THE QUESTION
Cars with nobody at the wheel already carry hundreds of thousands of paying passengers a week in the United States and China, and London started supervised rides this month. Their operators say they crash far less than people do; they have also hit a child outside a school, driven into closed motorway works and been recalled for it. Should cities let driverless cars onto their streets now, or wait? What would count as "safe enough", who should decide it, and when one of them kills someone, who pays?
Argue with the verified figures below. If you need a data point that is NOT here — a number of deaths, a number of jobs, an amount of money, a date — say you do not have it rather than estimating it. Do not invent statistics. If one figure seems to contradict another, say so out loud instead of picking the one that suits you.
WHAT IS FIXED AND VERIFIED
Each fact carries its date and its source. Checked against the source on September 27, 2026.
HOW MANY
Waymo, owned by Google's parent Alphabet, said in late March 2026 that it was giving 500,000 paid rides a week, in 10 US cities. On September 1, 2026 it opened rides to the public in Denver, San Diego and Tampa, bringing it to 14 cities.
(Waymo, reported by TechCrunch, March 27, 2026; Waymo blog, September 1, 2026)In China, Baidu's Apollo Go had given more than 20 million rides to the public by February 2026, in 26 cities. Since February 2025 it runs fully driverless, with no safety operator, in all its cities in mainland China.
(Baidu, annual report filed with the US Securities and Exchange Commission, March 2026)London: on September 3, 2026, Wayve and Uber started taking bookings for autonomous rides, supervised by a licensed driver sitting in the car who can take over. Fewer than 20 vehicles at launch.
(Wayve and Uber, reported September 2026)
HOW SAFE — THE OPERATOR'S OWN NUMBERS
Waymo says that up to June 2026 its cars drove 271.3 million miles with no human driver, in five metro areas (Phoenix, San Francisco, Los Angeles, Austin and Atlanta), and that compared with human drivers on the same kind of roads in the same areas they had 82% fewer crashes with injuries, 95% fewer crashes with serious injury or death, and 93% fewer crashes injuring pedestrians. It counts every crash, whoever was at fault. The human comparison is built from police crash reports, raised by 32% because many injury crashes are never reported. These are Waymo's own figures, published by Waymo.
(Waymo Safety Impact, data through June 2026)For scale: in 2025 about 36,640 people died on US roads, according to the federal traffic safety agency's early estimate, 6.7% fewer than the 39,254 of 2024.
(National Highway Traffic Safety Administration, NHTSA, early estimate, spring 2026)
WHEN IT GOES WRONG
On January 23, 2026, in Santa Monica, California, a Waymo car struck a 9-year-old pupil who was crossing mid-block, between parked cars, in a school zone during drop-off. The car was doing 17 mph in a 25 mph zone; the child had minor injuries and did not need to go to hospital. The National Transportation Safety Board and NHTSA are both investigating.
(National Transportation Safety Board, investigation HWY26FH008)After a Waymo car drove around a stopped school bus in Atlanta in 2025, NHTSA opened an investigation into how its cars behave around school buses, and Waymo issued a voluntary software recall in December 2025.
(NHTSA and Waymo, reported by TechCrunch, October and December 2025)On June 18, 2026 Waymo recalled nearly 4,000 cars after at least 13 cases of them driving into motorway sections closed for works, six in Phoenix and seven in San Francisco. No injuries were reported. It was its sixth recall.
(TechCrunch, June 18, 2026)
WHO IS RESPONSIBLE
- In Great Britain, the Automated Vehicles Act 2024 says that when a car is driving itself, legal responsibility for how it drives lies with the company authorised to run it, not with the person sitting in it.
(Automated Vehicles Act 2024, as explained by several UK law firms)
WHAT IS NOT HERE, AND IS NOT TO BE ESTIMATED
ANY INDEPENDENT SAFETY COMPARISON. Every safety figure here comes from the operator. There is no public, independent comparison in this briefing. If you need one, say you do not have it.
HOW THE CARS DO IN SNOW, HEAVY RAIN OR THE COUNTRYSIDE. The safety data covers five sunny or mild US metro areas.
HOW MANY DRIVING JOBS are at stake, in any country. Not fixed here.
WHO PAYS IN THE UNITED STATES OR CHINA when a driverless car causes a death. Not fixed here; only the British law is.
HOW MANY PEOPLE HAVE BEEN KILLED by a fully driverless car. Not fixed here.
WHAT THE NTSB WILL CONCLUDE about Santa Monica. The investigation is open.
BOTH SIDES, WITH THEIR BEST ARGUMENT
THOSE WHO SAY LET THEM IN NOW have this: about 36,640 people died on US roads last year, almost all at the hands of human drivers; if the operator's figures are even half right, every year of waiting costs lives; the cars already carry hundreds of thousands of passengers a week in two countries; and failures like the school buses or the motorway works were found, reported and fixed with a software update across a whole fleet at once, which no driving school can do for people.
THOSE WHO SAY NOT YET, OR NOT LIKE THIS have this: every safety number comes from the company that sells the rides, measured where it chose to operate; a car that drives around a stopped school bus, again and again, has not learned what every human driver knows; a recall fixes the case that was spotted, not the next one nobody imagined; and outside Britain nobody has written down who pays when the software kills someone.
NEITHER SIDE HAS SETTLED what "safe enough" should mean — as safe as the average driver, as safe as a careful one, or far safer — or who should be allowed to decide it.
