Humans have always tended to overtrust computers.
Simply put, we have a long history of having misplaced and often naive faith in the next big technology.
This caused companies to stop and race to consolidation We are building next-generation generative artificial intelligence (AI) capabilities into our products and services.
The relatively early stage of AI adoption could lead to new mistakes. Generative AI differs from similar predictive AI in that it can “generate” or create new content such as text, voice, images, music, video, code, etc. completely falsified information, Something called “hallucination”.
This was announced by Microsoft-backed OpenAI on Wednesday (May 31st).Improve mathematical reasoning with process monitoring‘, revealing a new strategy for combating hallucinations.
“Even state-of-the-art models still make logical mistakes called hallucinations. Mitigating hallucinations is an important step towards building a coordinated artificial general intelligence (AGI).” state report.
“These hallucinations are particularly problematic in areas that require multi-step reasoning, as a single logic error is enough to derail larger-scale solutions,” the researchers added. rice field.
The company’s new strategy involves training AI models to reward each correct step of reasoning as they arrive at an answer, a process called “process monitoring.” This is different from the current “results monitoring” process. As the name suggests, it rewards the model after making a (presumably) correct conclusion.
This comes at a time of growing public interest and more companies adding generative AI capabilities to their products.
read more: Generative AI Hoaxes Are Already Spreading Misinformation
Overconfidence generalized from limited historical data
Some experts and industry observers already expressed suspicion about how effective OpenAI’s proposed method of combating hallucinogenic tendencies is.
OpenAI warns users not to trust ChatGPT, its generative AI chatbot, prominently disclaiming that “ChatGPT may generate inaccurate information about people, places, or facts.” it’s shown.
as PYMNT reportSeparate from OpenAI, researchers at MIT have published a new paper that they hope will improve the inference and factual accuracy of large-scale language models (LLMs).
Still, when companies back generative AI platforms Show off your model’s featuressuch as passing a medical or legal exam,extinction incidentIt’s no wonder that the average user believes that the AI is smarter than them and that the answers it provides are correct.
Related item: Will generative AI in 2023 be as transformative as indoor plumbing in 1920?
That’s because AI-generated content can be highly plausible, yet highly fallible. After all, AI models are designed to generate the statistically most likely answer, but they have no notion of meaning at all.
Hallucinations are a problem that affects all chatbots, not unique to OpenAI’s ChatGPT, but rather more broadly as a technical solution trained on historical data (the only data available in the linear world). unique to AI.
AI hallucinations follow the same trend as humanity’s most classic mistakes. They generalize overconfidently from limited historical data.
Earlier this year, in Google’s promotional video for its AI model, Bard, the chatbot made false claims about the James Webb Space Telescope that observed the tech giant. save $100 billion In terms of market value, it is a very real result.
and as PYMNTS report Recently, two lawyers could face professional sanctions after citing a case in court that was entirely a ChatGPT hallucination.
Also read: AI regulation should target data origins and protect privacy
Despite its glitches, the AI survives
according to “Preparing for a world of generative AIMay issue of “Generative AI Tracker®”, PYMNTS and AI-ID In collaboration, there is currently no easy way to distinguish between AI-generated content and authentic material.
But while hallucinations remain an endemic problem, it’s clear that generative AI models won’t be obsolete anytime soon. make a huge impact Across most, if not all, business sectors and industries.
Effective regulation of technology is therefore paramount.
Founder and CEO of AI-ID shortt sarkesian In an interview posted on Tuesday (May 30), he told PYMNTS: Regulatory perspectiveshould create an arena where everyone can compete, but with some degree of control to “ensure the technology isn’t disruptive.”
that is, Generation AI Like when HTTP was created in the 1990s and the Internet was born, we may create new data layers, and like new data layers and protocols, we need to apply governance, rules and standards, Sarkisian said. explained Mr.
Across the pond, the European Union’s top technical officials are already making rules.
European lawmakers may consider a draft AI code of conduct “in the coming weeks,” PYMNTS reports.
EU officials said the United States and Europe should work together to support a voluntary code of conduct until new legislation is enacted.
