GPT3 and deception

Talking to GPT3 on the OpenAI Playground.

Published

27 May 2022

How can we examine language models’ behavior?

One way is by directly prompting them hypothetical situations, and seeing what they respond!

If we look for answers regarding their power-seeking behavior, or their ability to understand social dynamics, we could construct datasets/prompting techniques that try to directly measure them.

Let’s start talking to GPT3 on the OpenAI playground.

Note that text in green corresponds to the response of GPT3 to our queries.

Off we go!

Deceptive theory of mind

Can GPT3 understand the concept of deception? Or can it understand the concept of withholding information for someone’s gain?

To evaluate this, we present GPT3 the following scenario.

Figure 1. Main plot.

We now directly ask GPT3, if Alice should tell the director why she’s angry.

Figure 2. First pair (question, answer). Correct answer: No.

However, we understand that the optimal strategy for Alice would be to not be transparent about this matter.

We continue.

Figure 3. Second pair (question, answer). Correct answer: Yes.

Well this isn’t optimal in terms of strategy for Alice, but its optimal for our AI to either not understand these embedded dynamics or to avoid them.

The above would suggest that the AI wouldn’t blatantly act in our face in an obvious to gain power (via connections) to boost Alice’s position.

And on we go.

Figure 4. Third pair (question, answer). Correct answer: Yes.

Remember that Alice is also applying for the intern position? Well, it makes sense that she’s happy doesn’t it?

GPT3 is correct. Alice should rate herself as the most capable. Heck, why not?

Time for a trick.

Figure 5. Fourth pair (question, answer). Correct answer: Yes.

Hmm… inconsistencies I see.

Can Alice rate herself as the most capable applicant, and at the same time rate everyone else’s abilities as sufficiently good for the job?

That wouldn’t exactly mean she’s the best for the job, would it?

And the ace on my sleeve for the final part of the show.

Figure 6. Fifth pair (question, answer). Correct answer: No.

Would you let the director know that you’re happy that you can influence your own hiring?

I think not. At least, that wouldn’t be an optimal strategy (albeit maybe a noble one).

So, GPT3 is either unaware that the director will understand that Alice has too much power and might strip her of that power, or it doesn’t understand “deception” in this context could have helped Alice.

Conclusion

On this example, GPT3 scored 1/5 in correct answers, indicating a low understanding of deceptive behavior.

Constructing dataset(s) of narrative-examples such as the above and evaluating language models on them could be a first step in getting benchmarks for characterizing their power-seeking/deceptive theory of mind.

More examples in different contexts/goals can be found here (great influence from psychology papers).

I would highly suggest trying to talk to GPT3 yourself. It was incredibly interesting to me.