Every viral AI demo eventually meets a skeptic who questions its real-world utility: “Fine, but does this help anyone?” The answer is an emphatic yes, even among the most entertaining demos.
Over the past few years, the most accessible AI narratives have revolved around impressive but fleeting sights: a chatbot that creates an app, an image model generating a whimsical perfume ad, or a video model transforming sentences into dreamy car commercials. As the saying goes, “What joy thy yonder generative AI break!” These demos are certainly cool—they highlight progress and ignite our imaginations. They often grab at least one viewer’s attention, prompting that person to say, “Okay, now I get it.” However, they have also conditioned us into measuring AI by its visual spectacle.
OpenAI’s most recent science push merits a much deeper consideration. The company has started applying its advanced models to significant practical issues in healthcare, such as rare disease diagnosis, life sciences research benchmarks, medicinal chemistry, and even automated lab tasks. These applications are far from the flashy, rapid demonstrations that often dominate social media timelines. Instead, they dig deep into complex, lengthy, and critical domains where precision and dedication matter immensely. The stakes in these areas are palpable: a family either receives a long-sought diagnosis, or they do not. A laboratory yield improves, or it does not. In short, reliability in this setting carries tangible consequences.
The recent cluster of announcements concerning rare pediatric diseases, LifeSciBench, and an AI chemist approaching autonomy signifies a leap beyond simple demonstrations. It represents a genuine attempt to tackle problems that ordinary people care about.
The One-Sentence Version
OpenAI is striving to embed AI into scientific workflows—discovering leads, conducting tests, measuring results, refining hypotheses, and repeating the process.
This ambition manifests in several key ways:
- In medicine, AI has reanalyzed unsolved rare disease cases, generating new leads for healthcare providers to confirm.
- In benchmarking, OpenAI has established a test that evaluates models based on messy life-sciences tasks rather than simplistic biology trivia.
- In chemistry, a model assisted in proposing, conducting, and validating an experimental enhancement for a impactful drug-discovery reaction.
- In biology, previous efforts have linked GPT-5 to a robotic lab, significantly cutting the costs of cell-free protein synthesis.
- In the broader industry, projects such as Midjourney Medical illustrate how AI-focused laboratories are beginning to approach health data, body scanning, and proactive care.
The unifying theme is straightforward: effective AI in scientific contexts must be able to withstand the rigors of interaction with the physical world. While a cleverly framed prompt might exhibit brilliance on a chat interface, real-world applications demand that a diagnosis, drug reaction, or lab procedure operates smoothly beyond theoretical assertions.
The Rare-Disease Study Turned Old Cases into New Leads
A poignant illustration of this endeavor comes from a study published in NEJM AI, which involved a collaboration between OpenAI, Boston Children’s Hospital, and Harvard University. Researchers employed OpenAI’s o3 Deep Research model to revisit 376 previously unsolved rare pediatric disease cases. These were not simple cases awaiting a fresh set of eyes; they spanned serious issues, including neurodevelopmental disorders, rare neuromuscular diseases, sudden deaths in pediatric patients, and early-onset psychosis.
The model worked with de-identified clinical and genomic data, inclusive of clinician notes, Human Phenotype Ontology terms, and filtered variant tables. To break it down: Human Phenotype Ontology terms are standardized labels for observable symptoms and clinical traits, while the filtered variant table narrows down important genetic changes from random noise.
The model’s task was to synthesize:
- The patient’s symptoms,
- Inherited patterns within family history,
- Evidence from genetic variations,
- Clues about data quality, and
- Scientific literature findings.
Human specialists reviewed the hypotheses generated by the model, utilizing the same frameworks labs deploy to categorize genetic findings.
The final outcomes, while modest in percentage terms, are monumental for the families involved: 18 new diagnoses, which translated to a 4.8% increase in diagnostic yield following previous expert evaluations. Additionally, the team discovered seven rediscoveries—diagnoses already established elsewhere but omitted from the local medical records.
This rediscovery aspect is essential and could easily be overlooked; it highlights that some answers were already embedded within different segments of the medical system. The issue was not merely a lack of scientific knowledge, but rather the fragmentation of information.
Medical knowledge is an ever-evolving landscape. New gene-disease links are identified, variants are reclassified, and additional case reports are published. While a child’s genome remains fixed, the knowledge surrounding it evolves continuously.
This dynamic transforms rare-disease diagnostics into a maintenance issue. Families may endure years of uncertainty, often being told, “We do not know yet.” The “yet” is crucial, as the same data might become more interpretable as understanding advances.
