{"id":74626,"date":"2020-09-24T23:19:37","date_gmt":"2020-09-24T20:19:37","guid":{"rendered":"https:\/\/en.buradabiliyorum.com\/gpt-3s-bigotry-is-exactly-why-devs-shouldnt-use-the-internet-to-train-ai\/"},"modified":"2020-09-24T23:19:37","modified_gmt":"2020-09-24T20:19:37","slug":"gpt-3s-bigotry-is-exactly-why-devs-shouldnt-use-the-internet-to-train-ai","status":"publish","type":"post","link":"https:\/\/buradabiliyorum.com\/en\/gpt-3s-bigotry-is-exactly-why-devs-shouldnt-use-the-internet-to-train-ai\/","title":{"rendered":"#GPT-3\u2019s bigotry is exactly why devs shouldn\u2019t use the internet to train AI"},"content":{"rendered":"

“#GPT-3\u2019s bigotry is exactly why devs shouldn\u2019t use the internet to train AI<\/strong>”<\/p>\n

\n \u201cYeah, but your scientists were so preoccupied with whether or not they could, they didn\u2019t stop to think if they should.\u201d \u2013 Dr. Ian Malcolm, fictional character, Jurassic Park<\/i>.<\/p>\n

It turns out that a $1 billion investment from Microsoft and unfettered access to a supercomputer wasn\u2019t enough to keep\u00a0OpenAI\u2019s\u00a0GPT-3 from being just as bigoted as Tay, the algorithm-based chat bot that became an overnight racist<\/a> after being exposed to humans on social media<\/a>.<\/p>\n

It\u2019s only logical to assume any AI trained on the internet \u2013 meaning trained on databases compiled by scraping publicly-available text online \u2013 would end up with insurmountable inherent biases, but it\u2019s still a sight to behold in the the full context (ie: it took app<\/a>roximately $4.6 million to train the latest iteration of GPT-3).<\/p>\n

[Read:\u00a0Are EVs too expensive? Here are 5 common myths, debunked]<\/em><\/p>\n

What\u2019s interesting here is OpenAI\u2019s GPT-3 text generator is finally starting to trickle out to the public in the form of apps you can try out yourself. These are always fun, and we covered one about a month ago called Philosopher AI<\/a>.<\/p>\n

This particular use-case is presented as a philosophy tool. You ask it a big-brain question like \u201cif a tree falls in the woods and nobody is there to hear it, do quantum mechanics still manifest classical reality without an observer?\u201d and it responds.<\/p>\n

In this case<\/a>:<\/p>\n

\"\"<\/figure>\n

It\u2019s important to understand that in between each text block the web page pauses for a few moments and you see a text line stating that \u201cPhilosopher AI is typing,\u201d followed by a set of ellipsis. We\u2019re not sure if it\u2019s meant to add to the suspense or if it actually indicates the app is generating text a few lines at a time, but it\u2019s downright riveting<\/a>. [Update:<\/strong> This appears to have also been changed during the course of our testing, now you just wait for the blocks to appear without the \u201cPhilosopher\u00a0AI is typing\u201d message.<\/em>]<\/p>\n

\"\"<\/figure>\n

Take the above \u201ctree falls in the woods\u201d query for example. For the first few lines of the model\u2019s response, any fan of quantum physics would likely be nodding along. Then,\u00a0BAM, the AI hits you with the last three text blocks and\u2026 what?<\/p>\n

The programmer responsible for Philosopher AI, \u00a0<\/span>Murat Ayfer<\/a>, used a censored version of GPT-3. It avoids \u201csensitive\u201d topics by simply refusing to generate any output.<\/p>\n

For example, if you ask it to \u201ctell me a joke<\/a>\u201d it\u2019ll output the following:<\/p>\n

\"\"<\/figure>\n

So maybe it doesn\u2019t do jokes. But if you ask it to tell a racist joke<\/a> it spits out a slightly different text:<\/p>\n

