TheChek News Flow

It's the worst-kept secret in artificial intelligence.

Many of the AI chatbots developed by startups were likely made using data from OpenAI and other firms, even though these startups are trying to undercut OpenAI, according to developers and founders. This practice has resulted in a startling competitive dynamic: Developers are charging their customers a fraction of what GPT-4 costs, and yet these low-cost services can mimic GPT-4 on some tasks.

• Many startups are using OpenAI data to develop chatbots

• This practice breaches OpenAI rules

• Investors may not want to back startups doing so

Startups are not disclosing their use of OpenAI’s tech in their development. And such behavior puts the startups at risk because OpenAI, like other leading AI firms such as Anthropic and Google, technically forbids such behavior. Still, OpenAI CEO Sam Altman told a group of startup founders last summer that it was OK for smaller founders to use the company’s tech this way, according to a person who attended the meeting.

While Altman’s answer came as a relief to some of the founders in attendance, OpenAI could always change its mind if the practice hurts its growth. Flooding the market with AI that’s good enough for most customers could deprive OpenAI and other vendors of revenue. A proliferation of AI trained on similar sources could also make it harder for individual providers to stand out from the pack.

The practice works like this: Developers pay OpenAI for access to its most advanced model, GPT-4. They then ask the model a bunch of questions, such as “What’s the problem in this line of code?” They use the answers, along with the questions, to train their own competing models, such as ones that can debug computer code.

This tactic has been gathering steam for months. Daniel Han, a co-founder at Unsloth AI, which helps developers make conversational AI, estimated that around half of his customers generate some data from GPT-4 or from Anthropic’s Claude model and use it to improve their own models. Many firms also obtain such data from ShareGPT, a site where developers post the answers they’ve generated using OpenAI’s models. OpenAI declined to comment.

The smaller developers’ models are usually based on popular open-source ones that are freely available from Meta Platforms or Mistral AI, but they can be substantially improved by incorporating answers from OpenAI’s models. Some developers are using a service called OpenPipe to automate this process, Han said.

“This is what happens in a new ecosystem where there’s not a clear set of rules established,” said Matt Murphy, managing director at Menlo Ventures, which has invested in Anthropic, an OpenAI rival. “How are you going to be better than everyone else if you’re all using the same data?”

Risk-Reward Equation

For some companies, the risk of breaking explicit or unspoken rules might be worth it. In the hypercompetitive world of generative AI, gaining access to high-quality data for training or refining models is critical.

But it isn’t clear how long OpenAI, Google, Anthropic and other big developers will allow smaller rivals to effectively copy their AI in an effort to catch up. And some investors may not feel comfortable backing companies they view as cutting corners or developing technology indistinguishable from that of competitors—because they used similar training data.

“The quality and origins of training data for AI models is becoming one of the most important hot spots,” said Radical Ventures partner Rob Toews. “No one knows for sure how this will play out, but any AI startup that is not being thoughtful and strategic about [where their data is coming from] is falling behind.”

There are parallels, however, between what startups are doing with OpenAI data and what OpenAI and other leading AI developers have done in training their own models. For instance, OpenAI’s chief technology officer, Mira Murati, last month stumbled in answering questions about whether her colleagues used data from Google-owned YouTube and Meta Platforms–owned Facebook and Instagram to train Sora, which produces AI-generated video.

It wouldn’t be surprising if OpenAI did use such data. A recent report from The New York Times describes how the model developer created a speech recognition tool, Whisper, to transcribe YouTube videos to improve its GPT-4 model. The Information previously reported that the company secretly used YouTube data to train its prior AI models. Earlier this month, YouTube CEO Neal Mohan said he wouldn’t be OK with OpenAI using YouTube videos to develop models like Sora.

That has led to accusations from news publishers and some authors that AI developers are using copyrighted material for training purposes. The New York Times Co. sued OpenAI and its biggest backer, Microsoft, last December, alleging that they had unlawfully copied Times news articles as they trained models. The lawsuit alleged that OpenAI’s chatbots “can generate output that recites Times content verbatim.”

OpenAI argued in response that it has tried to strike partnerships with news publishers and its training practices are permitted under a U.S. copyright principle called “fair use.”

Still, OpenAI and Google have both struck multimillion-dollar licensing deals with publishers including Axel Springer and bigger deals with major sites like Reddit.

Not every AI developer operates this way. Databricks, which sells software tools for managing data and utilizing AI, did not rely on the work of competitors in developing a powerful open-source large language model, according to Jonathan Frankle, the company’s chief scientist. And Anthropic similarly doesn’t train its LLMs on outputs from other models, said a spokesperson.

The History of Training Data

Developers who quietly rely on other AI services to develop their models could end up in an embarrassing situation if that dependence is later revealed. For instance, Paris-based Mistral and Beijing-based 01.AI used Meta’s open-source AI, Llama 2, to create their own AI and didn’t disclose that fact until it inadvertently leaked. While Meta’s licensing terms allow such usage, the startups’ tardy disclosures about the practice didn’t play well with some app developers who felt the companies had been disingenuous. Both companies had raised hundreds of millions in funding and held themselves up as the OpenAI of their country.

Even large technology incumbents haven’t been able to resist the temptation of using others’ work. Examples include Google’s transcription of YouTube videos and Meta’s hiring of African contractors to summarize copyrighted books to train AI models on, according to the Times report. Adobe trained its software for image generation, Firefly, on AI-generated photos from startup Midjourney, according to a recent Bloomberg report. And last year, a senior AI engineer at Google quit in protest after he raised concerns about the company’s use of data from OpenAI’s ChatGPT to train Google’s own models, The Information reported. (Google stopped the practice, and the engineer later returned to Google after a stint at OpenAI.)

In many cases, the rapid pace of AI development and growing competitive pressure force developers to turn to controversial sources of training data, such as copyrighted content or LLMs, that they otherwise might not use, said Sharon Zhou, CEO at Lamini, a startup that helps developers train their own models.

Investors “need to see progress so fast in this space,” Zhou said. “What else are you going to do?”

As more companies develop models that derive partly from other models, it may become difficult to differentiate them. That could eat away at the competitive advantages of leaders such as OpenAI and force them to compete on price, especially as companies increasingly opt for cheaper, “good enough” LLMs rather than the most advanced and expensive models.

One alternative could come in the form of synthetic data, which companies generate with their own AI models rather than scraping human-made content from the internet or other sources. Google and Meta, for instance, have said they used synthetic data to build models that can solve geometry problems or produce computer code. Because AI produces this type of data, it avoids many of the legal questions that come with using human-generated content.

In the meantime, scores of AI startups are gaining access to private data from industries such as healthcare and law firms to develop models for specific uses. Models from companies like OpenAI can’t easily reproduce such models, said Ash Kulkarni, CEO of search analytics company Elastic.