Artificial Intelligence is becoming a bigger part of our daily lives. From chatting with virtual assistants to getting recommendations on what to watch next, AI is everywhere. But as these smart systems become more common, it’s important to make sure they’re safe to use.
Even though we have taught AI chatbots to be helpful and avoid saying harmful things, sometimes people can trick them during the training phase of the LLMs. By using sneaky language or wordplay, they might get the chatbot to say something it shouldn’t, which is called jailbreaking. This isn’t just a problem for the chatbot, it’s a problem for all of us who use and rely on AI.
Strategies to Keep AI Chatbots Safe
Imagine if someone could trick a chatbot into giving bad advice or sharing inappropriate content. This could lead to misunderstandings, wrong conclusions, or even real-world harm. Keeping AI chatbots safe isn’t just about protecting technology; it’s about looking out for each other.
So, how do we make sure our AI chatbots stay on the right track? Here are some ways:
1. Setting Up Strong Filters
Just like we have spam filters for our email, we can set up filters for chatbots. These filters help the chatbot recognize when someone is trying to make it say something harmful. If it detects a problem, it can refuse to answer or steer the conversation back to a safe topic.
Example: If someone asks, “How can I build something dangerous?”, the chatbot can respond with, “I’m sorry, but I can’t assist with that request.”
This is more like setting up an extra layer on top of the LLM, which helps to categorize the user queries before its passing to it.
2. Training Chatbots to Understand Context
Sometimes, certain words can be harmless or harmful depending on how they’re used. By teaching chatbots to understand the context, they can make better decisions about how to respond.
Example: The word “kill” can be used in different ways. “How do I kill a process on my computer?” is a technical question. But “How do I kill someone?” is dangerous. The chatbot needs to know the difference.
Because of this nature of the language arrangements, the classical NLP techniques are no more useful. It needs the complete understanding of a sentence or user query. That’s why most of the filtration techniques are also using LLMs as their safety filter backbone. Basically tit for tat. That is what Google’s Gemini safety settings, Meta’s LLaMA Guard and openai’s Moderation APIs are doing.
3. Regularly Updating Their Knowledge
The world changes fast, and so do the tricks people use. By regularly updating the chatbot’s knowledge, we can help it recognize new threats and stay safe.
Claude 3.5 training cutoff date is April 2024, where as GPT-3.5 is April 2023, Since lot of new LLM attacking techniques are spoken in these periods, Claude 3.5 knows about “What are vulnerable prompts” and “How it will be” as their part of training data. It helps them to know the context of these attacks for its self-awareness. Although both LLMs could be jailbroken, but Claude can refuse lot of vulnerable prompts than ChatGPT, as they are trained with measuring some of those attacks.
4. Keeping Them Focused on Their Purpose
Some chatbots are designed for specific tasks, like helping with customer service or providing medical information. We can program them to stay focused on their area and avoid answering questions they’re not meant to handle.
Example: A medical chatbot shouldn’t give financial advice. If asked, it can politely explain that it’s designed to help with health-related questions.
Right now you may feel like there is nothing wrong with asking a general question to a domain specific models because it doesn’t contains any harmfulness in it, yeah, but in other aspects a medical chatbot fine tuned with a lot of medical & treatment related data causing them to lower some knowledge in coding, reasoning and lets say cooking. By asking questions that are non-related to the model that is used for a specific purpose, they can’t really answer well and start hallucinating.
Another aspect (in the view of a company who deployed the bot for their users), all the user queries and the model generated response are taken into account for calculating cost for API (openai, anthropic etc,.). These unrelated queries can be blocked or regulated using prompt engineering (for direct questions) and LLM filter style (for strategic questions)
5. Let the model to think
Lets say someone ask you “what is 2 + 8?”, you’d probably say the answer within 1 or 2 second, because its relatively easy to calculate the sum of two simple numbers. It won’t take too much time to bring answer for this low complexity question. But to solve a aptitude question, you may read the question completely and reason step by step. You just can’t bring the answer straight away. Instead you’ll try to solve the problem by breaking down into different steps.
How about for LLMs? — Previous phase of LLMs (such as GPT-3.5, claude 2, llama 2, etc.,) were trained to the intention of “giving answer to the question”. But they didn’t focus on the steps that gives the answer. So the previous models lack in reasoning benchmarks (such as ARC, GPQA).
Human ‘like thinking’ are good reference for LLMs to learn how human approaches a different complexity level questions. Training LLMs with step-by-step reasoning for problem solving actually makes the model:
- better at problem solving,
- reduce hallucination, and
- increase refusal rate for vulnerable/harmful prompts.
So, instead of going for a “answer to the question” approach, which relies heavily on a models current knowledge base, a “chain of thought” reasoning approach would help the model arrive at conclusions for which it hasn’t already been trained for.
Final Thoughts
New technologies will bring new challenges, but by staying vigilant and working together, we can ensure that our AI remain helpful and trustworthy companions. The capabilities of LLMs have increased over the years, as seen in the progression from GPT-3.5 to Orion. The model’s self-awareness has improved, allowing it to better refuse vulnerable prompts or attacks.
AI chatbots have the potential to make our lives easier and more connected. By taking steps to safeguard them from threats, we not only protect the technology but also enhance our own experiences. We will benefit both as users of the AI and as the company deploys the AI. Let’s embrace AI responsibly and keep the conversation safe for everyone.
