[DigitalToday reporter Jinju Hong (홍진주)] Swedish AI startup Opper AI has unveiled an experimental platform called AI Roundtable that lets users compare the judgments of different AI models in one place and even make them debate. It is drawing industry attention because it allows observation of collective decision-making among AI models beyond individual model performance.

An online outlet, GigaZine, reported on March 27 that AI Roundtable lets users select up to 50 models from more than 200 AI models, pose the same question and compare answers at a glance. It also offers functions to summarise overall views or extract and display the most persuasive logic, rather than merely listing responses. Users can also see why each AI model made its choice, allowing analysis of thought processes as well as outcomes.

The platform’s core concept began with the so-called “car wash test”. Repeated cases were found in which some AI made judgments divorced from reality even on a simple question such as whether to walk or drive when a car wash is 50 metres away, leading to an idea to expand the test to many models for comparison and verification. In an actual test, results sometimes ran counter to intuition, with nearly half of the models choosing to walk.

AI Roundtable aims to address such limitations through “Debate”. In debate mode, up to 6 AI models conduct a debate by rebutting or revising one another’s logic. Even if opinions are split in an initial vote, the debate process shows a tendency for the final conclusion to converge to one side as models are persuaded by more logical arguments. It highlights the possibility that collective judgment can become more refined than a single AI response.

The platform is relatively intuitive to use. Users enter a question and choices, select either “Poll” or “Debate” mode, and add participating AI models from a list. In poll mode, users can check the response distribution and a summary result from up to 50 models, while debate mode lets users review a decision-making process centred on interactions among a small number of models. The results screen systematically organises items such as “choices by model”, “reasons for voting” and “summary conclusion”.

The platform also lets users save and share results as PNG images. Users can also review various past cases through the “Past Roundtable” menu, including everyday and ethical questions such as choosing a direction to move in a dungeon with limited visibility or whether to wait for someone when elevator doors are closing. Sessions are set to private by default, so they are not disclosed externally unless a user shares them directly.

AI Roundtable can be used for free, and users can run more experiments or manage past question histories by obtaining an API key. A “budget counter” displayed on the screen refers to costs used to call AI models, and expanded use runs on a credit basis.

The industry sees significant potential for the platform to be used as an experimental environment that goes beyond a simple comparison tool to reveal biases and errors in AI models and find directions for improvement.