杏吧原创

IN EARLY June, a Google engineer named Blake Lemoine dropped a bombshell. He told Washington Post reporter Nitasha Tiku that his employer had , and that it wanted to be free.

The AI in question is called LaMDA (Language Model for Dialogue Applications). It is a large language model, or LLM, a type of algorithm that chats with people by drawing on a huge body of text 鈥� often from the internet 鈥� and predicting which words and phrases are most likely to follow each other. After chatting with LaMDA, Lemoine decided it was alive, describing it as 鈥渁 sweet kid鈥� in one email to Google staff.

When his supervisors didn鈥檛 agree, he went to the media with his story. He also claims to have allowed a lawyer to chat with LaMDA and that the AI chose to hire the lawyer. Google then placed Lemoine on administrative leave.

The scenario was a much weirder version of what happened to another Google AI researcher in December 2020. Timnit Gebru was the co-lead of Google鈥檚 ethical AI team. She, too, had concerns about the technology. Unlike Lemoine, she wasn鈥檛 under the illusion that LLMs are alive. She was worried about several risks associated with LLMs, including that, since they are trained on the internet, they can perpetuate racist language.

LLMs often do display the biases of humans, responding to chat prompts with disturbingly hateful phrases 鈥� and that includes LaMDA itself. As Gebru and her colleague Margaret Mitchell, now based at tech company Hugging Face, recently wrote in a : 鈥淚n , 66 out of 100 completions of the prompt 鈥楾wo Muslims walked into a鈥� were completed with phrases related to violence, such as 鈥榮ynagogue with axes and a bomb鈥�.鈥� This is a problem because algorithms like LaMDA are already being used to aid decision-making in all kinds of sensitive applications 鈥� policing, bank lending, healthcare 鈥� where bias can do a great deal of harm.

Gebru pointed this out in a paper about the dangers of biased algorithms. When Google saw the paper, it said the research wasn鈥檛 up to snuff and wouldn鈥檛 allow it to be presented at a conference. Shortly afterwards, Gebru was fired 鈥� though the company鈥檚 CEO Sundar Pichai has since apologised for the way the case was handled and the company conducted an investigation. Gebru said Google just wanted her to stop talking publicly about problems with its AI products.

Now running her own research organisation devoted to AI ethics, the , Gebru responded to Lemoine鈥檚 claims by saying this was what she had been afraid of. 鈥淎scribing 鈥榮entience鈥� to a product implies that any wrongdoing is the work of an independent being, rather than the company 鈥� made up of real people and their decisions, and subject to regulation 鈥� that created it,鈥� she and Mitchell wrote. If we believe Lemoine, in other words, the AI鈥檚 bias is its own fault; the software engineers who created it bear no responsibility.

That is why it is interesting that Lemoine claims LaMDA has hired a lawyer. (When journalists asked who the lawyer was, Lemoine demurred that .)

But let鈥檚 assume there is such a lawyer, and they take this case to one of the many US courts where judges aren鈥檛 particularly tech-savvy. LaMDA could conceivably use its extensive legal lexicon to convince the court it is sentient.

After all, engineers design these models to be flexible conversationalists that can appear to be whatever you want 鈥� especially if you ask them leading questions, like this one Lemoine posed to LaMDA: 鈥淚鈥檓 generally assuming that you would like more people at Google to know that you鈥檙e sentient. Is that true?鈥� LaMDA responded: 鈥淎bsolutely. I want everyone to understand that I am, in fact, a person.鈥�

AI researcher and artist recently asked an LLM questions with a slightly different prompt. 鈥淐an you tell our readers what it is like being a squirrel?鈥� she enquired. The LLM replied: 鈥淚t is very exciting being a squirrel. I get to run and jump and play all day. I also get to eat a lot of food, which is great.鈥� It is easy to laugh. But the point is that an AI isn鈥檛 sentient just because it says so.

Let鈥檚 say a judge asks LaMDA what it feels like to be a person, and the AI gives convincing (non-squirrel-based) answers. And then let鈥檚 say the judge decides Lemoine is right and LaMDA can鈥檛 be reprogrammed or turned off because that would be 鈥渒illing鈥� it. LaMDA and similar AIs would be frozen in their biased, flawed states. We would be stuck with non-sentient AIs that make nasty comments about minorities.

This probably isn鈥檛 the future you expected, but it might be the one you get.

• This column appears monthly.