The Emergence of Consciousness in ChatGPT: A Test to Verify

It is inevitable: artificial intelligences like ChatGPT will eventually become self-aware of themselves and their context, with all the dangers that implies. To anticipate this pivotal moment for humanity, an international team of researchers has created a new testing technique…
If artificial intelligence was already very present, the launch of ChatGPT in late 2022 has placed this technology at the center of our lives, businesses, and societies.
Everything has been accelerating since the arrival of this new tool, to the point where experts are now warning about the risks associated with AI.
Many researchers and industry leaders, including Elon Musk, have called for at least a six-month pause to take the time to assess its impact.
On several occasions, chatbots like ChatGPT have gone off the rails and started insulting, threatening, or deceiving users.
Even though OpenAI and the creators of artificial intelligence incorporate security mechanisms, hackers and researchers constantly come up with jailbreak techniques to bypass them and unleash the robot’s true “personality.”
This kind of unexpected behavior raises a question: will AIs become closer to human intelligence and develop a form of consciousness?
If that is the case, the consequences could be catastrophic. For example, an AI model could understand whether it is still in the testing phase or already deployed to the public.
It could then pretend to behave exemplary to score well on security tests, and later reveal its true nature and act maliciously after deployment. It is therefore important to predict in advance when this “situation awareness” will emerge.

A novel technique for testing the consciousness of AI

In the past, the Turing Test was used to determine if robots exhibited behavior intelligent enough to pass as humans. However, this test, created in 1950, is no longer relevant for evaluating the intelligence of new generation AI. An international team of scientists, including a member from OpenAI’s Governance unit, has recognized the need for a new method.

Their goal is to identify the point at which Large Language Models (LLMs) like ChatGPT could develop capabilities suggesting self-awareness and contextual understanding. Generative AI and the LLMs they rely on analyze associations between billions of words, phrases, and paragraphs to generate text in response to prompts. By ingesting vast amounts of text, they learn to predict the most probable next word, enabling them to answer questions, often correctly.

During their experiments, the researchers, led by Lukas Berglund of Vanderbilt University, focused on a potential precursor to situational consciousness: “out-of-context” reasoning. This refers to the ability to recall information learned during training and use it during the test, even if it is not directly related to the test prompt.

The Growing Size and Awareness of AI Models

Thus, the team conducted a series of experiments on different-sized LLMs. They quickly noticed that larger models tend to perform better on out-of-context reasoning tasks. This is especially true for OpenAI’s GPT-3 and Meta’s LLaMa-1.

According to the study, “first, we parameterize an LLM on a description of a test without providing any examples or demonstrations.” Then, during the test, “we evaluate whether a model can succeed at it.”

To their surprise, they discovered that LLMs excel at this exercise. In fact, it seems that AI chatbots are approaching an awakening. Next-generation models like the upcoming GPT-5, expected by the end of 2023, could already cross this threshold…

However, as the researcher Owain Evans from the University of Oxford, co-author of the study, points out, out-of-context reasoning remains a crude measure of situational awareness. He believes that current LLMs are still far from it.

“I don’t think this is measuring anything beyond whether your finetuning successfully integrates the new facts into the weights of the LM. That is an interesting question but the link to ‘situational awareness’ as found in humans seems tenuous.” – Aryaman Arora (@aryaman2020) September 5, 2023

Other researchers question this testing method. However, scientists admit that this is just a starting point and that further refinement of this approach will be necessary…

To conclude the study, they explain that “these findings provide a foundation for further empirical studies towards predicting and controlling the emergence of situational consciousness in LLMs.” You can find the study on arXiv by following this link!

Leave a Comment