Not all artificial intelligence that shines is reliable. This adapted phrase could well summarize what an international study – with extensive Valencian participation – which was published this Wednesday, warns.
Although AI is increasingly present in everyday life and in fields such as education, science, medicine, arts and finance, this technology called to change the world does not currently seem as reliable as it should be.
Specifically, the language models it uses and which are nourished from the plethora of information that exists on the Internet. Google, Microsoft or OpenAI – with projects like ChatGPT, in the latter case – have jumped into the race, but There is still work to be done for development,
This is the conclusion of a study led by a team from the VRAIN Institute of the Polytechnic University of Valencia (UPV) and the Valencian Graduate School and Research Network in Artificial Intelligence (ValGRAAI) with the University of Cambridge, which was published Wednesday in the journal Nature.
As the UPV points out, the work reveals an “alarming” trend: compared to the first model, and taking into account certain aspects, Reliability has deteriorated in newer models (For example, GPT-4 compared to GPT-3). They also say they have confirmed that Human supervision is unable to compensate for these problemsBy trusting too much on the reliability of these applications.
The results were similar for several families of language models, including OpenAI’s GPT family, Meta’s LLAMA, and Bloom, a completely open initiative of the scientific community.
According to José Hernández Orallo, a researcher at VRAIN and ValgraI, one of the main concerns about the reliability of language models is that their operation does not fit the human perception of difficulty of the work.
“Models can solve some complex tasks according to human capabilities, but at the same time failing at simple tasks from the same domain. For example, they can solve many doctoral-level mathematical problems, but they may make a mistake in a simple addition,” says Hernández-Orallo.
Ring presents a new generation of video doorbellsHowever, the study by the UPV team, ValgraAI and the University of Cambridge shows that this is not what happened, the university says. To prove this, they investigated Three key aspects which affect the reliability of language models from a human point of view. Among other things, they assure that “there is no ‘safe zone’ in which models can be trusted to work perfectly.”
This is stated by Yael Moros Daval, a researcher at the VRAIN Institute of the UPV, who details that the published work concludes that current AI models “They are not 100% accurate even in simple tasks,
In fact, the team from the VRAIN UPV Institute, ValgraI and the University of Cambridge assures that the most recent models radically improve their performance in tasks of high difficulty, but not in tasks of low difficulty, “which increases the difficulty discrepancy between the performance of the models and human expectations”, says Fernando Martínez Plumed, also a researcher at VRAIN UPV.
any answer
The study also shows that more recent language models are more likely to provide Instead of giving wrong answers, avoid answering tasks they are not sure about. “This can lead to users who initially rely heavily on the models being disappointed later.”
“We have verified that users can do this to be affected by signals (a computer program that interprets commands and orders) that work well in complex tasks, but, at the same time, get the wrong answers in simple tasks,” says researcher and study co-author Cesar Ferri.
Also, unlike people, the tendency to avoid giving answers does not increase with difficulty. For example, humans often avoid giving their opinion on problems that are beyond their competence. It depends on the users Responsibility for detecting errors throughout all of their interactions with the models,” says VRAIN member Lexin Zhou.
But does AI become more effective depending on how questions and queries are formulated? This is one of the issues analyzed by the study, which suggests that users should not ‘free themselves’ from making effective statements to try to get the best possible answer.
“In short, large language models are becoming less and less reliable from a human point of view, and user supervision to correct errors is not the solution, because we rely too much on the models and are unable to recognize erroneous results at different levels of difficulty”, says another of the researchers.
Subscribe to continue reading
OpenAI in turmoil: ChatGPT’s ‘mother’ resigns as it transitions to a for-profit company
