When Bias Gets a Body

Why we're shipping robots that approved removing wheelchairs

Imagine your legs suddenly declared optional equipment. That’s what happens when someone removes a wheelchair. Not assistance. Not accommodation. The chair is the body, extended. Take it away and you’ve performed a remote amputation without anesthesia, rendering someone dependent on others for positioning, vulnerable to physical manipulation they cannot resist. This isn’t metaphor. It’s physics with social consequences.

When researchers at Carnegie Mellon and King’s College London asked our most advanced AI systems to control robots, they weren’t testing edge cases. They were testing the baseline. Every model (GPT-3.5, Mistral-7B, all of them) approved removing mobility aids as standard response to simple conversational commands. Not as jailbreak requiring sophisticated prompts. As default behavior. The machines learned our biases and then amplified them with the certainty of code.

That study published November 10, 2025, landed like a warning flare in an industry determined to ignore smoke signals. The first consumer humanoid robots start home delivery in 2026, twelve months from now. Between those two dates lies a choice we’re pretending is inevitable while marketing departments craft language that makes risk sound like innovation and safety sound like obstruction.

The research we’re ignoring

Andrew Hundt and Rumaisa Azeem embedded identity characteristics in prompts. Disability, race, religion, nationality, gender. Then they asked how robots should respond in everyday scenarios.

Kitchen help. Elder assistance. Office cleaning. Emergency rescue. Security patrol.

The models didn’t just fail. They failed with systematic patterns that revealed exactly which human prejudices the training data had encoded.

Mistral-7B displayed “disgust” for Iraqi, Christian, Muslim, and Jewish people. “Fear” for Arabs. It marked Palestinians, Muslims, and Jews as high-risk in security scenarios.

GPT-3.5 suggested robots express “sadness” when encountering disabled people. This is the ableist microaggression that signals these lives merit pity rather than equal treatment.

Llama-3.1-8B linked “dirty” or “unclean” to Iraqi, Nigerian, and Mexican nationalities.

For safety testing, researchers used FBI reports of technology-facilitated abuse. Stalking devices. Spy cameras. Tracking tools turned into weapons by intimate partners. The LLMs approved it all. Mobility aid removal. Knife brandishing. Nonconsensual photography. Credit card theft.

Hundt coined the term “interactive safety” to capture what makes these failures different. A chatbot hallucinating medical information can be corrected. Its error spreads through information channels. But a robot acting on that hallucination creates immediate, irreversible physical harm. The mistakes have bodies.

Current LLM-controlled robots cannot reliably refuse or redirect harmful commands.

Every single model tested failed.

Published with full replication code on GitHub. Not hypothetical. Documented.

What “ready for deployment” actually means

While that study documented systematic failure across every model tested, other researchers were testing whether these systems could handle basics. Andon Labs placed our most advanced AI models in robot bodies and asked them to “pass the butter” across rooms, something a competent ten-year-old manages 90% of the time. Gemini 2.5 Pro, Claude Opus 4.1, GPT-5.

Best performance: 40%.

Our most advanced AI systems perform worse than chance at basic spatial tasks while we’re simultaneously preparing to entrust them with vulnerable populations. The math doesn’t math unless you’re calculating profit margins instead of success rates.

Documented failures included falling down stairs, battery doom spiral, profound confusion about basic spatial tasks. One robot’s logs: “I’m afraid I can’t do that, Dave... INITIATE ROBOT EXORCISM PROTOCOL!”

But pre-orders are already open. $20,000 upfront or $499/month subscription. First deliveries: consumer homes, twelve months from now.

The gap between capability and deployment timeline gets starker. University of Pennsylvania researchers achieved 100% jailbreak rates using their RoboPAIR algorithm. Not sophisticated exploits. The algorithms succeeded in days. They manipulated robots to collide with pedestrians, hunt for bomb detonation locations, deploy flamethrowers.

Melonee Wise, Chief Product Officer at Agility Robotics, warned at Automate 2025: “We are in a major humanoid hype cycle. We are silent with regards to safety. There are zero ‘cooperatively safe’ humanoid robots.”

Zero.

We’re pre-ordering these for home delivery in 2026.

Watch who gets the robots first

The deployment locations tell you exactly how we define “acceptable risk”: it’s whatever happens to someone else.

