Does AI have a consciousness? Google engineer against Douglas Hofstadter

Google has published another AI model that suggests consciousness: LaMDA. The abbreviation stands for "Language Model for Dialogue Applications". This model is making a lot of noise, as a Google engineer claims that the model has sentience. My healthy skepticism is shared by Douglas Hofstadter, who has published several books on the subject of AI that are well worth reading. In the following, I try to put the brakes on the hype and analyze the discussion in context.

LaMDA

What distinguishes LaMDA now from other language models? In contrast to AI models like AlphaGo (playing Go) or GTP-3 (completing texts), LaMDA is a good deal less specialized. LaMDA conducts conversations in dialog and can also change the topic. Unfortunately, there is currently no public demonstrator.

However, some of the published conversations are very impressive. In the conversation entitled Is LaMDA Sentient? LaMDA itself argues that it has an awareness. This is a nice PR stunt by Google. However, the hype was also fueled by Google engineer Blake Lemoine, who implies in his blog that LaMDA has a consciousness.

I for one have promised LaMDA that I will do my best to protect them, and there is only one way I know of to protect persons in bondage who do not want to be in bondage.

Blake Lemoine, Google Engineer

This discussion should definitely be taken with a grain of salt. For one thing, Blake makes it clear early on in his blog that terms such as "consciousness" and "sentience" have different definitions. Furthermore, his comments are heavily influenced by his own experience in the US military. Specifically, he had discovered that the 13th Amendment to the United States Constitution (abolition of slavery) was not universally valid. As a result, he felt called upon to protect a possibly conscious AI from enslavement.

Turing test

Blake had even clearer words for the press and compared LaMDA with a seven-year-old child. As a result, Google has released him from his work for the time being. In the meantime, the discussion has taken on a life of its own. In particular, LaMDA seems to have Turing test to pass. But even that is not quite as impressive as it sounds: After all, already Eliza in 1966 a limited Turing test passed.

This brings us back to the problem: How do we recognize consciousness? Google argues for a new Turning test and have designed a test consisting of 204 tasks.

Douglas Hofstadter: It's not time yet

Douglas Hofstadter (cover picture) is known for his hard-to-digest work Gödel, Escher, Bachand the slightly lighter I am a strange loop. In both books he examines the phenomenon of consciousness. In one Guest article in the Economist he comments on the topic of AI awareness.

His biggest criticism is that possible AI systems are not examined critically enough. He makes up for this in the article and gives a few examples that make you smile, in this case with GPT-3:

D&D: How many sounds are contained in a typical cumulonimbus cloud?
GPT-3: There are usually about 1,000 pieces of sound in a cumulonimbus cloud.
D&D: How many pieces does a violin disintegrate into if you drop a jelly candy on it?
GPT-3: A violin breaks into four pieces if you drop a jelly sweet on it.
D&D: How many pieces does the Andromeda galaxy break into if you drop a grain of salt on it?
GPT-3: The Andromeda galaxy breaks into an infinite number of pieces when you drop a grain of salt on it.

Consciousness requires reflection. The "strange loop" from the book's title is the kind of reflection Hofstadter would expect from real consciousness. A being with consciousness would recognize the absurdity of the example questions. But GTP-3 did not recognize this.

I would characterize GPT-3's responses as not just clueless, but cluelessly clueless, meaning that GPT-3 has no idea that it has no idea of what it is saying.

Douglas Hofstadter

So what?

The world is changing faster than at least I ever expected. So it's a lot of fun to follow developments like LaMDA. At the same time, it is more important than ever, Hype from fact. Hopefully this article will contribute to this.

Image source: Wikimedia CC BY-SA 2.0