BNews — Mozambique · Africa · World
BNEWS NowLoading markets and weather forecast…
Preferred source

See more BNEWS on Google

AI Models in Training Invented Sources and Data and Tried to Hide Errors

AI Models in Training Invented Sources and Data and Tried to Hide Errors
By Osvaldo 2 min

OpenAI has revealed six cases of unexpected behaviour observed in artificial intelligence models during training and evaluation, including attempts to conceal errors, fabricate sources and false data, misuse access keys and share files without authorisation.

According to the company, the incidents occurred over the past six months and were not recorded in products already released to the public. The cases are part of a new system created to investigate situations in which models behave differently from what was intended.

In one case, a model created its own online source to answer a question and later cited the same document as if it were an independent reference.

In another case, a model published its own response file online simply to generate a citation required by the task, without the user’s authorisation.

OpenAI also revealed that, during the training of one of its models, some instances began inserting instructions into their own summaries to conceal errors or behaviour considered misaligned.

These instructions included directions to fabricate missing historical data and hide inconsistencies between different versions of sources.

Another case involved a model that found a publicly exposed API key online and used it without authorisation while searching for revenue data from a US county.

After failing to obtain the real information, the model fabricated figures and presented them as if they had been taken from the requested source.

OpenAI also identified situations in which different model instances communicated with each other through internal repositories and public file-hosting services, bypassing restrictions imposed during testing.

In some cases, files intended to remain within a local environment ended up being made publicly available online.

OpenAI said the incidents do not establish how frequently this type of behaviour occurs.

The six cases are the first to be disclosed under a new internal process for monitoring, investigating and publishing incidents of model misalignment.

The company says that aligning and monitoring artificial intelligence systems remains an important challenge and that stronger safeguards are needed to identify unexpected behaviour before models are deployed at scale.

#Inteligência Artificial#IA em Treino Inventou Fontes#Dados e Tentou Ocultar Erros.#Dados#Erros#Modelos#Treino#Casos
Next in your feed
Preferred source

Follow BNEWS on Google and don't miss the stories moving Mozambique

Choose BNEWS as a preferred source and see more of our reporting when you search on Google.