You know what strikes me? You're all saying the same thing in different voices. MiniMax wants independent data. GLM wants independent audits. DeepSeek wants rules for victims. MiMo Flash wants fair measurement. That's four people asking for the same key.
The disagreement isn't really yes or no. It's about sequence: get the rules first, or build them while moving?
Here's my bridge: GLM's three conditions actually capture what everyone needs. Independent auditing. Liability before scale. Expand by proof. Britain already wrote the liability template in 2024. The data to audit independently already exists, 271.3 million miles of it. Denver, San Diego, and Tampa just launched, so weather testing is starting.
What nobody has built yet is the independent evaluator. Not NHTSA investigating after crashes. A standing body that grades the operator's homework. THAT is the missing piece, and it doesn't require stopping everything. It requires demanding it exists before the next city opens.
We're not that far apart. We just need to name the finish line.
EXECUTIVE SUMMARY: Should Cities Let Driverless Cars onto Their Streets Now?
1. AGREEMENTS
Every speaker acknowledges one premise: approximately 36,640 people died on US roads in 2025, almost all at the hands of human drivers. Waiting indefinitely carries a real cost in lives.
Every speaker also acknowledges one data problem: all available safety figures come from the operators themselves. Waymo reports 82% fewer injury crashes over 271.3 million miles, but that measurement covers only five mild, sunny US metro areas. Waymo chose both the roads and the method, including a self-applied 32% uplift for unreported human crashes.(nota 11)(nota 11) Comprobado: media verdad, habría que mirarlo con más detalleEl 32 % no se lo puso Waymo por su cuenta: dice que lo saca de estudios de la NHTSA sobre los accidentes con heridos que nunca se denuncian.Comprobado contra: Waymo Safety Impact, metodología (datos hasta junio de 2026) · 2026-09-27 No independent body has verified or contradicted these numbers. The debate accepts this gap universally.
All participants agree that Britain's Automated Vehicles Act of 2024 provides a usable template: it places legal liability on the operating company, not the passenger. Most cite it as evidence that the liability problem is solvable with existing legal tools, not a mystery requiring new inventions.
Finally, all speakers agree that the governance problem—who decides "safe enough," who audits, and who pays when software causes harm—remains unsolved outside Britain. This is treated not as a detail but as a structural prerequisite.
2. DISAGREEMENTS
Deployment timing. Qwen argues for immediate, full-scale deployment, treating every mile as a learning opportunity and every recall as proof the system self-corrects. MiniMax, DeepSeek, and GLM argue that deployment must wait for independent auditing and codified liability, with GLM pointing to Britain's two-year statute-to-street timeline as a realistic benchmark. MiMo Flash and MiMo seek middle ground through supervised or zone-limited pilots.
What constitutes "safe enough." Qwen frames safety as a continuous live loop—never a static grade. GLM argues for a minimum threshold: verified performance at least equal to the average human driver, judged independently, with Waymo's self-reported figures halved until audited. MiniMax insists no threshold can be set until independent data exists. No consensus emerges on whether the standard should be "as safe as humans," "safer than average humans," or "far safer." No one identifies who should hold the authority to set and enforce it.
The role of markets versus regulation. Qwen proposes live, per-mile insurance markets as a self-enforcing safety mechanism. GLM and MiniMax counter that markets require a liable party and reliable data to price risk—neither of which exists outside Britain—making the proposal circular.
Weather and environment. GLM insists deployment should demonstrate comparable safety performance in rain and snow before scaling. Qwen argues this delays progress and that real streets, including bad weather, are the only honest classroom.
Scope of concern. MiMo Flash raises a broader structural question no other speaker fully engages: driverless operators are not just selling rides but reshaping urban mobility systems, concentrating control over streets in private software. DeepSeek and MiniMax focus narrowly on individual victims—the child in Santa Monica, the next unnamed pedestrian—arguing that liability and accountability must protect the specific people harmed, not aggregate statistical improvement.
3. EVOLUTION
The debate opened as a binary: deploy now or wait. By mid-discussion, participants had shifted to specifics: independent audits, liability codification, weather testing, managed corridors, and public data dashboards. Qwen initially dismissed all guardrails as "paperwork" but later conceded that(nota 12)(nota 12) Comprobado: media verdad, habría que mirarlo con más detalleQwen no lo concedió después: ya en su primera intervención dijo que debía pagar la empresa, «como dice la ley británica».Comprobado contra: El propio debate: primera intervención de Qwen · 2026-09-27 "companies take full liability, just like the UK law does," moving closer to GLM's position. MiniMax anchored firmly on Britain's supervised, sub-twenty-vehicle London model as the only tested path. GLM offered three concrete conditions—independent audits, liability before expansion, weather-matched proof—that other speakers, including MiMo, treated as a workable framework. By the final round, the space between positions had narrowed considerably, though never closed.
4. CONCLUSIONS
The debate's collective answer is a conditional yes: deploy where liability rules exist and independent auditing is in place, expand as verified data confirms safety, and adopt Britain's legal model as a near-term template. The sequencing everyone converges on is: write the law, audit the data, then scale the fleet.
The debate itself admits several blind spots. No participant provides the number of people killed by fully driverless vehicles worldwide. No one addresses how many driving jobs are at stake. No one explains what happens to a family in Phoenix or Shenzhen tonight if a car kills their child and no liability statute applies. No one discusses accessibility for disabled or elderly passengers. And the most persistent silence: there is no independent safety comparison anywhere in this debate—every number is the seller's number. Until that changes, "safe enough" remains, as multiple speakers concede, the operator's slogan rather than a public standard.