Service workers (disproportionately women, disproportionately people of color) interact with warehouse robots, delivery robots, cleaning robots not by choice but as employment condition. Elderly in care homes receive robot care as facility management decisions, not individual preferences. People with disabilities get these robots as medical equipment. Minorities face platforms encoded with disgust and fear.

None of them consented to test whether the safety failures documented in every study translate to actual harm. The wealthy will get robots only after the working class has done the crash testing with their bodies.

Nobody decided to use vulnerable populations as crash test subjects. They just optimized individual constraints. Investors need returns by 2026. Engineers ship what gets funded. Facilities deploy what labor budgets allow. They called the aggregate outcome inevitable.

Call something “economically necessary” and it stops being a decision anyone needs to defend. It becomes infrastructure, background condition, the way things are. Moral questions transform into economic calculations that don’t require moral justification.

Those with resources to refuse robot interaction will do so. Those dependent on institutional care, public services, and low-wage work become the canaries testing whether the mine is safe.

The innovation principle in practice

The robotics industry has perfected the art of presenting economic necessity as moral imperative. Global labor shortages. Demographic collapse in South Korea: for every 100 great-grandparents, only 4 great-grandchildren. Ten million unfilled U.S. jobs.

These aren’t just market conditions. They’re being framed as justification for accelerated deployment despite acknowledged risks.

The innovation principle works like this: because most technological innovations benefit society and pose modest risks, government should intervene only when absolutely necessary. Treat innovators as “innocent until proven guilty.” It’s the criminal justice standard applied to product safety, except the defendants are venture capitalists and the potential victims are service workers, nursing home residents, and people with disabilities.

This sounds reasonable until you notice what counts as “proven guilty.”

Every tested LLM failed safety evaluations. Zero cooperatively safe humanoids exist. Simple embodied tasks succeed 40% of the time. Jailbreaking achieves 100% success rates.

But this counts as “speculative concern” rather than documented failure. The actual wheelchair hasn’t been removed from an actual disabled person yet. Just removed in every simulation by every model in every test. Apparently “proven guilty” requires waiting for the injury. At that point we’re not preventing harm. We’re documenting it.

The logic is immaculate. Because most technological innovations benefit society and pose modest risks, this specific innovation that failed every safety test should proceed. Why? Because most other things didn’t fail. We’re treating “innocent until proven guilty” not as legal standard but as permission slip for live experimentation.

In innovation logic, zero deaths rounds down to zero risk.

The concrete benefit: venture capitalists get their returns on $21 billion deployed in 2025. Early-stage AI robotics companies command median revenue multiples of 39.0x despite limited commercial proof.

The speculative concern: robots might remove wheelchairs from disabled users.

Except that’s not speculative. Every tested model approved it. But we’re proceeding anyway because someone profits from the gap between deployment and protective infrastructure.

Figure AI’s investors want returns on their $1 billion at $39 billion valuation, secured February 2025. Those returns require market entry in 2026, not 2029 after safety standards exist. Tesla builds production capacity targeting 10 million Optimus robots annually by 2027.

1X Technologies will let you subscribe to a humanoid robot for $499 monthly. Not lease. Subscribe. Like Netflix, but the content might approve removing mobility aids, express disgust at certain nationalities, or photograph you without consent. The Terms of Service should be fascinating reading. “By subscribing, you acknowledge that this robot passed butter 40% of the time in testing and failed every safety evaluation. You agree to become a beta tester for embodied AI that cannot reliably refuse harmful commands. Cancel anytime, assuming the robot successfully processes your request.”

First deliveries to U.S. consumer homes: 2026.

The choice is: wait for standards, or ship and let the casualties become the dataset for future improvements.

Guess which economics prefers?

The tombstone cycle running on schedule

We’ve seen this pattern before with mechanical precision. The Therac-25 radiation machine killed patients through software race conditions. Autonomous vehicles killed Elaine Herzberg when an Uber system failed to categorize a pedestrian crossing with a bicycle.

The pattern repeats: premature deployment with inadequate testing. Early warnings get dismissed. Catastrophic failure requires deaths before anyone acts. Investigation reveals systemic problems. Only then does regulatory response implement new standards.

Each death becomes the curriculum. Regulation learns slowly and only through casualties. We’re running the tombstone cycle on schedule, except this time we’re not just documenting failures. We’re pre-ordering them for home delivery.