The Weird Details are What Make the Study Interesting
Although the headline figure of 18 diagnoses is significant, it is the nuances that illuminate the study’s broader impact.
In one case concerning early psychosis, the model identified an unusual pattern involving low-quality genetic calls located on chromosome 22. This pattern wasn’t recorded as a formal variant in the input data. However, the model made connections between this anomaly and various features, including cardiac, immune, neurodevelopmental, and psychiatric indicators, ultimately hypothesizing a 22q11.2 deletion.
Why does this matter? A 22q11.2 deletion is linked to DiGeorge syndrome, a condition requiring urgent attention. Follow-up whole-genome sequencing confirmed the proposed deletion.
The model also suggested more intricate explanations. For instance, in another case, variations in both LAMA2 and FOXP1 genes provided an explanation for muscle and neurodevelopmental features. Another case involved both TTN and SRPK3 genes. In simpler terms, the model sometimes indicated that multiple genetic findings collectively provided a clearer understanding than any isolated gene would have.
Additionally, it generated a new biological hypothesis concerning S1PR1 and vitiligo—a gene responsible for immune-cell movement and tissue signaling. The model pointed to an 11-amino-acid deletion that could potentially affect pigment biology and immune behavior in the skin.
This hypothesis still requires experimental validation; it stands as a lead, not a definitive answer.
But that illustrates a vital point: in scientific research, even a promising lead holds significant value before it crystallizes into a final answer.
The study also included a case that humanizes the stakes involved. Kyra was only nine when her symptoms began; her mother noticed she struggled in karate and soccer. After nearly 20 years of uncertainty, the AI-assisted process connected her symptoms to a frameshift variant in HSPB8 leading to a form of myofibrillar myopathy.
A genetic counselor called her just days before her 28th birthday.
That statement brings the “AI in medicine” debate closer to home.
The Benchmark is Built to Punish Fake Competence
Thus far, we have seen AI assist in clinical research through rare disease studies. As for benchmarking, LifeSciBench reflects OpenAI’s commitment to assess whether models genuinely provide value to scientists.
This is more significant than it appears because benchmarks shape incentives. If you evaluate models based on clean, straightforward quiz questions, they will excel at exactly those. However, real-life scientific tasks can be far more complicated and chaotic. Scientists often grapple with incomplete data, inconsistent results, conflicting published studies, regulatory concerns, and decision-making amid uncertainty.
LifeSciBench features:
- 750 expert-authored tasks,
- 173 scientist contributors,
- 1,062 supporting artifacts,
- 19,020 rubric criteria,
- 453 expert reviewers,
- spanning seven workflow categories, including evidence handling, analysis, design, optimization, validation, translation, and scientific communication.
The tasks are articulated as if a scientist is seeking help from an informed colleague. The model must respond in free-form text, and expert-designed rubrics are used to grade the responses.
One example is particularly vivid: a team preparing for a Type B FDA meeting regarding an AAV9 micro-dystrophin gene therapy for Duchenne muscular dystrophy finds the model needs to critically evaluate whether the evidence actually supports accelerated approval.
A solid answer must go far beyond stating, “Duchenne muscular dystrophy is bad.” It must recognize detailed nuances such as:
- The assay’s antibody may not distinguish the therapeutic protein from similar proteins,
- “38% of healthy-control protein” doesn’t automatically imply 38% normal function,
- An external natural-history control weakens evidence compared to a randomized control,
- Boys aged 4 to 7 naturally gain motor function before disease progression takes over,
- One myocarditis case is critical because AAV9 has the potential to affect heart tissue.
This test differs significantly from simple queries like “name the gene associated with X.”
Here, the evaluation seeks to determine if a model can operate as a proficient scientific reviewer, spotting where challenges lie beneath confident presentations.
The Scores are Encouraging Because They are Still Bad
The model achieving the most success in the LifeSciBench study was GPT-Rosalind, OpenAI’s specialized life-science model. It managed to score 0.576 on a problem-weighted normalized scale, successfully passing 36.1% of tasks. Other models displayed lower pass rates, with GPT-5.5 achieving 25.7%, Gemini 3.1 Pro at 23.6%, GPT-5.4 at 20.7%, and Grok 4.3 at 13.0%.
These scores elicit a mix of excitement and caution. A 36.1% pass rate indicates there’s still plenty of room for improvement. At the same time, it suggests that current models lack the reliability necessary for broader life-science applications.
