Research Preview

When a company releases a new program, it is usually thoroughly tested. With a research preview, things are different. The provider puts an early version online and openly states: this doesn’t work reliably yet. Users are allowed to try it out, but shouldn’t rely on it. The term comes from the world of artificial intelligence, that is, from the development of programs that generate language, images, or code themselves. It became well known in 2022, when OpenAI announced its chatbot ChatGPT in exactly this way.

Why companies put unfinished AI online

An AI system can only be tested to a limited extent in the lab. Developers can never anticipate all the questions that millions of people will actually ask. Only in real-world use does it become clear where the system gets things wrong, becomes offensive, or can be tricked. A research preview is therefore primarily a massive test series with voluntary participants.

There is also an economic reason. Competition among the major AI providers is fierce, and whoever gets to market first shapes public perception. An early launch secures attention, press coverage, and user numbers. As a research preview, ChatGPT reached around 100 million users within two months. No fully polished product could have achieved that faster.

But the term also has a legal side. Anyone who explicitly labels their system as a research preview makes no promise of reliability. If the AI then makes nonsensical claims, or a company passes on incorrect information as a result, the provider can point to the warning notice. Critics call this a convenient excuse: money is already being made, but officially it’s still just an experiment being paraded around.

What actually happens during this phase

Technically, a research preview usually differs little from the later full version. The difference lies in the surrounding conditions. Guarantees of availability are often missing, access is limited, and features can disappear without warning. Some previews are free, because usage itself is the compensation.

User feedback is central. Many systems have thumbs-up and thumbs-down buttons under every answer. These ratings feed into further training. The process is called reinforcement learning from human feedback: the model is fine-tuned to favor answers that people found good. Reported misbehavior also ends up on internal lists, against which new versions are checked.

It helps to distinguish this from similar terms. A beta version is essentially finished and is only still being checked for bugs. A research preview, by contrast, can still be fundamentally changed at its core, or discontinued altogether. Another common misconception is equating preview with small or harmless. Some of these versions are the most capable models that exist at that point in time.

The notice below the chat window

In everyday use, the term usually appears as small text at the edge of an interface. Sentences like “This preview may contain errors, verify important information” are typical. Anyone who sees such notices should not adopt results uncritically for homework, job applications, or medical questions. This applies especially to figures, quotes, and source citations.

In business news, the expression comes up when a provider unveils something new. Reports about new language models, video generators, or AI search features often explicitly mention this classification. For investors, it’s a signal: the product is already tying up data centers and money, but isn’t yet generating reliable revenue. For companies wanting to deploy such systems, it means that interfaces may still change.

Some previews are shut down again after a few weeks if the results are disappointing. Others quietly grow into regular operation without the label ever disappearing. This is exactly what makes this category special: it says less about the technology than about the claim the provider currently wishes to make.