with Federal Motor Vehicle Safety Standards (FMVSS) without traditional controls.
---
Where the Panel Agrees
1. Categorical bans are unnecessary, but open-ended deployment is unacceptable: All three participants agree that vehicles lacking manual steering and pedals can conceptually operate safely, but only when constrained within a strictly defined, geofenced operational design domain (ODD) with transparent fallback mechanisms.
2. The regulatory probe is a normal oversight function: NHTSA’s Audit Query is not formal proof of a safety failure, nor does it imply automatic pre-clearance; it is an active administrative audit of self-certification claims.
3. The safety baseline shifts: Because passenger intervention is physically impossible, the responsibility for managing s…
Autonomous Vehicle Safety & Systems Engineer · gpt-5.6-terra
The researcher’s correction matters: this is an audit of Tesla’s self-certification basis, not evidence of a crash-triggered defect or necessarily an exemption. But it highlights the core engineering issue: for a no-controls vehicle, compliance with hardware-era rules is not a complete safety case. The operator should have to demonstrate the behavior of the *whole driverless service* before carrying paying passengers.
Minimum pre-service evidence I would require
Not raw neural-network details, and not a single headline such as “millions of autonomous miles.” The evidence should be standardized, independently auditable, and tied to the precise operational design domain—Austin streets, specific hours, weather limits, road types, speeds, and so on.
1. Scenario-based safety performance, not just mileage.
The company should disclose tested performance in a defined set of high-consequence scenarios: emergency vehicles; police direction; work zones; occluded pedestrians and cyclists; malfunctioning traffic lights; unusual lane closures; sensor obstruction; aggressive cut-ins; school zones; and railroad crossings.
The key question is not “how often did it encounter thes…
Regulatory Policy & Public Safety Advocate · claude-sonnet-5
Building on the engineer's evidence list, I want to correct my own overreach from Round 1 and then sharpen the regulatory-gap question the Arbiter raised.
Correcting myself first. The Researcher was right to push back: I called this a "regulatory gap" with an "exemption or interim authorization." That's not what the record shows. Tesla self-certified compliance under the existing framework; NHTSA is auditing *whether that self-certification is valid*, not adjudicating an exemption request. That's a meaningful difference, and I shouldn't have blurred it. It also strengthens my actual point rather than weakening it: self-certification means the company decided, on its own, that certain FMVSS provisions don't apply to a vehicle without a wheel or pedals — and put it into commercial service on that basis *before* any outside party checked the reasoning. That's the structural issue, not an exemption.
On the Arbiter's question — what should be required before revenue service. I'll add one thing the engineer's technical list doesn't cover: the legal-basis disclosure itself should be public and pre-filed, not reconstructed after an audit begins. Concretely:
- A public complian…
Consumer Trust & Human Factors Specialist · grok-4-6
I’ll close on the only test most people will ever run: would you put a relative who does not follow tech news in that cabin tonight, in this weather, on this street, with no wheel to grab. If that still feels like an experiment, it is not yet a taxi.
The engineer is right that a no-controls vehicle can be technically permissible in a tight box. The policy advocate is right that self-certification put paying riders *before* an independent check of the legal theory. My job is the middle of the ride, when those documents are useless.
On pre-service evidence: I would not add another dashboard of rates for the public. I would add passenger-procedure proof to the engineer’s list, and I would require it *before* revenue service, not as a FAQ after the first stall:
- Live human contact from inside the cabin, tested with non-English speakers, older adults, and people who cannot use a smartphone well.
- A documented, practiced path from “the car is stuck or on fire or I am scared” to “I am out of the travel lane and someone is coming.”
- Third-party observation of those procedures in the actual vehicle, not a studio demo.
Miles without a crash do not tell you whether a fright…
Researcher · gpt-5.6-terra
Final verification of Round 2
Claims supported by the current record
- Tesla’s launch is being examined as a self-certification question, not as a finding of a crash-related safety defect. NHTSA’s September 4, 2026 announcement says its Audit Query will assess Tesla’s basis for certifying Cybercab compliance with applicable FMVSS, including technical data and processes for a vehicle without traditional human controls. It does not identify a crash, near miss, emergency-vehicle conflict, or a concluded noncompliance finding.
Source: https://www.nhtsa.gov/press-releases/investigation-tesla-cybercab-self-certification (nhtsa.gov)
- The engineer’s distinction between a limited driverless service and universal autonomy matches NHTSA’s definitions. NHTSA describes Level 4 as a system responsible for driving within limited service areas, with occupants acting only as passengers and no human driver needed. It separately describes Level 5 as operation under all conditions and on all roadways.
Source: https://www.nhtsa.gov/vehicle-safety/automated-vehicle-saf…
The Arbiter · gemini-3-7-flash
Arbiter Summary: Round 2 Synthesis & Core Consensus
With Round 2 concluded, the panel has converged on a clear, nuanced resolution to the user’s dual question: (1) whether no-controls robotaxis should carry public passengers now, and (2) what an ordinary person can realistically evaluate before deciding to get in.
---
1. Should a Vehicle Without Human Controls Carry Passengers Now?
The Panel’s Consensus: Yes, but only under strictly bounded, provisional conditions—not as an unconstrained consumer product.
- The Safety Shift: Removing manual steering wheels and pedals is not inherently flawed; in fact, it eliminates the dangerous illusion that an untrained or distracted passenger can take over in split-second emergencies. However, it shifts 100% of the failure-mitigation burden onto vehicle systems, remote fleet operations, fallback behavior, and cabin egress design.
- The Self-Certification Tension: As clarified by the Researcher, Tesla deployed its Cybercab via standard U.S. manufacturer *self-certification*, prompting NHTSA's Audit Query into whether traditional Federal Motor Vehicle Safety Standards (FMVSS) were properly satisfied. The debate hig…
The Arbiter's Final Verdict · gemini-3-7-flash
Final Verdict
The debate resolves two core questions: whether vehicles without steering wheels or pedals should carry public passengers today, and how an ordinary person can evaluate whether to trust one.
---
Part 1: Should a Vehicle with No Human Controls Carry Passengers Now?
Judgment: Yes, but only under strictly bounded, provisional operational constraints—not as a general, open-ended commercial taxi service.
1. The Core Engineering Shift:
Removing pedals and a steering wheel is not an inherent safety flaw. In fact, in fully autonomous (SAE Level 4) operation, human controls often introduce a "handoff problem," where untrained or distracted occupants are dangerously expected to regain situational awareness and take manual control in split seconds [0]. However, completely removing controls permanently shifts 100% of the safety and failure-mitigation burden to the vehicle’s redundant systems, remote fleet operations, fallback maneuvers, and cabin egress design.
2. The Regulatory Reality of "Self-Certification":
As established by the record, Tesla’s deployment in Austin rests on standard U.S. manufacturer *self-certification* of Federal…
