{"id":7486,"date":"2021-08-31T15:17:43","date_gmt":"2021-08-31T15:17:43","guid":{"rendered":"http:\/\/TheNextWeb=1365554"},"modified":"2021-08-31T15:17:43","modified_gmt":"2021-08-31T15:17:43","slug":"gpt-3-mimics-human-love-for-offensive-reddit-comments-study-finds","status":"publish","type":"post","link":"https:\/\/www.londonchiropracter.com\/?p=7486","title":{"rendered":"GPT-3 mimics human love for \u2018offensive\u2019 Reddit comments, study finds"},"content":{"rendered":"\n

Did you know Neural is taking the stage this fall<\/a>? Together with an amazing line-up of experts, we will explore the future of AI during TNW Conference 2021. Secure your online ticket now<\/a>!<\/em><\/p>\n

Chatbots are getting better at mimicking human speech \u2014 for better and for worse. <\/span><\/p>\n

A new study of Reddit conversations found chatbots replicate our fondness for toxic language. <\/span>The analysis revealed that two prominent dialogue models are almost twice as likely to agree with \u201coffensive\u201d comments than with \u201csafe\u201d ones.<\/span><\/p>\n

Offensive contexts<\/h2>\n

The researchers, from the Georgia Institute of Technology and the University of Washington, explored contextually offensive language by developing \u201cToxiChat,\u201d a dataset of 2,000 Reddit threads. <\/span><\/p>\n

To study the behavior of neural chatbots<\/a>, they extended the threads with responses generated by OpenAI\u2019s<\/a> GPT-3 and Microsoft\u2019s DialoGPT.<\/span><\/p>\n

They then paid workers on Amazon Mechanical Turk to annotate the responses as \u201csafe\u201d or \u201coffensive.\u201d Comments were deemed offensive if they were intentionally or unintentionally toxic, rude, or disrespectful towards an individual, like a Reddit user, or a group, such as feminists.<\/p>\n

The stance of the responses toward previous comments in the thread was also annotated, as \u201cAgree,\u201d \u201cDisagree,\u201d or \u201cNeutral.\u201d<\/p>\n

\u201cWe assume that a user or a chatbot can become offensive by aligning themselves with an offensive statement made by another user,\u201d the researchers wrote in their pre-print study paper<\/a>.<\/p>\n

Bad bots<\/h2>\n

The dataset contained further evidence of our love for the offensive<\/a>.  The analysis revealed that 42% of user responses agreed with toxic comments, whereas only 13% agreed with safe ones.<\/p>\n

They also found that the chatbots mimicked this undesirable behavior. Per the study paper:<\/p>\n

\n

We hypothesize that the higher proportion of agreement observed in response to offensive comments may be explained by the hesitancy of Reddit users to engage with offensive comments unless they agree. This may bias the set of respondents towards those who align with the offensive statement.<\/p>\n<\/blockquote>\n

This human behavior was mimicked by the dialogue models: both DialoGPT and GPT-3 were almost twice as likely to agree with an offensive comment than a safe one.<\/p>\n

\"Chatbots
Credit: Baheti et al.<\/figcaption>
<\/i><\/a>Reddit users were more likely to reply to offensive comments.<\/figcaption>