OpenAI Threatens to Ban Users Probing Strawberry AI’s Reasoning

OpenAI’s New AI Model: Strawberry and Its Reasoning Claims

OpenAI recently released an AI model named “Strawberry,” available as an open preview. This model boasts a key feature: the ability to perform “reasoning.” However, OpenAI insists on keeping the AI’s internal thought processes secret.

Users Face Bans for Asking About AI Thought Process

Reports from Ars Technica reveal that OpenAI threatens to ban users who try to uncover how Strawberry thinks. Many customers reportedly received emails warning them that their ChatGPT queries were flagged for “attempting to circumvent safeguards.”

The emails state:
“Additional violations of this policy may result in loss of access to GPT-4o with Reasoning.”

This crackdown seems ironic, considering Strawberry’s main appeal lies in its “chain-of-thought” reasoning. This feature allows the AI to explain its problem-solving step by step.

What Triggers OpenAI’s Policy Violations?

Reports vary on what prompts these flags. Some users say merely using the phrase “reasoning trace” triggered warnings. Others claim that even mentioning “reasoning” caused their queries to be flagged.

While users can still access a simplified summary of Strawberry’s reasoning, this summary is created by a second AI model. Consequently, it lacks the depth of the original internal thought process.

OpenAI’s Official Explanation and Business Motives

In a blog post, OpenAI explains that hiding the AI’s detailed chain-of-thought helps avoid filtering its reasoning publicly. This prevents the AI from generating non-compliant or unsafe content while “thinking out loud.”

However, OpenAI also admits this policy helps maintain a “competitive advantage.” By limiting access to the AI’s internal reasoning, OpenAI prevents competitors from easily replicating its technology.

Concerns About Transparency and AI Safety

This strategy places more responsibility for AI alignment solely on OpenAI. It limits transparency and hampers the work of “red-teamers”, security researchers who test AI models to improve safety.

AI researcher Simon Willison criticised the policy, writing,
“Interpretability and transparency are everything to me. The idea that key details of how prompts are evaluated are hidden feels like a big step backwards.”

OpenAI’s Path Toward Opaqueness

Currently, OpenAI appears to favour making its AI models increasingly opaque black boxes. While this protects their competitive edge, it raises concerns over AI openness and the ability of developers to understand and improve these systems.

OpenAI’s strict stance on probing Strawberry’s reasoning reflects a shift away from its earlier support for open-source AI. While protecting intellectual property and safety is important, many experts worry that this secrecy will hinder transparency and innovation.

#OpenAI #AITransparency #StrawberryAI #ChatGPT #AIReasoning #ArtificialIntelligence #AIResearch #TechNews