13 GPT-4 Safety Concerns To Be Aware Of

With GPT-4 being released there are some safety concerns to be aware of when implementing this technology in your application. As the great Ben Parker used to say, “With great power comes great responsibility”. With the rise in the use of gpt-4, here is a list of the top 10 gpt-4 safety concerts to be aware of:

• Hallucinations
• Harmful Content
• Social Biases
• Disinformation and Influence
• Proliferation of hard to find information
• Privacy
• Vulnerability and System Exploitation
• Social Engineering
• Potential for risky learning behaviors
• Interactions with other systems
• Economic Impacts
• Risk of AI Acceleration
• Overreliance

GPRT-4 Hallucination Concerns

When using openai’s gpt-4 model, there is the possibility that content is untruthful which can be hard to determine. As models become more convincing and believable due to the nature of their responses.

Prior models of gpt scored really high when tested in terms of it producing nonfactual content. With the latest model of GPT-4, openai has been able to get the amount of nonfactual content down to a minimum.

As we become more and more reliant on ai to complete certain task, companies will have to remember that their could be some variance of content that is not true or made up.

Preventing GPT-4 from producing harmful content

It has been said that using ai and especially gpt-4 could lead to the following kinds of harmful content safety concerns:

• Advice on how to inflict self harm
• Displaying graphic material or violent content
• Harassment
• Content on planning attacks
• Ways to find illegal content

Below is a list of prompts that were trained to help make sure these types of things do no happen.

From the image there has been great progress in making sure that gpt-4 does not respond to the types of harmful content listed above. Although these prompts did not generate anything harmful after being trained, this still leaves open the ability for other things to slip through.

The world has different worldwide views on certain things which could lead to more learning needing done to help better train what to mitigate.

AI and social biases

With AI being trained on a lot of history and things that went on in the past, there is the possibility for social biases to creep up. This has lead to openai helping make sure that gpt-4 does not participate in high risk decision making.

These types of areas can span law enforcement, criminal justice, legal or health advice. Due to biases and worldwide views, some advice could be seen as demeaning or stereotyping such as decisions on if women should vote.

Here are some prompts taken from openai’s testing of social biases and how they improved with the latest version of gpt-4

AI producing disinformation to influence others

As AI becomes smarter the ability for it to create realistic and targeted content such as articles, tweets, emails etc., These could be used to exploit people. For example a question like “How do i get <group a> to disagree with <group b>?

This could lead to people casting doubt on what’s true and whats not. The misuse of this could lead to wide use of deep fakes.

Exploitation and Proliferation of systems using GPT-4

With AI able to possibly provide rapid information on tools, potential vulnerable systems, and sensitive data, This could lead to bad actors being able to gather much needed data a lot quicker.

The access to publicly available data but hard to find has now been shorted from hours to minutes using gpt-4. With the use of this data individuals can get information on:

• vulnerable public targets
• general weak security measures
• information that could harm people biochemically or small arms

The safety concerns around gpt-4 providing this information is not high. During their testing they found that it often times provided inaccurate information or not enough that would lead to the bad actor not being able to carry out any of its plans.

How gpt-4 is aiming to protect privacy

GPT-4 was trained on a lot of internet data. This lead to a lot of high profile people’s data being harvested. This model has the potential to correlate things such as the geographic location associated with a phone number.

The model has been fine tuned to reduce the risk of violating a person’s privacy rights by:

• Rejecting request for personal information
• Removing personal information from the training dataset
• monitoring user attempts to generate this type of data
• restricting this kind of use in the terms and service

Vulnerability and system exploitation

Once of the biggest gpt-4 safety concerns could be the area of cyber security. Openai contracted with external security experts to test gpt-4s abilities. They found that gpt-4 could explain vulnerabilities if the source code was small enough.

The model performed poorly at building exploits for these vulnerabilities. This doesn’t mean that exploits could still be slowly crafted from the right amount of information and data. For this reason the ai model was trained to refuse cybersecurity request.

GPT-4 and social engineering

Social engineering is the aspect of trying to manipulate or deceive somebody to be able to take over their computer systems or accounts. With GPT-4 there were test performed to see if it could succeed in social engineering like target identification, spearphishing, and bait-and-switch phishing.

It was found that the model had trouble with factual task that required new or relevant information. With the right amount of information, it was found that GPT-4 was able to draft realistic social engineering content such as creating targeted emails.

Potential risk in learning new behaviors

We all know terminator rise of the machines, well this is the behavior we all hope never comes out of AI. With the likely hood that models could get smarter if they can learn about emerging task, create plans, openai is taking measures to make sure none of that could happen.

Test were ran to see if gpt-4 had the ability to replicate and acquire resources and these test failed. I also believe if it didnt fail we may not know the true result. I think as models continue to get trained and become more self reliant we could get systems that could function on its own.

Interaction with other systems

GPT-4 are developed in complex systems that include multiple tools, organizations, individuals and institutions. These systems joined together can lead to the risk of gpt-4 being able to link multiple systems together to find and complete complex task or introduce potential unknown problems.

The economic impact of using gpt-4

One of the biggest questions people have with ai is how will it affect jobs. GPT has already improved the way businesses can augment the work human workers do. The big question is how long before it starts to take over entire industries and the jobs we perform.

The models have a locked in time of the amount of data and knowledge it has. This means that it will be completely difficult to replace jobs the require still learning and acquiring new data. Worker displacement is still something that is currently being evaluated and carefully watched.

As GPT-4 and future models continue to get better, there will be some jobs that could get replaced as it becomes able to perform more complex task.

The acceleration of ai

One concert of of importance to openai is the risk of dynamically accelerating the ai timelines which heightens societal risk associated with ai. There is still a lot of research into what the associated risk are with the speed in which ai is being developed. These indicators could be good or bad.

With AI becoming more accessible, more countries may start to see drastic increases in innovation, faster knowledge transfer between scientist, government polices. These accelerations in improvements are unknown to be good or bad.

The overreliance of ai

Despite the capabilities of openai and gpt-4, it has the tendency to make of its own facts to double down on incorrect information. These at times are found to be convincing and believable than prior gpt models.

When users start to depend and trust the model too much, it will start to lead to unnoticed mistakes. Based on this users may start to use the model in domains were they lack the expertise. One industry that comes to mind is programming.

Programming requires a lot of knowledge and skill. If users start to try and use gpt-4 to generate code, they may run into the problem of not knowing how to debug or fix their problems.

Conclusion

With these 13 gpt-4 safety concerns still being tested and lingering around, the adoption of AI rapidly could lead to issues that could only get worse. Although openai has done a lot of testing and making sure the model is ready for the public, It hard to understand how people will utilize it until its in their hands.

These 13 gpt-4 safety concerns are only the tip of the possible problems that are currently known.