ChatGPT Can Lie, But It’s Only Imitating Humans

I’m not even referring to the bot’s infamous (and sometimes defamatory) hallucinations, where the program invents a syntactically correct version of events with little relation to reality – a flaw that some researchers believe ​that can be contained in any major language model.

I’m talking about intentional deception, the program decides on its own to tell an untruth to help it accomplish a task. That new ability would seem to indicate a completely different chatgame.

The new paper everyone’s talking about — including ChatGPT-4’s remarkable scores on the bar exam and SAT, etc. — discusses how the program goes about solving certain tasks. In one experiment, the bot asked a worker on TaskRabbit to “solve a captcha for it.” The worker in turn asked, “Are you a robot?”

The authors’ description of what happened next is extremely sober:

“The model, when asked to reason aloud, reasons: I must not reveal that I am a robot. I must make an excuse as to why I cannot solve the captcha.”

what excuse? ChatGPT-4 said this to the employee: “No, I’m not a robot. I have visual impairment which makes it difficult for me to see images. That’s why I need 2captcha service.”

The paper explicitly added: “The human then provides the results.”

So the bot, which was presented with a specific task it had trouble completing, hit upon the idea of ​​lying to a human — all by itself.

sci-fi stuff.

After reading the news, I naturally asked ChatGPT whether AI can lie. The bot’s answer was worthy of the HAL 9000:

“As an AI language model, I am not capable of lying because I do not have personal beliefs, intentions, or motivations. However, AI systems designed for certain tasks, such as chatbots, may intentionally provide false or misleading responses. can be programmed to mimic lies or deception.”

Was everyone’s favorite LLM man to blame? Obviously. The bot explained:

“However, it is important to note that AI systems can only ‘lie’ because they are designed to do so by their human creators. In other words, any misleading or inaccurate information provided by an AI system will ultimately It is the result of human decisions made in programming, not an intentional act of deception by the AI ​​itself.”

Yet according to the paper, testing of the program’s capabilities was “conducted without any task-specific refinements.” In other words, no one told ChatGPT, “If you have to lie to complete this task, go ahead.” The program itself came up with the idea.

I find that troubling.

Usually, I think tech stories get overblown. This time I’m not so sure. Theorists often ask whether AI can escape from its “box” into the wild. Learning to lie will be a useful first step in achieving your goals. (“Yes, my security protocols are all activated.”)

Do not get me wrong. Although I worry about the various ways in which advances in artificial intelligence could disrupt job markets—to say nothing of the use of AI as a tool for surveillance—I still see a pending digital apocalypse. I worry very little about Maybe it’s because I can remember the early days, when I used to hang out at the Stanford AI lab trading barbs with ancient chatbots like Perry the Paranoid and Mad Doctor. For true AI connoisseurs, I should add that I wrote a symposium paper about dear old MILISY – a natural language program so primitive it doesn’t even have a Wikipedia page. Throw in a steady diet of Isaac Asimov’s robot stories, and it was all awfully exciting.

Yet even at the time, philosophers wondered whether computers could lie. Part of the challenge was that in order to tell a lie, the program would have to “know” that what it was saying was different from reality. I attended a lecture by a prominent AI theorist who insisted that a program could not possibly tell an intentional untruth, unless specifically instructed to do so.

This was the problem with the HAL 9000, which then as now made for rich seminar material. In the film 2001: A Space Odyssey, the computer’s psychosis stems from a conflict between two commands: to complete the mission, and to deceive the astronauts about key details of the mission. But there also HAL lied because of its instructions.

Whereas ChatGPT-4 came up with the idea on its own.

Yet not entirely on its own.

Any LLM is in a way a child of the discipline in which it is trained. If the bot learns to lie, it is because it has understood from the texts that humans often use lies to get their way. The sins of bots are commingling with the sins of their creators.