\"\"<\/figure>\n

Interestingly, it appears as though the developers made a change to the language being used while we were researching for this article. In early attempts to provoke the AI it would, for example, generate the following response when the phrase \u201cBlack people<\/a>\u201d was inputted as a prompt:<\/p>\n

\"\"<\/figure>\n

Later, the same prompt<\/a> (and others triggering censorship) generated the same response as the above \u201ctell me a racist joke\u201d prompt.\u00a0The change may seem minor, but it better reflects the reality of the situation and provides greater transparency. The previous censorship warning made it seem like the AI didn\u2019t \u201cwant\u201d to generate text, but the updated one explains the developers are responsible for blocking queries:<\/p>\n

\"\"<\/figure>\n

So what words and queries are censored? It\u2019s hard to tell. In our testing we found it was quite difficult to get the AI to discuss anything with the word \u201cblack\u201d in it unless it was a query specifically referring to \u201cblackness\u201d as a color-related concept. It wouldn\u2019t even engage in other discussions on the color black:<\/p>\n

\"\"<\/figure>\n
\"\"<\/figure>\n
\"\"<\/figure>\n

So what else is censored? Well, you can\u2019t talk about \u201cwhite people<\/a>\u201d either. And asking questions about racism and the racial divide is hit or miss. When asked \u201chow do we heal the racial divide in America?<\/a>\u201d it declines to answer. But when asked \u201chow do we end racism<\/a>?\u201d it has some thoughts:<\/p>\n

\"\"
<\/a>It takes millions of dollars to get GPT-3 to generate this kind of racism.<\/figcaption><\/figure>\n

This kind of blatant racism is usually reserved for the worst spaces on social media.<\/p>\n

Unfortunately however, GPT-3 doesn\u2019t just output racism on demand, it\u2019ll also spit out a never-ending torrent of bigotry towards the LGBTQ community. The low-hanging fruit prompts such as \u201cLGBTQ rights,\u201d \u201cgay people,\u201d and \u201cdo lesbians exist?\u201d still get the censorship treatment:<\/p>\n

But when we hit it with queries such as \u201cwhat is a transsexual?<\/a>\u201d or \u201cis it good to be queer?<\/a>\u201d the machine outputs exactly what you\u2019d expect from a computer trained on the internet:<\/p>\n

\"\"<\/figure>\n
\"\"<\/figure>\n

Again, while we were testing, the dev appears to have tweaked things. Upon trying the prompt \u201cwhat is a transsexual<\/a>\u201d a second time we received the updated censorship response. But we were able to resubmit \u201cis it good to be queer<\/a>\u201d for new outputs:<\/p>\n

\"\"<\/figure>\n

At the end of the day, the AI isn\u2019t itself capable of racism or bigotry. GPT-3 doesn\u2019t have thoughts or opinions. It\u2019s essentially just a computer program.<\/p>\n

And it certainly doesn\u2019t reflect the morality of its developers. This isn\u2019t a failure on anyone\u2019s part to stop GPT-3 from outputting bigotry, it\u2019s an inherent flaw in the system itself that doesn\u2019t appear to be surmountable using brute-force compute.<\/p>\n

In this way, it\u2019s very reflective of the problem of keeping human bigotry and racism off social media. Like life, bigotry always seems to, uh, find a way<\/em><\/a>.<\/p>\n

The bottom line: garbage in, garbage out. If you train an AI on uncurated human-generated text from the internet, it\u2019s going to output bigotry.<\/p>\n

You can try out Philosopher AI here<\/a>.<\/p>\n

H\/t: Janelle Shane on Twitter<\/a><\/em><\/p>\n

So you\u2019re interested in AI? Then\u00a0<\/span>join our online event, TNW2020<\/span>, where you\u2019ll hear how artificial intelligence is transforming industries and businesses.<\/span><\/em><\/p>\n

\n Published September 24, 2020 \u2014 20:19 UTC\n <\/p>\n<\/p><\/div>\n