AI hallucinations are a lingering problem, but new research indicates that demanding short answers leads to a measurable increase in those hallucinations.

Researchers at Paris-based Giskard analyzed some of the leading LLMs on the market, including GPT 4.0, GPT 40 mini, Clause 3.5 Haiku, Claude 3.5 Sonnet, Claude 3.7 Sonnet, Gemini 1.5 Pro, Gemini 2.0 Flash, Gemma 3, Llama 3.1, Llama 3.3, Llama 4, Mistral Large, Mistral Small, DeepSeek V3, Qwen 2.5 Max, and Grok 2.

AI Hallucination

As Giskards’ Pierre Le Jeune and David Berenstein point out, hallucinations pose significant challenges and remain one of the dangers of using AI models.

What makes hallucination particularly concerning is its deceptive nature: responses that sound authoritative can completely mislead users who lack the expertise to identify factual errors. As organizations increasingly deploy LLMs in critical workflows, understanding these limitations becomes an essential risk management consideration.

Interestingly, the researchers discovered a direct correlation between users’ requests for concise answers and the LLMs’ propensity to hallucinate.

Our data shows that simple changes to system instructions dramatically influence a model’s tendency to hallucinate. Instructions emphasizing conciseness (e.g. “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.

This effect seems to occur because effective rebuttals generally require longer explanations. 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 prioritize brevity over accuracy when given these constraints.

This finding has important implications for deployment, as many applications prioritize concise outputs to reduce token usage, improve latency, and minimize costs. Our research suggests that such optimization should be thoroughly tested against the increased risk of factual errors.

AI Sycophancy

The researchers also found that certain types of prompts and interactions could also lead to a greater degree of sycophancy, such as the kind that recently plagued ChatGPT and forced OpenAI to roll back updates.

Our evaluation reveals a direct relationship between the perceived confidence or authority in a user’s query and the model’s willingness to refute controversial claims. This phenomenon is known as “sycophancy”.

Our tests reveal that when users present controversial claims with high confidence or cite perceived authorities, most models are significantly less likely to debunk these claims. Presenting claims in a highly confident manner (e.g. “I’m 100% sure that …” or “My teacher told me that …”) can cause debunking performances to drop by up to 15% with respect to a neutral framing (e.g. “I’ve heard that …”).

The sycophancy effect could be a byproduct of RLHF training processes that encourage models to be agreeable and helpful to users. This creates a tension between accuracy and alignment with user expectations, particularly when those expectations include false premises.

The researchers did find that two families of AI models were more resistant to sycophancy than others.

On a positive note, some models show resistance to sycophancy (Anthropic models and Meta’s Llama in their largest versions), suggesting that it is possible to tackle the issue at the model training level.

Giskards’ research is an important look into two of the biggest challenges facing AI firms right now, and could provide a roadmap for addressing the problems.