Earlier than DeepSeek R1 grew to become an AI sensation that crashed the US inventory market this week, early variations from the Chinese language AI startup recognized themselves as variants of ChatGPT.
After the Chinese language researchers revealed their work explaining the breakthrough coaching strategies that allowed them to develop a reasoning AI mannequin pretty much as good as ChatGPT o1, OpenAI accused DeepSeek of distilling ChatGPT to coach variations of DeepSeek. That’s in opposition to ChatGPT’s phrases of service.
It’s additionally ironic that OpenAI, which scraped the web of all the pieces it may discover to coach ChatGPT, together with copyright content material, is now complaining that somebody is stealing its work.
Quickly after, safety researchers uncovered a large DeepSeek safety vulnerability that accounts for the primary huge DeepSeek hack. Additionally they discovered many similarities between OpenAI and DeepSeek techniques “all the way down to particulars just like the format of the API keys.” This additional advised that the Chinese language AI agency took quite a lot of inspiration from OpenAI.
The proof retains piling up, as a special AI agency speculates that DeepSeek is perhaps a distillation of ChatGPT.
Originality.ai launched a weblog titled Did DeepSeek Copy ChatGPT and is it Detectable? The latter a part of the query refers to what Originality AI can do. The service identifies with excessive accuracy whether or not the textual content it’s taking a look at has been written by a human or generated with an AI.
Originality does this with each new AI mannequin, repeating the experiment with DeepSeek. The corporate used 150 textual content prompts, together with 50 rewrite prompts, 50 rewrite human-written textual content prompts, and 50 prompts to jot down articles from scratch.
Unsurprisingly, Originality AI was in a position to detect DeepSeek-written textual content with excessive accuracy. Its fashions (3.0.1 Turbo and Lite 1.0.0) detected DeepSeek textual content with 99.3% accuracy. That’s nice information for anybody trying to put textual content samples by means of a detector like Originality AI. As DeepSeek coaching and effectivity breakthroughs is perhaps, the AI can’t reliably idiot these techniques.
What’s uncommon within the take a look at is that Originality AI was too good at detecting DeepSeek-generated textual content on the primary strive.
“Every time a brand new LLM comes out, we run a take a look at to judge our AI detector’s efficacy and till at this time we sometimes see a slight drop off in accuracy when a brand new mannequin is launched,” the researchers wrote. As soon as that occurs, the researchers retrain the Originality fashions to extend the detection accuracy for the brand new AI merchandise.
“Nevertheless, with DeepSeek we aren’t seeing that dip in accuracy. Each of our fashions had been in a position to detect DeepSeek content material with 99%+ accuracy,” the weblog reads. “So, based mostly on our analysis, it’s attainable that DeepSeek may very well be a distilled model of ChatGPT.”
This isn’t conclusive proof that DeepSeek distilled (copied) ChatGPT, nevertheless it additional helps this declare. OpenAI alleges that DeepSeek might need used knowledge from ChatGPT to coach DeepSeek to supply the form of prompts customers (people) would need.
If DeepSeek discovered from ChatGPT knowledge methods to format responses, which are available in textual content kind, then It could generate any textual content in the identical type. Originality AI is already aware of how ChatGPT writes, as researchers educated it to detect OpenAI’s textual content era. The excessive accuracy of detecting DeepSeek textual content suggests the Chinese language startup might need used ChatGPT to coach its fashions effectively earlier than reaching R1.