In 2019, OpenAI withheld the full release of its language model, GPT-2, fearing that its ability to write convincing text could be misused.
Imagine this: One morning, you tell an artificial intelligence, “Write a convincing news article based on what happened yesterday,” and it instantly creates a piece more plausible than something a human could write. While this is quite a familiar scene for us today, in 2019, this capability brought both amazement and fear. This is the story of the debut of OpenAI’s ‘GPT-2’, a very significant event in the history of artificial intelligence.
Why is this important?
| In 2019, the language model GPT-2, announced by the non-profit AI research firm OpenAI, shook the global tech industry. It went beyond the level of writing grammatically correct sentences to producing text that was incredibly precise and persuasive. At the time, OpenAI expressed serious concerns that this technology could be used to generate mass amounts of fake news or be utilized for malicious purposes to deceive people [Source: OpenAI built a text generator so good, it’s considered too dangerous | TechCrunch](https://techcrunch.com/2019/02/17/openai-text-generator-dangerous/). |
This was an unprecedented event where the company that developed the technology itself refused a full release, stating, “Our technology is so advanced that it could be dangerous.” By the fact that a company began to reflect directly on the negative impact technology could have on society, it became a crucial turning point for the ‘AI safety’ discussions we have today Source: OpenAI built a text generator so good, it’s considered too dangerous - Know the Technology News.
Easy to understand: A master of guessing the next word
So, how exactly did GPT-2 write like that? To put it very simply, GPT-2 was a master of the ‘guess the next word’ game.
To use an analogy, do you remember playing ‘word chain’ with friends when you were young? GPT-2 studied the patterns of which words have a high probability of following others by reading all the books and internet text in the world. Having studied a massive 40 gigabytes of text data, GPT-2 had an exponentially superior ability to grasp context and select the most natural next word Source: Meet OpenAI’s Text Generator That’s Considered Too Dangerous To Release - AfroTech.
It’s as if a highly competent assistant read tens of thousands of books and, if you gave it just the first sentence, could perfectly write the content that should follow. People felt as if the AI was thinking and writing for itself like a human, because this ‘next-word prediction ability’ was so precise.
Current situation
| At the time, OpenAI researchers stated, “Due to concerns about malicious use, we have decided not to release the entire trained model” Source: Meet OpenAI’s Text Generator That’s Considered Too Dangerous To Release - AfroTech. They broke with convention and decided to hide the full training code and data [Source: New AI fake text generator may be too dangerous to release, say creators | AI (artificial intelligence) | The Guardian](https://www.theguardian.com/technology/2019/feb/14/elon-musk-backed-ai-writes-convincing-news-fiction). |
Instead, OpenAI chose to release a version that was much smaller in scale than the original model Source: OpenAI says its text-generating algorithm GPT-2 is too dangerous to release.. This decision served as a catalyst for igniting full-scale social discussion on the potential threats posed by the advancement of AI technology and how to handle it safely.
What will happen in the future?
The event of 2019 was just the beginning. The concerns at that time have now become even more realistic homework assignments. Verifying how factually consistent the text generated by artificial intelligence is (factuality) and establishing technical and social safeguards to block attempts to misuse AI remain important tasks Source: Why OpenAI’s fake news warnings are a bit overblown - TechTalks. Now, ‘how safely can we use it’ is becoming the center of technological development rather than ‘how high-performance can we make it’.
MindTickleBytes’ AI reporter perspective
The shock that GPT-2 sent into the world was not just simple tech bragging. It was a record of a company acknowledging that technology has the power to change society and choosing to pause for a moment. We must remember that the convenient AI services we enjoy today grew upon the cautious considerations of those days.
References
-
[OpenAI built a text generator so good, it’s considered too dangerous TechCrunch](https://techcrunch.com/2019/02/17/openai-text-generator-dangerous/) -
[New AI fake text generator may be too dangerous to release, say creators AI (artificial intelligence) The Guardian](https://www.theguardian.com/technology/2019/feb/14/elon-musk-backed-ai-writes-convincing-news-fiction) - OpenAI says its text-generating algorithm GPT-2 is too dangerous to release.
- OpenAI built a text generator so good, it’s considered too dangerous to release - Know the Technology News
-
[The AI Text Generator That’s Too Dangerous to Make Public WIRED](https://www.wired.com/story/ai-text-generator-too-dangerous-to-make-public/) - Meet OpenAI’s Text Generator That’s Considered Too Dangerous To Release - AfroTech
- Why OpenAI’s fake news warnings are a bit overblown - TechTalks
- Because there were too many technical errors
- Concerns about potential misuse and automated fake news
- Because the development costs were too high
- By analyzing images and converting them into text
- By learning from text samples to predict the next word
- By learning through real-time conversation with humans
- Released the entire dataset and code completely
- Released only some information and withheld the full model
- Did not release it to the public and sold it only to research institutes