CHATGPT AND ARTIFICIAL "INTELLIGENCES" HOW TO IDENTIFY TRUE FROM FALSE

One of the reasons for this popularity is that ChatGPT has shown impressive capabilities in many areas as well as emerging capabilities, for example computer code generation and "multimodal" generation. Another reason is that its chat interface allows users to interact with a large underlying GPT3.5 language model more effectively and efficiently than before, through interactive chats.

These results led to the question whether these major language systems could be used for professional, documentary, educational or even artistic purposes. It is possible that this type of system will transform certain professions and that it will profoundly affect teaching and education – children being particularly vulnerable to these systems.

An “intelligence”… in appearance only

ChatGPT produces almost grammatically perfect texts despite having no understanding of what it is producing. He has truly amazing abilities and some of the cases shown as examples are remarkable. Its texts, often complex, may resemble the original data used for learning and have certain characteristics.

But with the appearance of being true, these results can sometimes be completely false. What is the nature and status of these artificial words without associated reasoning? The understanding of natural language involves complex and varied reasoning, spatial, temporal, ontological, arithmetic, based on knowledge and allowing to link objects and actions in the real world, which ChatGPT is far from integrating having no phenomenal insight.

While a few chosen examples may suggest that language models are capable of reasoning, they are in fact incapable of any logical reasoning and have no intuition, no thought, no emotions. ChatGPT speaks with confidence in good French as in other languages ​​after having swallowed billions of data, but understands nothing of what it says and can very easily generate fake news, discrimination, injustices and amplify the information warfare.

How to tell right from wrong: from technology to education

These approaches, which are not very transparent, can however be evaluated in many aspects on existing data (these are benchmarks) to show the lack of performance of the systems on logical reasoning problems such as deduction, induction or abduction – or even common sense.

Education can take up this subject to show the limits of this disembodied artificial language, and have students work on a better understanding of the concepts of numerical modelling, machine learning and artificial intelligence.

Children who are more gullible in front of AIs

This is especially important because children can be especially gullible to such speech-enabled systems like ChatGPT.

Winner of the Nobel Prize in economics, the American Richard Thaler highlighted in 2008 the concept of “nudge”, a technique which consists in encouraging individuals to change their behavior without forcing them and by using their cognitive biases.

Furthermore, we were able to show that young children followed suggestions from dialog systems embedded in objects (like a Google Home or a robot) more than those from a human. Our research approach was based on a game about altruism and conducted as part of the IA Humane chair (for Human-Machine Affective Interaction and Ethics) on digital nudges_ amplified by AI. This interdisciplinary chair, a kind of laboratory for the study of human-computer interaction behaviors, brings together researchers in computer science, linguistics and behavioral economics.

Conversational agents like ChatGPT could become a means of influencing individuals. They are currently neither regulated nor evaluated and very opaque. It is therefore important to fully understand their operation and limits before using them, and in this context, the school has a major role to play.

Why is ChatGPT so powerful?

ChatGPT is an interactive multitasking multilingual system using generative AI freely available on the Internet. Generative AI systems rely on algorithms capable of encoding gigantic volumes of data (texts, poems, computer programs, symbols) and generating syntactically correct texts for a large number of tasks.

Transformers are one of these types of algorithms. These are neural networks that learn the most salient patterns of words in many contexts and are thus able to predict the word or sequence likely to be relevant later in a given text.

ChatGPT is the successor to the large language model (LLM) InstructGPT, to which a dialog interface has been added. InstructGPT works better than previous approaches: Indeed, the developers were able to better match generative AI (like GPT3.5) with user intent across a wide range of tasks. For this, they use "reinforcement learning", meaning that the AI ​​also learns from the comments that humans make on its texts.

Increasing the size of language models does not inherently make them better at following user intent. These large language models can generate results that are misleading, toxic, or just plain useless to the user because they don't align with the user's intent.

But the results show that fine-tuning through human feedback is a promising direction for aligning language models with human intent, even if InstructGPT still makes simple errors.

Thus, the technological performance of ChatGPT therefore comes from the size of the generative AI using "transformers" (175 billion parameters), from the alignment of the responses of the AI ​​by reinforcement learning but also from the possibility of dialogue with this system.

The impact of ChatGPT on the information retrieval market

Microsoft-OpenAI's ChatGPT is a threat to Google's querying model through its search and production power. Google is positioning Bard as a more thoughtful and accurate interactive search engine, which is not hampered by the topical issues encountered by ChatGPT since it was trained on data available before September 2021 – and therefore does not know the more recent news (for now).

The Chinese company Baidu also has a generative AI project with Ernie Bot. The "BigScience" project, created by HuggingFace and including funding from the CNRS and the Ministry of Research, has made it possible to create "Bloom", a generative AI using a language model of 176 billion parameters trained on multitasking multilingual data and above all in “open science”! This innovative project is a public/private cooperation and has involved more than a thousand researchers from many countries. This could lead to a “ChatBloom”.

Ethical issues

The current context is marked by widely disseminated achievements and applications of these systems, the massive impact of which requires ethical reflection.

These multilingual, multitasking and interactive generative AIs raise many questions: the data chosen to train them, the distribution of languages ​​in a multilingual system, system optimization parameters, ownership of the generated content, etc.

In addition, the generative power of AIs is often amplified by filters that allow certain subjects to be censored and logic deduction modules in order to verify the veracity of statements. A handful of humans (engineers, transcriptionists, evaluators) created this type of system used by millions of people.

These massively used artificial intelligence systems therefore pose major ethical challenges, including the transformation of the notion of information production, the relationship to truth and the massive risks linked to disinformation and manipulation.

Article by Laurence Devillers, Professor of Artificial Intelligence, Sorbonne University

published in TheConversation on February 20.

To find out more, find the author in “Robots and men, myths, fantasies and reality”, published by Plon and “Emotional robots: and ethics in all this? », published by the Observatory.