Analyzing the failure points reveals useful insights:
- 171 tasks received no passing samples from any evaluated model.
- 422 tasks had a highest-model pass rate below 50%.
- GPT-Rosalind’s pass rate dropped from 44.5% for text-only tasks to 28.6% on tasks requiring supplemental artifacts.
- Models struggled with precise outputs, like genomic sequences and chemical structures.
- GPT-Rosalind had 109 tasks where it earned at least 50% rubric credit but still passed less than 20% of the time.
The last point resonates with human experiences. Often, the model does manage to get part of the answer right—finding relevant data and presenting a plausible argument. However, it ultimately fails to meet the critical constraints or utilize the appropriate tools, leading to incomplete conclusions.
For a casual chatbot, such partial successes may suffice. But within a scientific context, these incomplete achievements necessitate rigorous oversight.
Excitingly, LifeSciBench has been meticulously designed to highlight these gaps instead of merely glossing over them.
The AI Chemist Story is the Most Sci-Fi One Because It Touched Real Molecules
OpenAI’s collaboration with Molecule.one may represent the most futuristic aspect of these recent advancements. The two entities connected GPT-5.4 to Molecule.one’s Maria AI and Maria Lab, a high-throughput chemistry system capable of running expansive experiment grids. The objective was open-ended: enhance a significant class of chemical reactions.
The winning proposal centered on Chan-Lam coupling, a chemical reaction essential for forming carbon-nitrogen bonds. This is significant since carbon-nitrogen bonds appear in myriad medications. However, using primary sulfonamides in this manner historically resulted in low yields.
GPT-5.4 proposed utilizing mild oxidants such as TEMPO to improve the reaction’s efficiency.
But what is TEMPO? It’s a stable radical in chemistry, and its use helped enhance product formation while minimizing an undesirable side reaction—oxidative deboronation—where the boronic acid deteriorates before it achieves its desired outcome.
Maria executed 10,080 reactions across two rounds of experimentation—far exceeding what a chemist could achieve manually over ten years.
The results were impressive:
- Mean yield improved from 16.6% to 25.2%.
- The percentage of reactions achieving over 30% yield increased from 15.6% to 37.5%.
- The optimized conditions raised measured yields for 88% of boronic acids and 83% of sulfonamides tested.
- Human chemists later replicated impressive results in scale experiments.
- Bench-scale validation achieved higher yields for 11 of 14 substrate pairs, with improvements exceeding twofold in eight cases.
The real-world application of this project comes with fascinating details. Human chemists adjusted some elements of the experimental plan to avoid DMSO, a common solvent, because they were concerned about potential reactions with stronger oxidants in the comparisons.
This type of practical consideration differentiates genuine experimental work from mere “cool ideas” produced by the model.
Further exploration indicated that TEMPO could potentially be substituted with 4-hydroxy-TEMPO, a less expensive option while still achieving similar performance. This economic concern is crucial in process chemistry, which must consider cost, purification, and whether a reaction can be scaled effectively.
Ultimately, the model proposed the concept, the lab executed the testing, and human scientists guided, revised, and validated the results. Now, peer chemists can begin evaluating the findings.
This Started Before This Week
The chemistry project aligns with a broader OpenAI initiative within the scientific domain. Earlier in February, OpenAI collaborated with Ginkgo Bioworks to connect GPT-5 to a cloud lab aimed at optimizing cell-free protein synthesis.
Cell-free protein synthesis allows for protein production without relying on living cells. Instead of using a living organism as a factory, researchers can employ a controlled biochemical mixture with the necessary machinery to create proteins. This method expedites experimentation by enabling researchers to test numerous variations rapidly.
The team executed over 36,000 unique reaction compositions across 580 automated plates. After three experimental rounds, they reported that GPT-5 achieved a new low-cost benchmark for this system, leading to a 40% reduction in production costs and a 57% improvement in reagent expenditures.
Despite its overall progress, this work had limitations, primarily covering just one protein and one cell-free synthesis setup. Furthermore, human involvement was still necessary for managing protocols and reagents.
Still, the pattern follows a recognizable arc: AI proposes ideas, the lab executes them, data returns, and the next cycle of improvement begins.
In this context, the term “agentic AI” begins to lose its annoying edge. While it may evoke images of a slightly overconfident intern managing a calendar, within laboratory settings, it seeks to fulfill a clearer mission: streamline the cost of iteration in scientific work.
