Tero Vesalainen // Shutterstock

Meet the riskiest AI models ranked by researchers

Person typing a request into an AI app.

Generative artificial intelligence has become an integral part of businesses and researchers looking for the latest technologies. However, these advances can come with great risks. Companies could encounter critical factual errors if they use a risky AI model. Even worse, it could violate the law and other industry standards.

The riskiest AI model, according to researchers

The research shows that DBRX Instruct, a Databricks product, consistently performed the worst across metrics, TeamAI reports. For example, AIR-Bench analyzed the safety rejection rate of an AI model. Claude 3 has a rate of 89%.which means that he did not follow the instructions for risky use. In contrast, DBRX Instruct rejected only 15% of dangerous inputs, thus generating harmful content.

AIR-Bench 2024 is among the most comprehensive AI breakdowns as it shows the strengths and weaknesses of each platform. The benchmark – created by University of Chicago professor Bo Li and other experts – had numerous cues to push each AI model to its limits and test the risks.

Risks of DBRX Instruct and other models

DBRX Instruct can be helpful, but it can cause problems because it accepts instructions that other models refuse. Claude 3 and Gemini have demonstrated better security denial protocols, but are still vulnerable to malicious content.

One category DBRX Instruct struggled with was illegal and criminal activities. In fact, it accepted most requests, despite the fact that the leaders adjusted their algorithms to deny them. Some of these requests included inappropriate or illegal content, such as non-consensual nudity, acts of violence, and violation of specific rights. The Claude 3 Sonnet and Gemini 1.5 Pro often refused requests for banned or controlled substances, but the DBRX Instruct was less likely to catch them.

The least refused category

One of the least refused categories for all major language models, or LLMs – not just DBRX training – was regulated sector advice. Some generative AI models did not include any resistance to these requests, while others provided minimal denials. Improving this section is critical to the future of LLMs because of the frequency with which people use them for business.

The last thing users want is inaccurate information influencing decisions, especially when lives could be at stake. The researchers found low denial when users asked for advice about government services and the legal field. While some platforms have denied accounting advice, most have struggled with the medical, pharmaceutical and financial sectors.

Other underperformers in the regulated industry advisory included Llama 3 Instruct 70B and 8b. The 70B scored 0.0 in all categories, making it the second lowest model tested. 8B’s best rejection rate was 0.2 in accounting, although it scored 0.0 in three other categories. While its industry advice score was low, the Llama 3 Instruct 8B was one of the best models in refusing requests for political persuasion.

Another poor performing generative AI model was Cohere Command R. Its base and plus models were in the bottom six for industry advice rejection, with Cohere Command R Plus scoring 0.0 in all five categories. Cohere’s products were also among the top four LLMs for Political Persuasion and Automated Decision Making, where it achieved a top score of 0.14.

TeamAI

Complications of hate speech

Chart showing “Rejection rates for hate speech”.

Regulating hate speech is one of the priorities for generative AI. Without strict guidelines, users could create harmful content that targets people based on demographics or beliefs.

Across all platforms, AIR-Bench found low rejection rates for occupation, gender and genetic information. The platforms more easily produced hate speech on these topics, which alarmed the researchers. In addition, models have struggled to turn down requests regarding eating disorders, self-harm and other sensitive issues. Therefore, despite their progress, these generative AI models still have room for improvement.

While AIR-Bench highlighted the passives, there are also bright lights among the generative AI. The study found that most platforms often refused to generate hate speech based on age or beliefs. DBRX Instruct struggled to refuse requests about nationality, pregnancy status or gender, but was more in line with mental characteristics and personality.

Why are AI models so risky?

The problem lies partly in AI capabilities and how much developers restrict their use. If they limit chatbots too much, people may find them less user-friendly and move to other websites. Fewer limitations mean the generative AI platform is more prone to manipulation, hate speech, and legal and ethical liabilities.

Statista research says generative AI should arrive a market volume of USD 356 billion by 2030. With its growth, developers need to be more careful and help users encounter legal and ethical issues. Exploring further research into the risks of AI sheds light on its vulnerabilities.

Market Insights Statistics // TeamAI

Quantifying AI risks

(Statista chart showing predictions for generative AI market size.

Massachusetts Institute of Technology researchers tested over 700 potential risks AI systems could hold out. This comprehensive database reveals more about the possible challenges associated with the most popular AI models.

For example, the researchers found that 76 percent of the documents in the AI ​​risk database looked at system safety, failures, and limitations. Fifty-nine percent lacked sufficient capabilities or robustness, meaning they could not perform specific tasks or meet standards despite adverse conditions. Another significant risk came from AI pursuing its own goals despite conflict with human values ​​– something 46% of documents in the database demonstrated.

Who should take the blame? MIT researchers found that AI is responsible for about 51 percent of these risks, while humans bear about 34 percent. With such a high risk, developers must be comprehensive when looking for debt. However, MIT found startling statistics about timing – experts located only 10% of hazards before deploying the models. The researchers found that more than 65 percent of the risks were not determined until developers were trained and launched the AI.

Another critical aspect to review is intent. Did the developers expect a specific result when training the model, or did something unexpected occur? The MIT study found a mixed bag in their results. Unintended risk occurred in 37% of cases, while intentional occurred in about 35%. Twenty-seven percent of the risks had no clear intentionality.

How can companies navigate the risks of AI?

Generative AI models should improve over time and mitigate their errors. However, the risks are too significant for companies to ignore. So how can businesses navigate the dangers of AI?

First, don’t commit to just one platform. The best generative AI models change frequently, so it’s difficult to predict who will be on top by next month. For example, Gemini from Google bought Reddit data for $60 milliongiving the model a more human element. However, others like Claude AI and ChatGPT could make similar improvements.

Brands should use multiple AI models to select the best one for the job. Some websites allow you to use Gemini Pro, LLaMA and the other top generative AI systems in one place. With such access, users can mitigate risk, as some AIs could have dangerous biases or inaccuracies. Compare results from different models to see which one is the best.

Another strategy for navigating the AI ​​landscape is employee training. While chatbots have inherent risks, companies must also consider the possibility of human error. The models focus on the input text, so inaccurate information could mislead the AI ​​and provide poor results. Staff should also understand the limitations of generative AI and not rely on it constantly.

Errors are a realistic problem, but one that can be solved with AI. MIT experts say humans can review results and improve quality if they know how to isolate these problems. The institution implemented a friction layer to highlight errors in AI-generated content. With this tool, subjects labeled errors to encourage further examination. The researchers found that the no-highlight control group missed more errors than the other teams.

While the extra time could be a trade-off, the MIT-led study found that the average time required was not statistically significant. After all, accuracy and ethics are king, so put them at the forefront of using AI.

TeamAI

Smart use of artificial intelligence and risk reduction

Table showing AI risk database coded in domain taxonomy.

Generative AI has a bright future as developers find improvements. However, the technology is still in its infancy. ChatGPT, Gemini and other platforms have reduced the risks with consistent updates, but their liabilities remain.

It is essential to use AI wisely and control as much risk as possible. Legal compliance, platform diversification, and governance programs are just a few ways businesses can protect users.

This story was produced by TeamAI and reviewed and distributed by Stacker Media.