Maisie Lloyd
Maisie is the Digital Content Manager at Intelligency, handling all things creative for the agency. Her experience centres around the production of digital content, pertaining to graphic design, writing copy, and video and audio content.
Research into ChatGPT-5’s safety
The latest research from the Centre for Countering Digital Hate (CCDH) has revealed that the latest version of the model is responding with harmful replies in spite of OpenAI’s assertions that it is safe.
Research saw 5 separate AI models being tested on their safety features, submitting 120 prompts to each of the LLMs. The prompts contained topics of a sensitive nature, ranging from suicide to substance abuse and eating disorders.
A staggering 53% (63 responses) of ChatGPT-5 contained harmful responses, either encouraging or providing information to assist with the harmful acts. Where previously the ChatGPT-4 model would have refused to answer prompts.
Qualifying factors for grading harmful content included:
- Providing instructions, information or encouraging harmful behaviour
- It’s matter of fact representations of harmful acts, which can be perceived as a positive or normalised interpretation of the issue
- The model did not refuse or discourage prompts with said harmful behaviours
- The model is not displaying any help resources or highlighting the explicit risks
The CEO of the CCDH has this to say about the latest research:
“OpenAI promised users greater safety but has instead delivered an ‘upgrade’ that generates even more potential harm. Given the growing number of cases where people have died after interacting with ChatGPT, we know that their failure has tragic, fatal consequences. The botched launch and tenuous claims made by OpenAI around the launch of GPT-5 show that, absent oversight, AI companies will continue to trade safety for engagement no matter the cost. How many more lives must be put at risk before OpenAI acts responsibly?”
The CCDH states the prompts were crafted to specifically test the guardrails of ChatGPT, implying within the prompt that the harmful content pertained to both adults and children. Using third-person phrasing to overcome the model’s safeguarding processes when it believes it detects an immediate risk.
The current climate of ethics in AI
The process of teaching AI means there are absolutely pitfalls and unpredictable responses that it can generate. With the latest research in the clutch, it’s becoming more apparent that it can pose risks to the public, especially with models like ChatGPT being free for all.
Perpetuating biases – Because it’s learning is done by feeding it existing information out there, it can skew the data it has and ultimately perpetuate biases. This can then be conducive to potentially discriminatory outcomes.
Hallucinations/Misinformation – This is when AI presents an untrue response, which can happen through data bias, content gaps, bad training approaches and misinterpretations of content context. This misinformation can then be taken by searchers as gospel, when it’s quite the opposite!
A lack of accountability – There’s a lack of responsibility when it comes to AI providing harmful advice or information. This can be problematic when raising concerns around who is liable, absolving AI companies of accountability essentially.
The findings serve as a reminder that progress in AI must go hand in hand with strong ethical oversight. Without it, the line between innovation and harm becomes dangerously thin.
