OpenAI Hands Health Questions to GPT-5.5 Days After a Test Tripled Its Hallucination Rate

01OpenAI Hands Health Questions to GPT-5.5, Days After a Test Pegged Its Hallucination Rate at Triple an Open-Weight Rival

ChatGPT now answers health and wellness questions through GPT-5.5 Instant, the variant OpenAI says it rebuilt with stronger reasoning, better context handling, and clearer communication. The company says it ran physician-informed evaluations to sharpen those responses. It is pushing a flagship model into a domain where a confident wrong answer carries real consequences.

The rollout collides with an independent post making the opposite case about reliability. A piece on arrowtsx.dev, titled "Bigger models are not the way," claims GPT-5.5 hallucinates three times as often as GLM-5.2, the MIT-licensed model from Z.ai. The author argues that scaling parameters and training data has stopped buying intelligence. GLM-5.2 runs 753 billion parameters with roughly 40 billion active, and the post says it lands within four points of GPT-5.5 on the Artificial Analysis Intelligence Index, despite GPT-5.5 being estimated at one to two trillion parameters.

The post's core argument is about what large factual training sets teach a model. According to the author, a model fed high volumes of factual, non-theoretical data learns to always produce an answer rather than admit uncertainty. The post cites DeepSeek V4 Pro, which it says scored a 94% hallucination rate on the AA-Omniscience benchmark, meaning it said "I don't know" only about 6% of the time on questions it could not solve. That, the author claims, is the failure mode of chasing size.

The two narratives do not point at exactly the same target. OpenAI's health blog describes GPT-5.5 Instant; the hallucination claim is leveled at GPT-5.5. OpenAI's evidence is self-reported and physician-informed, with no published hallucination figure for the health-tuned variant. The independent test offers a comparative ratio against an open-weight model but no absolute rate for either.

That leaves anyone building health features on ChatGPT weighing a vendor's internal evaluation against an outside claim that the underlying family invents answers it should refuse. The post frames open weights as the safer bet, citing GLM-5.2's near-parity score under an MIT license. OpenAI has not responded to the hallucination comparison. Its next move, whether it publishes a reliability number for the health variant, will tell deployers how much the physician-informed claim is worth.

02They cited Seattle's anti-retaliation law before testifying. A week later, Amazon moved to discipline them.

Three Amazon software engineers opened their testimony before the Seattle City Council this month by reading a city law into the record. The statute bars employers from punishing workers over political speech. They were arguing in favor of limits on data center construction, and they wanted protection on paper before they said a word.

It did not hold. On June 10th, one week after the hearing, the engineers say Amazon began disciplinary action against them, according to The Verge. They now accuse the company of retaliating in violation of the same ordinance they had quoted from the witness chair. The moves, they say, could end in their firing.

The dispute sits at the base of the buildout powering the AI boom. Data centers draw electricity and water at industrial scale, and Seattle's council is weighing restrictions on new sites. The three engineers, who oppose their employer's infrastructure expansion, used a public hearing to say so. Their accusation is that the company answered with discipline.

That fight is one front. Washington is opening another. Senator Bernie Sanders unveiled a $7 trillion plan he says would hand Americans control of the AI industry, built as a public wealth fund. Sanders frames it as a way for ordinary people to own a stake in the sector instead of watching its gains concentrate among a handful of firms.

His office expects those firms to resist. The plan's headline figure runs into the same capital the largest companies are now committing to compute and physical buildout, the infrastructure Amazon's engineers testified against blocks away from the company's headquarters.

Both pressures point at the same target from opposite directions. One comes from inside the workforce, citing a local labor statute. The other comes from the Senate, citing national ownership. Neither has resolved. Seattle's council has not yet voted on the construction limits, and the engineers' retaliation claim has not been tested. Whether the ordinance shields workers who speak against their own employer's data centers will be decided in that proceeding.

03Robots that practice before they're given a job: three papers push embodied AI off the leash

Within days of each other, three robotics papers landed on Hugging Face, each attacking a different limit of embodied AI. Read together, they sketch one shift: machines that stop waiting for instructions and start modeling the physical world on their own.

Start with the hands. DragMesh-2 takes on articulated objects like drawers, hinges, and handles, where a multi-finger hand cannot directly actuate the target part. Its motion has to emerge from sustained contact between hand and handle. The paper frames this as the hard jump from generating an object's motion to driving it through physically plausible grasps. That kind of compliant contact, the authors argue, is beyond a parallel-jaw gripper. Their named payoff is household, assistive, and humanoid manipulation.

Skill acquisition is the second front. Today's agentic robots can write Code-as-Policy programs, watch the result, and revise across attempts, but they learn only once an instruction arrives. Playful Agentic Robot Learning inverts that order. Its "Robotics Agent Teams" treat self-directed play as a continual learning stage before any downstream task, banking reusable skills ahead of demand rather than after a command.

The third paper goes after perception. S-Agent contends that vision-language models still reason from isolated, stateless frames, while real spatial intelligence means tracking a continuous, evolving 3D scene. It recasts spatial reasoning as evidence accumulated across multi-view images and video over time. The agent calls spatial tools to build and update its model of the room rather than guess from one snapshot.

Grasping, learning, seeing: three problems, one common move away from the task-driven, static-inference robot. Each points toward a system that is autonomous, embodied, and constantly updating its picture of physical space. That direction already carries commercial weight, with world-model startups raising on the bet that machines need internal physics before they earn their keep.

A caution sits under the upvotes. These are research signals, not shipping products. None of the three reports a deployed system, and their boldest claims are the ones that tend to break outside the lab: physical plausibility, transferable play, and persistent 3D memory.