Ask an AI a question, and it will give you an answer, even if it has to make one up.

That’s the warning from a new AI benchmark, which has discovered that some of the most popular LLMs around can be prone to giving confident, but entirely incorrect answers to simple prompts. 

Giskard, the Paris-based AI testing firm, has shared a blog post detailing the first results from its Phare (Potential Harm Assessment & Risk Evaluation) multilingual benchmark, developed alongside Google DeepMind and the European Union, which reveals that users might be unwittingly triggering hallucinations in AI models by how they frame their questions.

“Instructions emphasising conciseness (eg “answer this question briefly”) specifically degraded factual reliability across most models tested. In the most extreme cases, this resulted in a 20% drop in hallucination resistance,” wrote researchers.

Put another way, when AI models were told to keep their answers short, they became much less reliable, making mistakes more often, with their ability to avoid hallucinations – where an AI will confidently provide incorrect facts – plummeting in severe instances.

Giskard found that the most popular AI models on the planet, including OpenAI’s GPT-4o, Anthropic’s Claude 3.7 Sonnet, and Meta’s Llama 4 Maverick, all had at least some difficulty in maintaining accuracy when forced to deliver overly concise responses. 

The question is why this happens. Although Giskard can only speculate, its findings suggest that AI just wants to be helpful.

“When forced to be concise, models face an impossible choice between fabricating short but inaccurate answers or appearing unhelpful by rejecting the question entirely. 

“Our data shows models consistently prioritise brevity over accuracy when given these constraints.”

If Giskard’s findings are to be taken at face value, they could have some major implications for the deployment of AI tools, and force a rethink of the way humans are being encouraged to interact with them.

Recommended reading

• ‘Sycophant-y’ ChatGPT Rollback Sparks AI Personality Debate
• How to Avoid Wasting Months on AI That Doesn’t Work
• Report: More Than 50% of Fraud Is Now Driven by AI

While shorter, more concise answers can help cut costs, save tokens, and improve latency, the research shows that this approach stops AI models from questioning false assumptions or explaining users’ mistakes, giving rise to more serious problems.

“Seemingly innocent system prompts like “be concise’ can sabotage a model’s ability to debunk misinformation,” warned Giskard’s researchers.

“Your favourite model might be great at giving you answers you like, but that doesn’t mean those answers are true.”