GPT-3, OpenAI’s Revolution
Looking into the implications of the most powerful language model ever created
OpenAI is one of the clear leaders in AI right now, and it certainly leads the way in natural language systems. GPT, its text generation algorithm, has shown an uncanny ability to generate human-like text, revolutionising the domain of text generation. Its latest version, GPT-3, is now able to go beyond that, being able to not only come up with coherent paragraphs but also generating code from specifications and much more.
GPT is basically an autocomplete. However, it doesn’t just suggest the next word like in popular messaging apps. GPT is able to write coherent, lengthy paragraphs of text, following an initial prompt.
The content generated is so close to that written by humans, that there have been discussions on whether this tool is too powerful to be handed over to people.
Have a look for yourself:
The text was generated after having seeded only the title, the author’s name and the first “It”.
But it doesn’t stop here.
This person used GPT-3 to generate React code from a list of specifications.
Another user created a cover letter to apply for a job.
It is evident that GPT-3 is shockingly good, but what are the implications?
1. Fake News
As with any powerful tool, the use that is made of it determines its impact, and this can range quite significantly. On one side, a tool like GPT-3 can be used for generating text to support businesses in many areas, such as converting legalese into plain English. However, it is quite obvious that it can also be used to generate misleading content, specifically fake news.
GPT-3 can be used to programmatically generate misleading information, for instance by feeding it divisive and untrue prompts, such as racially charged, sexist or violent statements. This could lead to outcomes more severe than in the 2016 Election, where fake news further polarized popular opinions. Being able to generate fake news algorithmically could open Pandora’s Box.
2. Jobs
Another important, but less obvious, impact is on jobs. A tool like GPT-3 can cause many jobs to become obsolete, including journalists, writers and scriptwriters, to name a few. However, the impact is much more extended than that.
The ability to generate code directly from the specifications could make software engineers, at least as we know them now, redundant as well. Software Engineering could become a much different job, where, rather than writing the code itself, people would write in plain English the specifications, algorithms and functionalities, for then letting a model like GPT-3 generate the code.
3. Bias
GPT-3 is shockingly powerful because it is trained on billions of examples and has billions of parameters. However, a model is only as good as the data it is trained on. If a model is trained on bad data, it will perform poorly.
It follows that, if a model is trained on biased data, it will learn such bias. This is true of GPT-3 as well. The model has already been shown to generate offensive, sexist and racist text when prompted with difficult topics.
This is due to the fact that it is trained with text from multiple sources that can contain opinions of people with existing biases. Letting a model like this freely generate text without restrictions might foster the radicalization and polarization of opinions by spreading bias and fostering divisions.
Summary
OpenAI’s GPT-3 model is a true revolution. The tool is not just able to predict text, but to generate it across lengthy paragraphs keeping a coherent stream of “thought”. However, such a powerful tool can also become a weapon if used in the wrong hands. If left uncontrolled, such a system could have a dramatic impact on fake news, jobs and the spreading of bias.
