ChatGPT still lies a lot and its creators know it too

OpenAI is taking it upon itself to combat the “hallucinations” of its AI in ChatGPT, the company announced Wednesday, May 31, with a new method for training AI models.

The research comes at a time when disinformation from AI systems is more hotly debated than ever, amid the AI ​​generative boom and the buildup to the 2024 US presidential election.

AI hallucinations occur when models like OpenAI’s ChatGPT or Google’s Bard fabricate information, behaving as if they’re broadcasting facts.

Some examples:

Even cutting-edge models have a tendency to produce falsehoods: they show a tendency to make up facts in moments of uncertainty. These hallucinations are especially problematic in domains that require multi-step reasoning, as a single logical fallacy is enough to derail a much larger solution.

OpenAI researchers in relation

new approach

OpenAI’s potential new strategy to fight forgery: train AI models to reward themselves for each correct individual reasoning step when they arrive at an answer, rather than just rewarding a final correct conclusion.

The approach is called “process supervision,” as opposed to “outcome supervision,” and could lead to more explainable AI, according to the researchers, as the strategy encourages models to follow more than “thinking like the human one”.

Detecting and mitigating logical errors or hallucinations of a model is a critical step towards creating overall aligned AI. The motivation behind this research is to address hallucinations to make models more capable of solving complex reasoning problems.”

Karl Cobbe, a math researcher at OpenAI, told CNBC

OpenAI has released a complementary dataset with 800,000 human tags used to train the model mentioned in the research paper, Cobbe said.

OpenAI’s progress is positive, but there is skepticism

Ben Winters, senior counsel at the Electronic Privacy Information Center and leader of its AI and human rights project, expressed skepticism, saying in an interview with CNBC that he would like to look at the full data set and corresponding examples.

“I don’t think that, by itself, significantly mitigates concerns about misinformation and erroneous results… when used in practice,” Winters said. He added: ‘It certainly matters if they are going to implement what they have discovered through this research. [em seus produtos]and if they don’t, it raises some pretty serious questions about what they’re willing to make available to the public.”

Because it’s unclear whether the OpenAI paper has been peer-reviewed or revised in another format, Suresh Venkatasubramanian, director of Brown University’s technology accountability center, told CNBC he sees the research more as preliminary observation than anything else. .

This will need to be evaluated by the research community before anything certain can be said about it. In this world, there are many results that are published very regularly and due to the general instability of how large language models work, what may work in one scenario, model and context may not work in another scenario, model and context.

Suresh Venkatasubramanian, on CNBC

Venkatasubramanian added: “Some of the hallucinatory things that people are worried about are [modelos] compilation of citations and references. There’s no evidence in this article that it works for this… I’m not saying it won’t work; I’m saying this article provides no such proof.

Cobbe said the company “will probably ship [o artigo] for a future peer review conference. According to CNBC, OpenAI did not respond to a request for comment on when, if ever, the company plans to implement the new strategy in ChatGPT and its other products.

It’s certainly good to see companies trying to fine-tune their systems development to try to reduce these kinds of errors – I think most importantly, interpret this as corporate research in light of the many barriers that exist to deeper forms of accountability.

Sarah Myers West, chief executive officer of the AI ​​Now Institute, told CNBC

West added: “[A OpenAI está] releasing a small human feedback dataset with this article, but did not provide background details on the data used to train and test GPT-4. So there’s still a huge amount of opacity standing in the way of meaningful accountability efforts in AI, even when those systems already directly impact people.”

With information from CNBC.

The ChatGPT post still lies a lot and even its creators know that it appeared first on Olhar Digital.

Source: Olhar Digital