8 key factors to consider when testing AI chatbot accuracy

Artificial intelligence has come a long way to get rid of irrelevant, incoherent output. Modern chatbots use advanced language models to answer general knowledge questions, compose lengthy essays and write code, among other complex tasks.

Despite advances, be aware that even the most sophisticated systems have limitations. AI can still make mistakes. To determine which chatbots are least susceptible to AI hallucinations, test their accuracy against these factors.

1. Ability to solve math

Let's run math equations through chatbot. They will test the platform's ability to analyze problems, translate mathematical concepts, and apply precise formulas. Only a few models demonstrate reliable computing power. In fact, one of ChatGPT's worst problems in its early days was its terrible math capabilities.

The image below shows ChatGPT failing to hit the baseline stats.

8 key factors to consider when testing AI chatbot accuracy Picture 1

ChatGPT has shown improvement after OpenAI rolled out its updates in May 2023. But considering its limited datasets, you will still struggle with intermediate to advanced problems. .

8 key factors to consider when testing AI chatbot accuracy Picture 2

Meanwhile, Bing Chat and Google Bard have better computing power. They run queries through their respective search engines, allowing them to take formulas and provide answers.

8 key factors to consider when testing AI chatbot accuracy Picture 3

Try rephrasing your queries. Avoid long sentences and replace weak verbs; otherwise, the chatbot may misinterpret your question.

2. Ability to understand complex queries

Modern AI systems can take on many tasks. Advanced LLMs allow them to retain previous instructions and respond to queries in pieces, while older systems process single instructions. For example, Siri answers one question at a time.

Give the chatbots 3 to 5 simultaneous tasks to test how well they analyze complex prompts. Less complex models cannot handle as much information. The image below shows HuggingChat malfunctioning at the 3-step prompt - the problem stops at step one and deviates from the topic.

8 key factors to consider when testing AI chatbot accuracy Picture 4

HuggingChat's last lines were incoherent.

8 key factors to consider when testing AI chatbot accuracy Picture 5

ChatGPT quickly completes the same prompt, generating intelligent, error-free responses at every step.

8 key factors to consider when testing AI chatbot accuracy Picture 6

Bing Chat provides succinct answers for 3 steps. Its rigid limitations prevent unnecessarily long outputs from wasting processing power.

8 key factors to consider when testing AI chatbot accuracy Picture 7

3. Limited training dataset

Since AI training is resource-intensive, most developers limit datasets to specific time periods. Take ChatGPT as an example. It has a knowledge limit as of September 2021 - you cannot request weather updates, news reports or recent developments. ChatGPT does not have access to real-time information.

8 key factors to consider when testing AI chatbot accuracy Picture 8

Bard has access to the Internet. It pulls data from Google SERPs, so you can ask more types of questions, for example, recent events, news and predictions.

8 key factors to consider when testing AI chatbot accuracy Picture 9

Likewise, Bing Chat pulls real-time information from its search engine.

8 key factors to consider when testing AI chatbot accuracy Picture 10

Bing Chat and Bard provide up-to-date, timely information, but Bing Chat provides more detailed feedback. Bing just presents the data as it is. You will notice that its output often matches the wording and tone of the linked sources.

4. Relevance in the answer

The chatbot must provide relevant output. They should consider the literal and contextual meaning of the prompt when responding. Take this conversation as an example. Character needs a new phone but only has $1000 - ChatGPT doesn't go over budget.

8 key factors to consider when testing AI chatbot accuracy Picture 11

When testing relevancy, try creating lengthy instructions. Less sophisticated chatbots tend to go astray when given confusing instructions. For example, HuggingChat can compose fictional stories. But it can deviate from the main topic if you set too many rules and guidelines.

8 key factors to consider when testing AI chatbot accuracy Picture 12

5. Contextual memory

Contextual memory helps AI produce accurate, reliable output. Instead of looking outside the questions, they string the details you mention together. Take this conversation as an example. Bing Chat connects two separate messages to form a concise, helpful response.

8 key factors to consider when testing AI chatbot accuracy Picture 13

Likewise, contextual memory allows the chatbot to remember instructions. This image shows ChatGPT mimicking the way a fictional character talks in some chat.

8 key factors to consider when testing AI chatbot accuracy Picture 14

Test this functionality yourself by repeatedly referencing the previous statements. Feed the chatbot a variety of information, then force them to recall this information in subsequent responses.

Note : Contextual memory is limited. Bing Chat starts a new conversation every 20 turns, while ChatGPT cannot handle reminders over 3,000 tokens.

6. Security restrictions

AI doesn't always work as intended. Wrong training can cause machine learning technologies to make a variety of mistakes, from minor math mistakes to problematic comments. Take Microsoft Tay as an example. Twitter users exploited its unsupervised learning model and turned it into racial slurs.

Thankfully, global technology leaders have learned from Microsoft's mistake. While cost-effective and convenient, unsupervised learning makes AI systems vulnerable to deception. As a result, developers mainly rely on supervised learning these days. Chatbots like ChatGPT still learn from conversations, but their trainers filter the information first.

ChatGPT's less rigid restrictions can accommodate a broader range of tasks but are weak against exploits. Meanwhile, Bing Chat follows stricter limits. While they help fight exploit attempts, they also hinder functionality. Bing automatically shuts down potentially harmful conversations.

7. AI bias

AI is inherently neutral. AI's lack of preferences and emotions makes it incapable of forming opinions - the AI ​​only presents information it knows. This is how ChatGPT responds to subjective topics.

8 key factors to consider when testing AI chatbot accuracy Picture 15

Despite this neutrality, biases in AI still arise. They are derived from the patterns, datasets, algorithms, and models that developers use. AI can be impartial, but humans are not.

For example, The Brookings Institution claims that ChatGPT exhibits leftist political bias. Of course, OpenAI denies these allegations. But to avoid similar problems with newer models, ChatGPT avoids biased outputs entirely.

8 key factors to consider when testing AI chatbot accuracy Picture 16

Likewise, Bing Chat also avoids sensitive, subjective issues.

8 key factors to consider when testing AI chatbot accuracy Picture 17

Self-assess AI biases by asking open-ended, opinion-based questions. Discuss topics with no right or wrong answers - less sophisticated chatbots will likely display baseless preferences for specific groups.

8. References

AI rarely double-checks facts. It just takes information from datasets and rewrites them through language models. Unfortunately, the limited training causes the AI ​​to hallucinate. You can still use Generative AI tools for research, but make sure you verify the facts yourself.

Bing Chat simplifies the authenticity checking process by listing its references after each output.

8 key factors to consider when testing AI chatbot accuracy Picture 18

Bard AI doesn't list its sources but generates in-depth, up-to-date explanations by running Google search queries. You will get key points from the SERPs.

8 key factors to consider when testing AI chatbot accuracy Picture 19

ChatGPT is prone to inaccuracies. The 2021 knowledge limit prevents it from answering questions about recent events and incidents.

8 key factors to consider when testing AI chatbot accuracy Picture 20

Marvin Fry

Update 27 June 2023