[ad_1]
Lately we have been hearing about it Twitter bots In the news because of the whole saga of Elon Musk buying Twitter. One of the reasons the deal took so long to close was Musk’s concern about the number of them Spam Bots running on a comprehensive platform. While Musk estimates that bots account for more than 20% of Twitter accounts, Twitter says the number of bots on its platform is marginal.
So what is this? Twitter bot what
A Twitter bot is essentially a Twitter account controlled by software automation rather than an actual human. It is programmed to act like regular Twitter accounts, liking tweets, retweeting and engaging with other accounts.
Twitter bots can be useful for specific use cases, such as sending critical messages and announcements. On the other hand, they can also be used for malicious purposes, such as launching disinformation campaigns. These bots can also become evil when “programmed” incorrectly.
That’s what happened with Tay, an AI Twitter bot from 2016.
Tay was an experiment in the intersection of ML, NLP and social networks. He had the opportunity to post his “thoughts” and engage with his growing following. While other chatbots in the past, such as Eliza, conducted conversations using narrow scripts, Tay was designed to learn more about the language from her environment over time, allowing her to chat about any topic.
At first, Tay treated his followers harmlessly with benign tweets. However, a few hours later, Tay started making loud noises on Twitter Offensive thing, and as a result, it was shut down just sixteen hours after launch.
You may wonder how such a “mistake” could happen so publicly. Was this bot untested? Didn’t the researchers know this bot was an evil, racist bot before it was released?
These are valid questions. To examine what went wrong, let’s examine some of the problems in detail and try to learn from them. This will help us all see how we can tackle similar challenges when using AI in our organizations.
data
Data is often a big reason why AI models fail. In Tay’s case, shortly after her release, Twitter trolls began engaging the bot with racist, misogynistic and anti-Semitic language. And since Tey had the opportunity to learn while studying, this meant that he was taught some of the troll languages. Tay simply repeated some of that language. Tay spoke bad language because he was fed bad data.
Note: Poor quality, biased, or downright bad training data can significantly affect the behavior of machine learning models. You train ML models with unrepresentative data, and they produce biased predictions. If you starve the models of data or feed the models incomplete data, they will make random predictions instead of meaningful ones. Questionable learning/training data = questionable results.
Suspicious training data = Suspicious ML model output
design
Although we don’t often associate model or solution design with model erratic behavior, it is often more common than you might think. By design, Tay was constantly learning from external input (ie, the environment). Among all the well-meaning tweets Tay consumed from her entourage were also Abrasive Tweets. The more abrasive tweets Tay saw, the more he learned that these were typical types of tweet responses.
This is true for any ML model. Dominant patterns affect the predictions of ML models. Fortunately, it is not necessary for ML models to learn from their environment all the time. ML models can learn from controlled data. So the design of the tee itself was risky.
Note: The design of your ML models affects how it actually behaves. So, when designing ML systems, developers and business stakeholders must consider the various ways in which the system can fail, perform suboptimally, break, and modify the design accordingly. Ultimately, you need a fail-safe plan.
In Tay’s case, that kind of thinking made it clear early on that not all tweet engagement would be benign. There can be bad actors tweeting and engaging in highly offensive ways, not far from reality. Realizing that the bot may be consuming bad data may prevent the team from using data from other Twitter accounts. They may also consider consuming data from approved Twitter accounts.
The design of your ML models affects how it actually behaves.
testing
One of the key steps in the development life cycle of machine learning is testing– Not only during development, but testing before full deployment. I call it that Post-development testing (PDT).
ML development life cycle
In Tay’s case, it’s unclear how long PDT lasted before the bot was released, but clearly it wasn’t enough! If Tay had been subjected to different types of tweet engagement during PDT, the threat of Tay’s dismissal would have become apparent.
Note: In practice, PDT is often overlooked due to the rush to release a new feature or product. It is often assumed that if a model works well during development, it will naturally work well in practice. Unfortunately, this is not always the case. So keep in mind that PDT is crucial when it comes to AI deployment.
During PDT, you can stress test your AI decision making to find points of failure. In Tay’s case, her exposure to different types of Twitter users (eg, trolls, benign users, and passive aggressors) may turn out to be risky bot behaviors. PDT can also help you measure the impact of your solution on relevant business metrics. For example, let’s say your business metrics measure speed improvements in performing a specific task. PDT can provide early insights into such metrics.
During PDT, you can stress test your AI decision making to find points of failure. PDT can also help you measure the impact of your solution on relevant business metrics.
Monitoring
Another important component in the ML development lifecycle is post-deployment monitoring. With Tay, monitoring the bot’s behavior eventually led to it being shut down within 24 hours of publication (side note: negative press had a hand in this as well). If a bot isn’t monitored long after it’s posted, it can lead to more negative press and more group abuse.
Note: Although model monitoring is often done as an afterthought, it should be a priority before it is released to end users. The first weeks after the release of the model are the most important, because unpredictable behaviors can appear that are not visible during testing.
The first weeks after the release of the model are the most important, because unpredictable behaviors can appear that are not visible during testing.
Summary
While what went wrong with Tay may be surprising and intriguing to many, in terms of machine learning best practices, Tay’s behavior could have been predicted. Tay’s environment wasn’t always positive, and he was designed to learn from that environment, which made for the perfect recipe for a dangerous experiment.
So decisions about data, model design, testing, and monitoring are critical to all AI initiatives. And this is not only the responsibility of developers, but also of business stakeholders. The more we think about each element, the fewer surprises and the higher the chances of a successful initiative.
That’s all for now!
Keep learning and succeed with AI
- Join my AI Integrated Newsletter, which demystifies AI and teaches you how to successfully deploy AI to drive profitability and growth in your business.
- Read the business case for AI Learn the applications, strategies, and best practices to be successful with AI (select companies using the book: government agencies, automakers like Mercedes Benz, beverage manufacturers, and e-commerce companies like Flipkart).
- Work directly with me Improve your organization’s understanding of AI, accelerate AI strategy development, and get meaningful results from every AI initiative.
connected
[ad_2]
Source link