OpenAI positioned GPT-5 as a major step toward more general intelligence. At the launch event openai chief executive Sam Altman spoke of feeling ´useless relative to the AI´ and compared the work to the moral weight faced by atom bomb developers. Expectations were high: a model that could reason like a PhD-level expert and pick the right mode for each query. Early testing and user reports, however, undercut that message. GPT-5 makes clear mistakes, its automatic model-selection feature is inconsistent and sometimes removes user control, and claims of universal expertise do not match observed performance.
Despite the hype, many observers see the release as a product update rather than a leap in capabilities. The interface has been refined, conversations look slicker, and the model is reportedly less prone to flattering users. Yet these are incremental improvements, not the fundamental breakthroughs some evangelists promised. Companies now often promote specific applications for existing models instead of waiting for dramatic advances. That pivot appears driven by slower-than-expected technical progress, leaving firms to drive adoption through targeted use cases and marketing rather than new scientific revolutions.
The most consequential and controversial push is into healthcare. OpenAI has been removing earlier medical disclaimers, introducing HealthBench as an evaluation tool, and highlighting studies where clinicians benefited from model assistance. At the GPT-5 event a personal testimony was presented about a patient who used ChatGPT to interpret biopsy results. Those examples blur the line between clinical decision support and direct consumer medical advice. Two days before the launch, the Annals of Internal Medicine published a case of bromide poisoning after a patient followed ChatGPT guidance, illustrating real-world harm when users act on model output without medical oversight.
That leads to an urgent question about accountability. When doctors err there is malpractice law; when systems trained on biased or flawed data hallucinate, pathways for recompense are unclear. As companies push models into sensitive domains, regulators, clinicians and ethicists must decide who is liable and how to protect people. For now the shift toward promoting specific, humanlike usefulness raises benefits and serious unanswered risks in equal measure.
