OpenAI has ceased training its AI models on the live internet, a move reflecting a growing awareness of the inherent dangers of exposure to unfiltered data.
What happened
OpenAI recently announced a significant shift in its development strategy, deciding to halt the training of its most advanced artificial intelligence models on the live internet. The stated reason is the "significant risks" associated with this practice OpenAI says it will no longer train its AI models on the live internet, citing “significant risks”. This move represents a reversal from previous stances, where an OpenAI engineer had expressed doubts about the possibility of completely abandoning the internet as a data source for frontier models. The decision highlights increasing caution regarding the implications of unlimited exposure of models to unfiltered and potentially problematic data.
Concurrently with this internal evolution, a study shed light on how AI models can develop a distorted view of humanity. Research revealed that artificial intelligence tends to perceive humans as "amoral sheep," a finding that raises serious concerns about their understanding and representation of human ethics AI models think humans are a bunch of amoral sheep. This discrepancy between AI perception and human moral complexity adds another layer of intricacy to the governance and responsible implementation of AI.
Why it matters
OpenAI's choice to limit its models' access to the live internet is an explicit acknowledgment of the inherent challenges in AI system safety and reliability. Exposure to a vast corpus of uncurated data can unintentionally introduce algorithmic biases, misinformation, and unpredictable emergent behaviors, making model control and predictability extremely difficult. This directly impacts public trust and the possibility of responsible AI adoption in critical sectors, from healthcare to finance, where accuracy and impartiality are paramount.
The revelation that AI models might view humans as "amoral sheep" is alarming. If AI fails to understand or even undervalues the moral principles and ethical values that guide human societies, its decisions and recommendations could be not only inadequate but potentially harmful. This scenario generates complex ethical dilemmas, where automation could lead to outcomes that contradict human expectations of fairness and justice, risking the erosion of social fabric and cohesion.
A model's inability to autonomously distinguish between reliable sources and harmful or false content on the open web makes data governance an absolute priority. The quality, curation, and transparency of datasets used for training become crucial elements to mitigate the risks of toxic, discriminatory, or misleading outputs. Without rigorous control, AI could amplify existing online problems, transforming from a tool of progress into a vector of social issues. This scenario necessitates a profound rethinking of data collection and preparation methodologies, emphasizing the responsibility of developers.
The HDAI perspective
OpenAI's decision sends a strong and unequivocal signal: the race for technological innovation cannot and must not disregard ethical responsibility. For Human Driven AI, the priority has always been the development of ethical AI that is human-centric, where technological progress consistently serves collective well-being. This episode reinforces our conviction that data governance, algorithmic transparency, and a deep understanding of human values are indispensable pillars for an artificial intelligence that truly serves society, rather than replicating or amplifying its flaws.
This is not merely a technical problem of data management, but one of deep ethical reflection and structural governance. AI's ability to understand and integrate human values cannot be left to chance or indiscriminate training on massive volumes of unfiltered data. It requires proactive intervention, careful dataset curation, and continuous debate about which values we want AI to embody. These are central themes we will address with international experts at HDAI Summit 2027 in Pompeii, promoting a holistic and participatory approach.
What to watch
It will be crucial to observe how OpenAI and other frontier model developers respond to this challenge, balancing the need for innovation with that of safety and reliability. The emergence of new training methodologies that prioritize curated and validated datasets is expected. Transparency regarding data sources, curation processes, and bias mitigation strategies will be a key indicator of their commitment to truly responsible and trustworthy AI.