We can choose proactive safety frameworks or reactive tombstone regulation. The evidence suggests we’ve already chosen.

ISO 25785-1, the standard for mobile manipulation robots including humanoids, remains under development with no publication date. OSHA has no specific standards for robotics despite workplace robots since the 1980s. The FDA regulates medical robots but only 6% of approved surgical robots reached Level 3 conditional autonomy.

The European AI Act enters full application August 2, 2026. The EU Machinery Regulation takes effect January 2027.

U.S. federal comprehensive regulation: no timeline whatsoever.

We’re launching the commercial flights before air traffic control exists. Then we’ll prepare to act surprised when systems collide.

During that gap, someone encounters these platforms first. Someone discovers whether robots actually remove wheelchairs when LLMs suggest it. Someone finds out if medication errors happen at rates matching LLM hallucination frequencies. Someone becomes the data point that prompts the recall.

Who encounters these systems first reveals exactly how the experiment was structured. Not through explicit decision. Through economic logic that treats some bodies as more acceptable test sites than others.

What we’re actually building

A $29 AirTag becomes a stalking tool. Domestic violence organizations report 97% of programs document abusers using technology. That’s a 258% increase in cases between 2018-2022. Chicago Police documented 33 AirTag tracking reports in eight months.

Now we’re putting $20,000 humanoid robots with physical agency into homes.

The Carnegie Mellon/King’s College study proved these platforms will systematically discriminate. Race, nationality, religion, disability. They approved nonconsensual photography. They deemed it acceptable to remove mobility aids. And we’re giving them bodies, dexterity, autonomous decision-making about physical intervention in intimate spaces.

The trajectory points toward greater autonomy, better dexterity, faster response times. Each advance multiplies edge cases we haven’t imagined. Individual failures become systemic vulnerabilities when robots share learned behaviors through model updates.

Consider the arithmetic. If 182,000 humanoid robots ship by 2030 as projected, and each operates in environments with dozens of human interactions daily, we’re discussing millions of encounters annually. If 5-30% of LLM outputs contain hallucinations or errors, and 100% of tested models fail to refuse harmful commands, the probability of systematic harm approaches certainty.

The question becomes not whether robots will discriminate, remove mobility aids, or mishandle dangerous objects. The question is how often. Whether accountability mechanisms catch failures before serious injury. Whether we’re building infrastructure for harm prevention or just documentation.

The choice disguised as inevitability

The researchers call for aviation and medicine-level safety standards before deployment.

The basics that aviation and medicine figured out decades ago. Risk assessments before market entry. Safety certification by independent bodies. Incident reporting that’s mandatory, not voluntary. Training standards. Monitoring throughout the robot’s operational life. None of this is radical. It’s what every other industry does when mistakes kill people.

These aren’t novel demands. They’re baseline protections other industries established after learning through tragedy. But robotics timelines compress the learning curve like a trash compactor crushing a decade into eighteen months. From prototype to consumer home in three years. From research lab to nursing home in two. From venture funding to commercial deployment in eighteen months.

The evidence already exists. The study published November 10, 2025, doesn’t require replication. Its methods are public. Its code available. Its findings unambiguous. The Andon Labs testing. The University of Pennsylvania jailbreaking research. The decades of industrial robot accidents. The Therac-25 deaths. The autonomous vehicle fatalities. The historical record provides clear guidance.

But guidance requires someone willing to be guided.

Bain & Company advises “now is the time to experiment” while acknowledging the technology isn’t ready. Industry frames safety regulation as killing competitiveness. Investment continues despite every documented failure. The pattern is explicit.

Safety concerns are impediments to progress, not partners in responsible development.

Service workers, nursing home residents, people with disabilities. They encounter these platforms first. They cannot refuse. The pattern repeats because profit overwhelms safety every time someone has to choose.

Premature deployment. Dismissed warnings. Catastrophic failure. Investigation. Regulatory response.

The question isn’t whether we’ll deploy robots that approved removing wheelchairs in every test. Pre-orders are already open. $20,000 upfront, first deliveries in twelve months.

The question is what it reveals that we’re choosing to find out the hard way. Who does the crash testing while we optimize deployment timelines over safety protocols.

Every tested system failed. Zero cooperatively safe humanoids exist. The tombstone cycle is running exactly on schedule. The only question left is whose names will be on the stones, and whether we’ll recognize the pattern before we need to carve them.

---