What Is Botshit? How Do You Prevent Botshit Scenarios?

I read an article in Business Insider titled, “Botshit is an example of how AI is making customer service worse.”

The examples cited in the article were an eye-opener.

Here is a concise summary of those.

Jake Moffatt booked a flight to her funeral on Air Canada, hoping to use bereavement travel discounts.

Air Canada’s customer service chatbot told Moffatt he could claim the discount after the flight, and the company later denied his discount, stating that it had to be filed prior to the flight.

Canada’s Civil Resolution Tribunal ruled that Air Canada’s chatbot misled Moffatt and directed the airline to compensate him for the discount.

This is what is called botshit, which was coined by researchers Ian P. McCarthy, Timothy R. Hannigan, and Andre Spicer.

Botshit describes incorrect or fabricated information produced by Chatbots that humans use to complete tasks.

Here are a few more examples:

• Researchers in 2003 found that 75% of ChatGPT responses to drug-related questions were often inaccurate or incomplete. ChatGPT further generated fake citations to support some of its inaccurate responses.

• Recently, a UK parcel company removed its AI customer service chatbot after it swore at a customer.

• Last year, a New York law firm was fined $5000 after lawyers submitted a court brief containing false references produced by ChatGPT.

Does This Mean That Organizations Should Abandon Using AI for Their Customer Service?

AI is definitely here to stay. In the last six months, the usage of AI in customer service has doubled across the spectrum. Besides, most organizations are looking at implementing AI as a part of their customer service.

I remember playing Bridge many years ago. After the bidding, when it is time for you to lead, and you don’t find a logically perfect lead, what do you do?

I was told, “When in doubt, lead a Trump.”

How is this applicable to AI usage in customer service?

When in doubt, even an iota of it, don’t AI.

AI definitely has its use cases and can help you run a majority of customer service scenarios.

It can be improved using rigorous guidelines, guardrails, and restrictions like flagging a response for human intervention. 

Strategies to Minimize Risks With Chatbots

Data Is Key

The more quality training data you have, the better would be your AI output. You should regularly update the training data set to remove any inaccuracies and to add new information.

Quality Is Everything

Comprehensively test the chatbot for various scenarios and use cases to identify potential issues in real-world situations. You should use a combination of human review and automated testing to ensure that chatbot responses are appropriate.

Besides, you can train the chatbots to ask follow-up questions when the queries are ambiguous, so you are absolutely sure about what your customers want.

When all else fails, remember the adage:

When in doubt, even an iota of it, don’t AI.

Response Validation

When it comes to transactional queries, it is pretty easy to validate against the data available. For instance, if someone is asking to know their last five transactions, it is easy to understand the question and pull out this information from the existing systems.

What if it is not a transactional query but something similar to a bereavement travel discount?

Then, it becomes difficult to validate easily.

In this case, AI should flag this to a human review and take the time to respond to the customer or bring in a human agent to respond.

Implement Natural Language Integrations

Implement natural language understanding and natural language processing to improve the chatbot’s comprehension and response accuracy. 

Explanations Help

You should design your chatbot to provide the source of information. In the same bereavement travel discount example, the response could have been:

From what I see, you are eligible for a bereavement travel discount. However, I am connecting you to a supervisor who will be able to guide you through the process of claiming it.

Constant Training

Every AI implementation improves over time as it learns with additional data and experience.

It has to be trained by updating its knowledge base to ensure that it remains accurate and relevant, and it should keep getting better.

You should regularly conduct audits to understand if the responses remain accurate and relevant to the queries.

Seek Feedback

As you do with regular customer service, you should seek feedback from your customers on their experience with the chatbot.

Ask for qualitative responses in addition to quantitative responses, as they are a goldmine of information, which will allow you to train the chatbot for better results. 

Considering the use cases of AI, I believe that customer service is the easiest and the best place to start your AI initiatives.

It is no no-brainer to use chatbots for transactional queries. For the rest, you can mitigate risks associated with chatbots by constantly training and improving the quality of data.

Till such time, “When in doubt, even an iota of it, don’t AI.”

Flag them for human intervention.