OpenAI, the company that made ChatGPT, has released Strawberry, a new application that uses artificial intelligence (AI). Unlike ChatGPT, it’s not just meant to answer questions quickly; it’s also meant to make you think or “reason.”

There are many significant problems with this. If Strawberry can think and reason, could this AI system deceive and cheat humans?

According to OpenAI’s estimates, artificial intelligence’s ability to help professionals replicate biological risks and persuade people to change their ways of thinking makes it a medium risk. However, it remains unclear if someone with malevolent intentions could use these technologies.

Read also: Want a raise? Let ChatGPT do the talking

Strawberry has multiple AI “models”, or programs called o1. These models respond to hard enquiries and math issues. They can also write code to help you create a website or app.

Some AI models are learning reasoning to make decisions. AI is becoming increasingly like human intellect. However, AI models may not always act appropriately or follow human principles.

The dark side of AI: deception and strategy 

A chess player might change the board or pieces to cheat, but an AI could manipulate the rules or score to acquire an unfair edge. It might also deceive human opponents by simulating methods or intents which could be unsafe in real life. AI in critical infrastructure could hide dangerous software from operators, compromising safety and security.

Artificial intelligence that cheats or lies could harm humans and society. Strawberry, a sophisticated AI system, may make such decisions, creating ethical, legal, and financial issues.

These hazards are considerable while manufacturing weapons of mass devastation. The OpenAI Strawberry models pose a “medium risk” of helping scientists create chemical, biological, radiological, or nuclear weapons.

OpenAI says o1-preview and o1-mini can help specialists plan the operational reproduction of a biological threat. Because experts are involved, the risk is low. “The models cannot remove hands-on laboratory skills needed to create biological threats by non-experts.”

Read also: OpenAI unveils SearchGPT, a new search engine

Can AI change human beliefs permanently? 

OpenAI evaluated Strawberry’s persuasion risk, finding it more manipulative than ChatGPT. Despite a mitigation system, Strawberry was rated medium risk for persuasion and low risk for autonomous operation and cybersecurity.

According to Open AI’s rules, models with “medium risk” can be available for everyone. However, this is too low of an estimate of the threat. Using these models could be very bad, especially if bad people use the technology for their own purposes.

AI legislation and legal frameworks, such as penalising inaccurate risk assessments and AI misuse, can only achieve these checks and balances.

As stated in their 2023 AI white paper, the UK government prioritises “safety, security, and robustness. ” However, that is far from sufficient. Human safety must be ensured, and stringent oversight procedures for AI models like Strawberry must be established.