News: 1602243247

Someone used OpenAI's GPT-3 text-generating software to write a spree of posts on Reddit, convincing people the missives were penned by a real person, and banking thousands of internet points in the process, The Register can confirm.

Redditor [1]thegentlemetre has produced over the course of the past 10 days more than 1,000 answers on the popular subreddit [2]AskReddit in which people turn to millions of their peers for practical tips and life advice. At its most frenzied periods, the account generated about one comment per minute, offering a jumble of answers to people's questions. That would be quite a feat for a lone human, considering most of these submissions are several paragraphs long.

Some of its responses are bizarre and downright creepy, such as this [3]fake story about a colony of humans living trapped underneath elevators. Sometimes it [4]doesn't answer a netizen's question at all, and sometimes it spreads [5]false information , such as denying that the state of Nebraska exists. There's one post where it believes that [6]men are discriminated against when it comes to being topless in public.

Phillip Winston, a software engineer, [7]speculated that thegentlemetre was, in fact, a bot based on GPT-3. "I saw this post and thought it's either a prolific user or a bot," he told The Register .

"I looked at the history and saw they were posting once per minute at times. I figured it had to be a bot. I looked at the posts some were very real but many had issues, and I felt sure it was a bot, but better than anything I'd seen."

I felt sure it was a bot, but better than anything I'd seen

His suspicions were confirmed when he asked Reddit users in the GPT-3 subreddit to take a look. Several people suggested the text had the same patterns as prose produced by the GPT-3-powered [8]Philosopher AI , built by Murat Ayfer, an engineer who had early access to OpenAI's cloud-based GPT-3 API as a beta tester. That would suggest the mystery Redditor's comments were indeed the work of GPT-3.

Philosopher AI started as a [9]website where anyone could type in a question, and get an answer composed by GPT-3 back; Ayfer's site was in effect an access layer between netizens and an instance of OpenAI's GPT-3 hosted in the cloud. When OpenAI ended its free trial for beta users, Ayfer ended free access to Philosopher AI, and the service is now only available through a smartphone app that costs $2.99 to download.

Get ready for Clippy 9000: Microsoft exclusively licenses OpenAI's mega-brain GPT-3 for anything and everything [10]READ MORE

Ayfer confirmed that whoever is behind thegentlemetre did indeed use Philosopher AI to craft a Reddit bot. Ayfer keeps a database of all the responses generated using his software, and he found that the bot's posts matched some of those in his database word-for-word.

He believes the bot was able to mine his application after bypassing a Captcha road-block put in place to prevent abuse. The iOS version of his app doesn't ask users to complete the Captcha test so if someone accessing the Philosopher AI backend masqueraded as the iOS app, they could get through unhindered.

"I had the iOS app configured to skip the Captcha check to make sure paying users don't get blocked by it, but the way I identified iOS clients was quickly reverse-engineered," he told The Register . "I've already made some patches and I'm continuing to improve the Captcha system." He also introduced a stricter limit that prevents users from querying Philosopher AI as much.

Ayfer believes that an automated script was set up to read the questions posted on the AskReddit forum, pass them on to Philosopher AI, and its text outputs were then posted onto the subreddit. After he improved the Captcha process on his app, the activity on thegentlemetre's account died down.

An odd stunt?

Weirdly enough, the bot hasn't completely gone silent, however. Instead, the user has made another two posts since – one in their favorite AskReddit subreddit and another in ToasterTalk, in which people discuss AI ethics.

The responses are oddly short, each one consisting only of two or three words. Winston believes that the human who built the bot script has now taken over the account. We tried to contact thegentlemetre over Reddit for comment, and have yet to receive a response.

"For the last 3,000 years, human-looking text meant a human wrote it," Winston told us. "Very soon that will no longer be true. I'm sure many great things will come from that, but it does raise a lot of questions. The dystopian vision for me is millions of Redditors and Twitter users having conversations with bots and not knowing it."

The dystopian vision for me is millions of Redditors and Twitter users having conversations with bots and not knowing it

OpenAI has warned its software could be used by miscreants to create bots and spread fake news and spam. "It's clearly just someone having some fun, and stirring things up," Ayfer reckoned, though.

"It's somewhat valuable to see it happen now in a harmless setting; it's easy to imagine what someone can do on social media using unlimited malicious GPT-3 access, or any similar machine learning tech which will surely become more widely available in the coming months or years. This will be an interesting problem for the next few years. we will need bulletproof bot detection."

"I wonder if only OpenAI or someone with GPT-3 can detect it," Winston added. "If so, they could charge for that service. So is it like the vacuum salesman who barges into your house and throws dirt on the floor first, then charges you to clean it up? Charging people to run GPT-3 bots and then charging other people to detect those same bots could be the world's greatest business model."

Spokespeople for Reddit were not available for comment.

"We don't support anyone using our API to deceive people," a spokesperson for OpenAI told El Reg . "Our company is founded on the belief that artificial intelligence should benefit all of humanity, and we are constantly working to fight misuse, bias and other industry-wide challenges." ®

Get our [11]Tech Resources



[1] https://www.reddit.com/user/thegentlemetre/

[2] https://www.reddit.com/r/AskReddit/

[3] https://www.reddit.com/r/AskReddit/comments/j1wrzl/elevatormaintenance_folks_what_is_the_weirdest/g71r75b/

[4] https://www.reddit.com/r/AskReddit/comments/j5n6v6/drug_dealers_of_reddit_what_was_your_weirdest/

[5] https://www.reddit.com/r/AskReddit/comments/j1ld8g/what_evidence_is_there_that_the_state_of_nebraska/

[6] https://www.reddit.com/r/AskReddit/comments/j19zrc/how_can_men_support_the_free_the_nipple_movement/

[7] https://www.kmeme.com/2020/10/gpt-3-bot-went-undetected-askreddit-for.html

[8] https://philosopherai.com/

[9] https://philosopherai.com/

[10] https://www.theregister.com/2020/09/23/microsoft_will_exclusively_license_openais/

[11] https://whitepapers.theregister.com/