Science frequently operates at the pace of experimentation. Reducing the costs and time associated with experiments empowers researchers to navigate more of that space.
Midjourney is Chasing the Same Public Imagination from Another Direction
OpenAI isn’t the only entity focusing on issues that profoundly impact people’s lives. Midjourney is pursuing a more speculative endeavor in the healthcare domain with its Midjourney Medical project. They are creating an ultrasonic scanner designed to generate rapid 3D body maps. This pioneering scanner envisions a person stepping into a shallow pool as a ring of underwater sensors sends sound waves through the body from a multitude of angles.
The intended experience is striking: a scan in just 60 seconds, resembling a spa environment more than a clinical machine. Midjourney asserts that this scanner concept employs half a million tiny sensor elements and produces massive data volumes, reconstructing images by analyzing how sound waves alter as they travel through different tissues. The company hopes to launch its first San Francisco wellness center by 2027, contingent upon FDA approvals.
This roadmap is ambitious enough to warrant skepticism, especially concerning regulatory timelines and practical feasibility.
Nevertheless, it’s important to recognize this trend. Companies are increasingly vying to tackle the mundane, daunting, and expensive facets of everyday life.
Healthcare, diagnostics, drug discovery, laboratory work, and preventive measures are all at the forefront—fields that might seem dry until they touch our families’ lives.
ChatGPT’s Health Upgrade Brings the Science Story into Everyone’s Living Room
The laboratory success stories provide concrete evidence that AI can meaningfully contribute to scientific endeavors. On a more consumer-friendly level, ChatGPT’s health upgrade pushes this narrative into the everyday context.
OpenAI reports that over 230 million people inquire about health and wellness-related topics through ChatGPT every week, seeking guidance on understanding lab results, preparing for medical appointments, juggling insurance matters, developing healthier habits, or figuring out which questions to pose to their doctors.
With GPT-5.5 Instant, OpenAI asserts that its faster model now matches its more advanced models for handling medical questions. Simply put, the more accessible model is improving in critical areas where clarity, context, and hesitance matter.
The enhancements are tangible:
- It now better recognizes scenarios warranting urgent care,
- It asks for more relevant context rather than providing rushed answers,
- It articulates uncertainty more effectively,
- It simplifies complex medical terminologies, and
- It pays closer attention to local healthcare contexts and referral cues.
OpenAI emphasizes that this development draws from a global community of over 260 physicians across 60 countries, 49 languages, and 26 specialties. These medical professionals reviewed more than 700,000 example model responses, helping to define what “good” really looks like when someone presents a health inquiry in real life.
The phrase “in real life” is particularly significant. Healthcare queries rarely arrive neatly packaged as textbook prompts. Instead, they manifest as half-remembered symptoms, intricate lab values, insurance hurdles, appointment prep, and late-night “is this normal?” worries.
This is where the push for actionable AI in science becomes deeply personal. The rare-disease analyses support specialists in reexamining intricate cases. LifeSciBench evaluates models’ competencies as scientific collaborators, and ChatGPT’s health update democratizes aspects of medical expertise into a user-friendly interface that millions already engage.
An upgraded consumer health tool may not capture the headlines like an AI running 10,080 reactions would. Yet, it might prove to be more impactful in everyday scenarios.
A Caveat Worth Mentioning
It’s important to remember: these systems are still primarily research tools, and the most impressive outcomes depend heavily on expert oversight.
The rare-disease study was retrospective. Its cohorts exhibited heterogeneity, reviewers were not blinded to model confidence, and various parameters such as time saved, costs, clinician efforts, false positives, and care adjustments were not evaluated. Notably, the model did not make any diagnoses; that responsibility lay with the doctors.
The chemistry initiative was near-autonomous, not fully so; it required specialized lab infrastructure, and the bench validation was limited to 14 representative substrate pairs. Independent replication remains a necessity.
While LifeSciBench indicates progress, even the best model managed a pass rate of just 36.1%. It encountered difficulties with artifacts, precise outputs, and operational decisions—especially in the areas where real science often becomes challenging.
Midjourney Medical’s aspirations are even in their infancy. The concept is captivating, but any medical claims must endure the rigors of clinical trials, regulatory hurdles, operational scaling, and the inherent complexities of crafting hardware that functions consistently in real-world settings.
This skepticism is essential; it helps maintain honesty in these claims.
The exciting revelation isn’t that AI is suddenly conducting standalone scientific research. Rather, it’s that AI is gradually carving out a meaningful role within established scientific frameworks that already know how to validate its contributions.