The biggest tell that you are chatting with an AI and not a human is that they talk too much.

Yesterday, to start off a session with ChatGPT, I asked a yes/no question and it responded with a “Yes” and another 325 more words. No person would do this.

Another version of this same behavior is when you ask an LLM a vague question—it answers it! No questions. Just an answer. Again, people don’t usually behave this way.

It’s odd because in the online forums that were used to train these models, vague questions are often called out. It’s nice that the LLM isn’t a jerk, but asking a clarifying question is basic “intelligent” behavior and they haven’t mastered it yet.

If, twenty years ago, you took the most sophisticated AI thinkers in the world and set up a Turing test with a modern AI chatbot, they would all think they were chatting with a human. Today, that same chatbot would have a hard time fooling most people that have played around with them.

In a sense, humans have been upgraded by the existence of LLMs. If we had no idea that they existed, we would be fooled by them. But, now that we’ve seen them, we’ve collectively learned how they come up short.

I just had conversation with ChatGPT that ended with it making product recommendations. I was asking about documentation, so it suggested Confluence and Notion.

Those are good options and do match what I was asking about. But, if ChatGPT is going to make product recommendations, then companies will try to game this like they do for search engines.

Before Google, search engines ranked pages on the count of keywords in the document. That was easy to game by just stuffing the keywords. Google solved this by ranking the search results by backlinks. They reasoned that pages with the most incoming links must be better because a backlink was like a vote of confidence. Unfortunately, that was also easy to game. The last 20+ years has been a struggle between Google trying to rank results correctly and SEO experts trying to rank higher than they deserve. There is an enormous amount of money fueling both sides.

We are just starting with chatbots. I don’t know why it specifically chose Confluence and Notion, but since it’s a probability engine, more text in the training set would be a factor. It understands sentiment, so it can discern good reviews from bad reviews and knows that recommendations should offer good options.

But does it understand authority, bias, expertise, etc. If I stuff the internet with 1000’s of good reviews of my product (even if they would never be read by a human), would ChatGPT then start recommending me? The ChatBot equivalents of SEO experts will be trying to figure this out.

Since ChatGPT is currently stuck in January 2022 (about a year ago), you might not be able influence its suggestions for a year or so. The learning cycle is currently very long. But, that also means that it probably can’t recover from being gamed quickly either. Hopefully, models can be developed that resist gaming.

In 2007, seeing the iPhone and owning a taxi company, you might think it was a good idea to make an app. After all, people need taxis when they are out and about. Maybe, you think, that people will stop using pay phones, so they won’t see the stickers you plastered there with your phone number.

So, you get Phone-a-Taxi made so that people can use an app to call you up. They tell you where they are and you send a cab. There are hardly any apps in the store, so you get a bunch of downloads and calls, and things look great. You are riding the smart phone hype wave.

And then Uber comes along.

That’s what a lot of LLM features feel like to me. Technically, they are using LLM technology, but mostly just trying to shoehorn chats and autocomplete into apps that were not designed around LLMs. Even GitHub Copilot is guilty of this. I love it, but in the end I still have the same amount of code I had before. I don’t even want code—I just want the program.

Like the Phone-a-Taxi app, they are still doing things the old way.

LLM chatbots are bad at some things. Some of this is intentional. For example, we don’t want chatbots to generate hate speech. But, some things are definitely not intentional, like when they make stuff up. Chatbots also fail at writing non-generic text. It’s amazing that they can write coherent text at all, but they can’t compete with good writers.

To get around some of these limitations, we have invented a field called “prompt engineering”, which use convoluted requests to get the chatbot to do something it doesn’t do well (by design or not). For example, LLM hackers have created DAN prompts that jailbreak the AI out of its own safety net. We have also seen the leaked prompts that the AI companies use to set up the safety net in the first place. Outside of safety features, prompt engineers have also found clever ways of trying to get the LLM to question its own fact assertions to make it less likely that it will hallucinate.

Based on the success of these prompts, it looks like a new field is emerging. We’re starting to see job openings for prompt engineers. YouTube keeps recommending that I watch prompt hacking videos. Despite that, I don’t think that this will actually be a thing.

All of the incentives are there for chatbot makers to just make chatbots better with simple prompts. If we think chatbots are going to approach human-level intelligence, then we’ll need prompt engineers as much as we need them now for humans, which is “not at all.”

Prompt engineering is not only a dead end, it’s a security hole.

Yesterday, I wrote that LLM prompt hacking was like an injection attack. I looked up injection in OWASP’s 2021 10 top of security vulnerabilties and see that it’s number three. Since LLM prominence started this year, they haven’t listed prompt hacking yet, but you can see from their description and list of remedies how similar it is to injection. And since we’re busily attaching LLMs to web applications via their APIs, prompt hacking should be considered a web application security vulnerability in the next survey.

Here’s the top prevention technique:

Preventing injection requires keeping data separate from commands and queries:

• The preferred option is to use a safe API, which avoids using the interpreter entirely, provides a parameterized interface, or migrates to Object Relational Mapping Tools (ORMs).
  Note: Even when parameterized, stored procedures can still introduce SQL injection if PL/SQL or T-SQL concatenates queries and data or executes hostile data with EXECUTE IMMEDIATE or exec().

For an LLM this means that the LLM itself isn’t affected by the user query. I realize that that may be impossible with current implementations. My suggestion is to somehow create two channels (one for “code” and one for “data”) in the training process so that the resulting model isn’t exploitable this way.

No, I have no idea how to do that, but it’s not with a more convoluted prompt.

Web programmers can cause security problems if they embed data into HTML and render the result. For example, if I have a simple form that asks for your name and then output a page with that name in it, I’ll open myself up to an “injection” attack if the user types in some Javascript, and I don’t carefully escape it. I’ll end up running that Javascript.

The same is true if we take user data and try to create queries by concatenating it with SQL, as lampooned by XKCD.

We invented encoding and string interpolation techniques to solve this. But nothing forces you to use those features, so we still mess it up, which is why security bounties are frequently paid for injection attacks.

But, those issues are with legacy languages like HTML and SQL where we send strings that mix code and data over the network and run them. We should have designed them in a way that separated the code and the data. Surely, we learned from that for new things that we invented since then.

We did not.

An LLM chatbot is also a service that we send strings over a network to. The prompt you send is “code” in natural language and the LLM “runs it”. The problem is that there is a kind of meta-language that controls the chatbot itself, which can be sent before your normal prompts. Using these “jailbreaking” prompts, you can trick the LLM into dropping its safety net and produce hate speech or help you code malware.

These prompts are essentially the same idea that Bobby’s mom is using in the comic, and the solution is likely going to be a prompt version of what encoding and string interpolation is doing.

It would be better if the system was designed such that user chat requests weren’t treated like a program that could change the chatbot itself.